Zebra logotype

CI Docker CI OSes Continuous Delivery Coverage codecov Build docs Build lightwalletd Build Zcash Params

License

Contents

About

Zebra is the Zcash Foundation's independent, consensus-compatible implementation of a Zcash node, currently under development. It can be used to join the Zcash peer-to-peer network, which helps keeping Zcash working by validating and broadcasting transactions, and maintaining the Zcash blockchain state in a distributed manner.

Zcash is a cryptocurrency designed to preserve the user's privacy. If you just want to send and receive Zcash then you don't need to use Zebra directly. You can download a Zcash wallet application which will handle that for you.

Please join us on Discord if you'd like to find out more or get involved!

Using Zebra

You would want to run Zebra if you want to contribute to the Zcash network: the more nodes are run, the more reliable the network will be in terms of speed and resistance to denial of service attacks, for example.

Zebra aims to be faster, more secure, and more easily extensible than other Zcash implementations.

Beta Releases

Every few weeks, we release a new Zebra beta release.

Zebra's network stack is interoperable with zcashd, and Zebra implements all the features required to reach Zcash network consensus.

Currently, Zebra validates all of the Zcash consensus rules for the NU5 network upgrade.

But it may not validate any:

  • Undocumented rules derived from Bitcoin
  • Undocumented network protocol requirements

Getting Started

Building zebrad requires Rust, libclang, and a C++ compiler.

Build and Run Instructions

zebrad is still under development, so there is no supported packaging or install mechanism. To run zebrad, follow the instructions to compile zebrad for your platform:

  1. Install cargo and rustc.
    • Zebra is tested with the latest stable Rust version. Earlier versions are not supported or tested, but they might work. (Rust 1.57 and earlier are not supported, due to missing features.)
  2. Install Zebra's build dependencies:
    • libclang: the libclang, libclang-dev, llvm, or llvm-dev packages, depending on your package manager
    • clang or another C++ compiler: g++, Xcode, or MSVC
  3. Run cargo install --locked --git https://github.com/ZcashFoundation/zebra --tag v1.0.0-beta.13 zebrad
  4. Run zebrad start (see Running Zebra for more information)

For more detailed instructions, refer to the documentation.

Optional Features

For performance reasons, some debugging and monitoring features are disabled in release builds.

You can enable these features using:

cargo install --features=<name> ...

System Requirements

The recommended requirements for compiling and running zebrad are:

  • 4+ CPU cores
  • 16+ GB RAM
  • 100 GB+ available disk space for building binaries and storing cached chain state
  • 100+ Mbps network connections

We continuously test that our builds and tests pass on:

The latest GitHub Runners for:

  • macOS
  • Ubuntu

Docker:

  • Debian Bullseye

Zebra's tests can take over an hour, depending on your machine. We're working on making them faster.

zebrad might build and run fine on smaller and slower systems - we haven't tested its exact limits yet.

For more detailed requirements, refer to the documentation.

Memory Troubleshooting

If Zebra's build runs out of RAM, try setting: export CARGO_BUILD_JOBS=2

If Zebra's tests timeout or run out of RAM, try running: cargo test -- --test-threads=2

(cargo uses all the processor cores on your machine by default.)

macOS Test Troubleshooting

Some of Zebra's tests deliberately cause errors that make Zebra panic. macOS records these panics as crash reports.

If you are seeing "Crash Reporter" dialogs during Zebra tests, you can disable them using this Terminal.app command:

defaults write com.apple.CrashReporter DialogType none

Network Ports and Data Usage

By default, Zebra uses the following inbound TCP listener ports:

  • 8233 on Mainnet
  • 18233 on Testnet

Zebra needs some peers which have a round-trip latency of 2 seconds or less. If this is a problem for you, please open a ticket.

zebrad's typical mainnet network usage is:

  • Initial sync: 40 GB download (in the longer term, several hundred GB are likely to be downloaded).
  • Ongoing updates: 10 MB - 1 GB upload and download per day, depending on user-created transaction size, and peer requests

Zebra also performs an initial sync every time its internal database version changes.

For more detailed information, refer to the documentation.

Network Troubleshooting

Some of Zebra's tests download Zcash blocks, so they might be unreliable depending on your network connection. You can set ZEBRA_SKIP_NETWORK_TESTS=1 to skip the network tests.

Zebra may be unreliable on Testnet, and under less-than-perfect network conditions. See our roadmap for details.

Disk Usage

Zebra uses up to 40 GB of space for cached mainnet data, and 10 GB of space for cached testnet data.

RocksDB cleans up outdated data periodically, and when the database is closed and re-opened.

Disk Troubleshooting

Zebra's state commits changes using RocksDB database transactions.

If you forcibly terminate Zebra, or it panics, any incomplete changes will be rolled back the next time it starts.

So Zebra's state should always be valid, unless your OS or disk hardware is corrupting data.

Known Issues

There are a few bugs in Zebra that we're still working on fixing:

  • No Windows support #3801
    • We used to test with Windows Server 2019, but not anymore; see issue for details

Performance

We are working on improving Zebra performance, the following are known issues:

  • Send note commitment and history trees from the non-finalized state to the finalized state #4824
  • Speed up opening the database #4822
  • Revert note commitment and history trees when forking non-finalized chains #4794
  • Store only the first tree state in each identical series of tree states #4784

RPCs might also be slower than they used to be, we need to check:

  • Revert deserializing state transactions in rayon threads #4831

Ongoing investigations:

  • Find out which parts of CommitBlock/CommitFinalizedBlock are slow #4823
  • Mini-Epic: Stop tokio tasks running for a long time and blocking other tasks #4747
  • Investigate busiest tasks per tokio-console #4583

Future Work

Features:

  • Wallet functionality

Performance and Reliability:

  • Reliable syncing under poor network conditions
  • Additional batch verification
  • Performance tuning

Currently, the following features are out of scope:

  • Mining support
  • Optional Zcash network protocol messages
  • Consensus rules removed before Canopy activation (Zebra checkpoints on Canopy activation)

Documentation

The Zebra website contains user documentation, such as how to run or configure Zebra, set up metrics integrations, etc., as well as developer documentation, such as design documents. We also render API documentation for the external API of our crates, as well as internal documentation for private APIs.

Security

Zebra has a responsible disclosure policy, which we encourage security researchers to follow.

License

Zebra is distributed under the terms of both the MIT license and the Apache License (Version 2.0).

See LICENSE-APACHE and LICENSE-MIT.

User Documentation

This section contains details on how to install, run, and instrument Zebra.

System Requirements

We usually build zebrad on systems with:

  • 2+ CPU cores
  • 7+ GB RAM
  • 14+ GB of disk space

On many-core machines (like, 32-core) the build is very fast; on 2-core machines it's less fast.

We continuously test that our builds and tests pass on:

  • macOS Big Sur 11.0
  • Ubuntu 18.04 / the latest LTS
  • Debian Buster

We usually run zebrad on systems with:

  • 4+ CPU cores
  • 16+ GB RAM
  • 50GB+ available disk space for finalized state
  • 100+ Mbps network connections

zebrad might build and run fine on smaller and slower systems - we haven't tested its exact limits yet.

Additional Features

Sentry Production Monitoring

Compile Zebra with --features sentry to monitor it using Sentry in production.

Lightwalletd Test Requirements

To test Zebra's lightwalletd RPC methods:

  • compile Zebra with the --features lightwalletd-grpc-tests
  • install a lightwalletd binary
    • Zebra's tests currently target adityapk00/lightwalletd
    • some tests might fail on other lightwalletd versions, due to differences in the logs
  • install the protoc Protobuf compiler:
    • the protobuf-compiler or protobuf package, or
    • cmake to automatically compile protoc in the zebrad build script
  • set the required test environmental variables:

Installing Zebra

zebrad is still under development, so there is no supported packaging or install mechanism. To run zebrad, follow the instructions to compile zebrad for your platform:

  1. Install cargo and rustc.
    • Using rustup installs the stable Rust toolchain, which zebrad targets.
  2. Install Zebra's build dependencies:
    • libclang: the libclang, libclang-dev, llvm, or llvm-dev packages, depending on your package manager
    • clang or another C++ compiler: g++, Xcode, or MSVC
  3. Run cargo install --locked --git https://github.com/ZcashFoundation/zebra --tag v1.0.0-beta.13 zebrad
  4. Run zebrad start (see Running Zebra for more information)

If you're interested in testing out zebrad please feel free, but keep in mind that there is a lot of key functionality still missing.

Build Troubleshooting

If you're having trouble with:

Dependencies:

  • use cargo install without --locked to build with the latest versions of each dependency
  • sqlite linker errors: libsqlite3 is an optional dependency of the zebra-network/tor feature. If you don't have it installed, you might see errors like note: /usr/bin/ld: cannot find -lsqlite3. Follow the arti instructions to install libsqlite3, or use one of these commands instead:
cargo build
cargo build -p zebrad --all-features

Compilers:

  • clang: install both libclang and clang - they are usually different packages
  • libclang: check out the clang-sys documentation
  • g++ or MSVC++: try using clang or Xcode instead
  • rustc: use rustc 1.58 or later
    • Zebra does not have a minimum supported Rust version (MSRV) policy yet

Dependencies

Zebra primarily depends on pure Rust crates, and some Rust/C++ crates:

They will be automatically built along with zebrad.

Running Zebra

zebrad generate generates a default config. These defaults will be used if no config is present, so it's not necessary to generate a config. However, having a config file with the default fields is a useful starting point for changing the config.

The configuration format is the TOML encoding of the internal config structure, and documentation for all of the config options can be found here.

  • zebrad start starts a full node.

Return Codes

  • 0: Application exited successfully
  • 1: Application exited unsuccessfully
  • 2: Application crashed
  • zebrad may also return platform-dependent codes.

Network Ports and Data Usage

zebrad's default ports and network usage are documented in the README.

If Zebra is configured with a specific listen_addr, it will advertise this address to other nodes for inbound connections.

Zebra makes outbound connections to peers on any port. But zcashd prefers peers on the default ports, so that it can't be used for DDoS attacks on other networks.

The major constraint we've found on zebrad performance is the network weather, especially the ability to make good connections to other Zcash network peers.

Zebra needs some peers which have a round-trip latency of 2 seconds or less. If this is a problem for you, please let us know!

Improving Performance

Zebra usually syncs in 4-12 hours, depending on the network connection.

If you're having trouble syncing, try the following config changes:

Release Build

Make sure you're using a release build on your native architecture.

If you're using an ARM machine, install the Rust compiler for ARM. If you build using the x86_64 tools, Zebra might run really slowly.

Run a release build using the cargo install command from the README.

Syncer Lookahead Limit

If your connection is slow, try downloading fewer blocks at a time:

[sync]
lookahead_limit = 1000
max_concurrent_block_requests = 25

Peer Set Size

If your connection is slow, try connecting to fewer peers:

[network]
peerset_initial_target_size = 25

Turn off debug logging

Zebra logs at info level by default.

If Zebra is slow, make sure it is logging at info level:

[tracing]
filter = 'info'

Or restrict debug logging to a specific Zebra component:

[tracing]
filter = 'info,zebra_network=debug'

If you keep on seeing multiple info logs per second, please open a bug.

Tracing Zebra

Dynamic Tracing

Zebra supports dynamic tracing, configured using the config's TracingSection and an HTTP RPC endpoint.

Activate this feature using the filter-reload compile-time feature, and the filter and endpoint_addr runtime config options.

If the endpoint_addr is specified, zebrad will open an HTTP endpoint allowing dynamic runtime configuration of the tracing filter. For instance, if the config had endpoint_addr = '127.0.0.1:3000', then

  • curl -X GET localhost:3000/filter retrieves the current filter string;
  • curl -X POST localhost:3000/filter -d "zebrad=trace" sets the current filter string.

See the filter documentation for more details.

journald Logging

Zebra can send tracing spans and events to systemd-journald, on Linux distributions that use systemd.

Activate journald logging using the journald compile-time feature, and the use_journald runtime config option.

Flamegraphs

Zebra can generate flamegraphs of tracing spans.

Activate flamegraphs using the flamegraph compile-time feature, and the flamegraph runtime config option.

Zebra Metrics

Zebra has support for Prometheus, configured using the prometheus compile-time feature, and the MetricsSection runtime configuration.

This requires supporting infrastructure to collect and visualize metrics, for example:

  1. Create the zebrad.toml file with the following contents:

    [metrics]
endpoint_addr = "127.0.0.1:9999"

  2. Run Zebra, and specify the path to the zebrad.toml file, for example:

    zebrad -c zebrad.toml start

  3. Install and run Prometheus and Grafana via Docker:

    # create a storage volume for grafana (once)
sudo docker volume create grafana-storage
# create a storage volume for prometheus (once)
sudo docker volume create prometheus-storage

# run prometheus with the included config
sudo docker run --detach --network host --volume prometheus-storage:/prometheus --volume /path/to/zebra/prometheus.yaml:/etc/prometheus/prometheus.yml  prom/prometheus

# run grafana
sudo docker run --detach --network host --env GF_SERVER_HTTP_PORT=3030 --env GF_SERVER_HTTP_ADDR=localhost --volume grafana-storage:/var/lib/grafana grafana/grafana

    Now the grafana dashboard is available at http://localhost:3030 ; the default username and password is admin/admin. Prometheus scrapes Zebra on localhost:9999, and provides the results on localhost:9090.

  4. Configure Grafana with a Prometheus HTTP Data Source, using Zebra's metrics.endpoint_addr.

    In the grafana dashboard:

    1. Create a new Prometheus Data Source Prometheus-Zebra
    2. Enter the HTTP URL: 127.0.0.1:9090
    3. Save the configuration

  5. Now you can add the grafana dashboards from zebra/grafana (Create > Import > Upload JSON File), or create your own.

Running lightwalletd with zebra

Zebra's RPC methods can support a lightwalletd service backed by zebrad.

Contents:

Download and build Zebra

cargo install --locked --git https://github.com/ZcashFoundation/zebra zebrad

Zebra binary will be at ~/.cargo/bin/zebrad.

Configure zebra for lightwalletd

We need a zebra configuration file. First, we create a file with the default settings:

zebrad generate -o ~/.config/zebrad.toml

The above command places the generated zebrad.toml config file in the default preferences directory of Linux. For other OSes default locations see here.

Tweak the following option in order to prepare for lightwalletd setup.

RPC section

This change is required for zebra to behave as an RPC endpoint. The standard port for RPC endpoint is 8232.

[rpc]
listen_addr = "127.0.0.1:8232"

Sync Zebra

With the configuration in place you can start synchronizing Zebra with the Zcash blockchain. This may take a while depending on your hardware.

zebrad start

Zebra will display information about sync process:

...
zebrad::commands::start: estimated progress to chain tip sync_percent=10.783 % 
...

Until eventually it will get there:

...
zebrad::commands::start: finished initial sync to chain tip, using gossiped blocks sync_percent=100.000 % 
...

You can interrupt the process at any time with ctrl-c and Zebra will resume the next time at around the block you were downloading when stopping the process.

When deploying for production infrastructure, the above command can/should be implemented as a server service or similar configuration.

For implementing zebra as a service please see here.

Download and build lightwalletd

While you synchronize Zebra you can install lightwalletd.

Before installing, you need to have go in place. Please visit the go install page with download and installation instructions.

With go installed and in your path, download and install lightwalletd:

git clone https://github.com/zcash/lightwalletd
cd lightwalletd
make
make install

If everything went good you should have a lightwalletd binary in ~/go/bin/.

Sync lightwalletd

Please make sure you have zebrad running (with RPC endpoint and up to date blockchain) to synchronize lightwalletd.

  • lightwalletd requires a zcash.conf file located somewhere, however this file can be empty if you are using the default zebra rpc endpoint (127.0.0.1:8232).

  • For production setups lightwalletd requires a cert.pem. For more information on how to do this please see here.

  • lightwalletd can run without the certificate (with the --no-tls-very-insecure flag) however this is not recommended for production environments.

With the cert in ./ and an empty zcash.conf we can start the sync with:

lightwalletd --zcash-conf-path ~/.config/zcash.conf --data-dir ~/.cache/lightwalletd --log-file /dev/stdout

By default lightwalletd service will listen on 127.0.0.1:9067

Lightwalletd will do its own synchronization, while it is doing you will see messages as:

...
{"app":"lightwalletd","level":"info","msg":"Ingestor adding block to cache: 748000","time":"2022-05-28T19:25:49-03:00"}
{"app":"lightwalletd","level":"info","msg":"Ingestor adding block to cache: 749540","time":"2022-05-28T19:25:53-03:00"}
{"app":"lightwalletd","level":"info","msg":"Ingestor adding block to cache: 751074","time":"2022-05-28T19:25:57-03:00"}
...

Wait until lightwalletd is in sync before connecting any wallet into it. You will know when it is in sync as those messages will not be displayed anymore.

Run tests

The Zebra team created tests for the interaction of zebra and lightwalletd.

Please refer to acceptance tests documentation in the Lightwalletd tests section.

Connect a wallet to lightwalletd

The final goal is to connect wallets to the lightwalletd service backed by Zebra.

For demo purposes we use zecwallet-cli.

Make sure both zebrad and lightwalletd are running and listening.

Download and build the cli-wallet

cargo install --locked --git https://github.com/adityapk00/zecwallet-light-cli

zecwallet-cli binary will be at ~/.cargo/bin/zecwallet-cli.

Run the wallet

$ zecwallet-cli --server 127.0.0.1:9067
Lightclient connecting to http://127.0.0.1:9067/
{
  "result": "success",
  "latest_block": 1683911,
  "total_blocks_synced": 49476
}
Ready!
(main) Block:1683911 (type 'help') >>

Developer Documentation

This section contains the contribution guide and design documentation. It does not contain API documentation, which is generated using Rustdoc:

Contributing

Running and Debugging

See the user documentation for details on how to build, run, and instrument Zebra.

Bug Reports

File an issue on the issue tracker using the bug report template.

Pull Requests

PRs are welcome for small and large changes, but please don't make large PRs without coordinating with us via the issue tracker or Discord. This helps increase development coordination and makes PRs easier to merge.

Check out the help wanted or good first issue labels if you're looking for a place to get started!

Zebra follows the conventional commits standard for the commits merged to main. Since PRs are squashed before merging to main, the PR titles should follow the conventional commits standard so that the merged commits are conformant.

Coverage Reports

Zebra's CI currently generates coverage reports for every PR with rust's new source based coverage feature. The coverage reports are generated by the coverage.yml file.

These reports are then saved as html and zipped up into a github action's artifact. These artifacts can be accessed on the checks tab of any PR, next to the "re-run jobs" button on the Coverage (+nightly) CI job's tab example.

To access a report download and extract the zip artifact then open the top level index.html.

Zebra RFCs

Significant changes to the Zebra codebase are planned using Zebra RFCs. These allow structured discussion about a proposed change and provide a record of the planned design.

To make a Zebra RFC:

  1. Choose a short feature name like my-feature.

  2. Copy the book/src/dev/rfcs/0000-template.md file to book/src/dev/rfcs/drafts/xxxx-my-feature.md.

  3. Edit the template header to add the feature name and the date, but leave the other fields blank for now.

  4. Write the design! The template has a suggested list of sections that are a useful guide.

  5. Create an design PR using the RFC template.

  6. After creating an RFC PR, update the RFC header and the PR description with the PR number.

  7. Make changes to the RFC in collaboration with the Zebra team.

  8. When the RFC is merged, take the next available RFC number (not conflicting with any existing RFCs or design PRs) and name the RFC file accordingly, e.g., 0027-my-feature.md for number 27.

  9. Make sure that book/src/SUMMARY.md links to the new number for the RFC.

  10. After the RFC is accepted, create an issue for the implementation of the design, and update the RFC header and PR description with the implementation issue number.

Design Overview

This document sketches the future design for Zebra.

Desiderata

The following are general desiderata for Zebra:

  • [George's list..]

  • As much as reasonably possible, it and its dependencies should be implemented in Rust. While it may not make sense to require this in every case (for instance, it probably doesn't make sense to rewrite libsecp256k1 in Rust, instead of using the same upstream library as Bitcoin), we should generally aim for it.

  • As much as reasonably possible, Zebra should minimize trust in required dependencies. Note that "minimize number of dependencies" is usually a proxy for this desideratum, but is not exactly the same: for instance, a collection of crates like the tokio crates are all developed together and have one trust boundary.

  • Zebra should be well-factored internally into a collection of component libraries which can be used by other applications to perform Zcash-related tasks. Implementation details of each component should not leak into all other components.

  • Zebra should checkpoint on Canopy activation and drop all Sprout-related functionality not required post-Canopy.

Non-Goals

  • Zebra keeps a copy of the chain state, so it isn't intended for lightweight applications like light wallets. Those applications should use a light client protocol.

Architecture

Unlike zcashd, which originated as a Bitcoin Core fork and inherited its monolithic architecture, Zebra has a modular, library-first design, with the intent that each component can be independently reused outside of the zebrad full node. For instance, the zebra-network crate containing the network stack can also be used to implement anonymous transaction relay, network crawlers, or other functionality, without requiring a full node.

At a high level, the fullnode functionality required by zebrad is factored into several components:

  • zebra-chain, providing definitions of core data structures for Zcash, such as blocks, transactions, addresses, etc., and related functionality. It also contains the implementation of the consensus-critical serialization formats used in Zcash. The data structures in zebra-chain are defined to enforce structural validity by making invalid states unrepresentable. For instance, the Transaction enum has variants for each transaction version, and it's impossible to construct a transaction with, e.g., spend or output descriptions but no binding signature, or, e.g., a version 2 (Sprout) transaction with Sapling proofs. Currently, zebra-chain is oriented towards verifying transactions, but will be extended to support creating them in the future.

  • zebra-network, providing an asynchronous, multithreaded implementation of the Zcash network protocol inherited from Bitcoin. In contrast to zcashd, each peer connection has a separate state machine, and the crate translates the external network protocol into a stateless, request/response-oriented protocol for internal use. The crate provides two interfaces:

    • an auto-managed connection pool that load-balances local node requests over available peers, and sends peer requests to a local inbound service, and
    • a connect_isolated method that produces a peer connection completely isolated from all other node state. This can be used, for instance, to safely relay data over Tor, without revealing distinguishing information.

  • zebra-script provides script validation. Currently, this is implemented by linking to the C++ script verification code from zcashd, but in the future we may implement a pure-Rust script implementation.

  • zebra-consensus performs semantic validation of blocks and transactions: all consensus rules that can be checked independently of the chain state, such as verification of signatures, proofs, and scripts. Internally, the library uses tower-batch to perform automatic, transparent batch processing of contemporaneous verification requests.

  • zebra-state is responsible for storing, updating, and querying the chain state. The state service is responsible for contextual verification: all consensus rules that check whether a new block is a valid extension of an existing chain, such as updating the nullifier set or checking that transaction inputs remain unspent.

  • zebrad contains the full node, which connects these components together and implements logic to handle inbound requests from peers and the chain sync process.

  • zebra-rpc and zebra-client will eventually contain the RPC and wallet functionality, but as mentioned above, our goal is to implement replication of chain state first before asking users to entrust Zebra with their funds.

All of these components can be reused as independent libraries, and all communication between stateful components is handled internally by internal asynchronous RPC abstraction ("microservices in one process").

zebra-chain

Internal Dependencies

None: these are the core data structure definitions.

Responsible for

  • definitions of commonly used data structures, e.g.,

    • Block,
    • Transaction,
    • Address,
    • KeyPair...

  • parsing bytes into these data structures

  • definitions of core traits, e.g.,

    • ZcashSerialize and ZcashDeserialize, which perform consensus-critical serialization logic.

Exported types

  • [...]

zebra-network

Internal Dependencies

  • zebra-chain

Responsible for

  • definition of a well structured, internal request/response protocol
  • provides an abstraction for "this node" and "the network" using the internal protocol
  • dynamic, backpressure-driven peer set management
  • per-peer state machine that translates the internal protocol to the Bitcoin/Zcash protocol
  • tokio codec for Bitcoin/Zcash message encoding.

Exported types

  • Request, an enum representing all possible requests in the internal protocol;
  • Response, an enum representing all possible responses in the internal protocol;
  • AddressBook, a data structure for storing peer addresses;
  • Config, a configuration object for all networking-related parameters;
  • init<S: Service>(Config, S) -> (impl Service, Arc<Mutex<AddressBook>>), the main entry-point.

The init entrypoint constructs a dynamically-sized pool of peers sending inbound requests to the provided S: tower::Service representing "this node", and returns a Service that can be used to send requests to "the network", together with an AddressBook updated with liveness information from the peer pool. The AddressBook can be used to respond to inbound requests for peers.

All peerset management (finding new peers, creating new outbound connections, etc) is completely encapsulated, as is responsibility for routing outbound requests to appropriate peers.

zebra-state

Internal Dependencies

  • zebra-chain for data structure definitions.

Responsible for

  • block storage API
    • operates on parsed block structs
      • these structs can be converted from and into raw bytes
    • primarily aimed at network replication, not at processing
    • can be used to rebuild the database below
  • maintaining a database of tx, address, etc data
    • this database can be blown away and rebuilt from the blocks, which are otherwise unused.
    • threadsafe, typed lookup API that completely encapsulates the database logic
    • handles stuff like "transactions are reference counted by outputs" etc.
  • providing tower::Service interfaces for all of the above to support backpressure.

Exported types

  • Request, an enum representing all possible requests in the internal protocol;
    • blocks can be accessed via their chain height or hash
    • confirmed transactions can be accessed via their block, or directly via their hash
  • Response, an enum representing all possible responses in the internal protocol;
  • init() -> impl Service, the main entry-point.

The init entrypoint returns a Service that can be used to send requests for the chain state.

All state management (adding blocks, getting blocks by index or hash) is completely encapsulated.

zebra-script

Internal Dependencies

  • ??? depends on how it's implemented internally

Responsible for

  • the minimal Bitcoin script implementation required for Zcash
  • script parsing
  • context-free script validation

Notes

This can wrap an existing script implementation at the beginning.

If this existed in a "good" way, we could use it to implement tooling for Zcash script inspection, debugging, etc.

Questions

  • How does this interact with NU4 script changes?

Exported types

  • [...]

zebra-consensus

Internal Dependencies

  • zebra-chain for data structures and parsing.
  • zebra-state to read and update the state database.
  • zebra-script for script parsing and validation.

Responsible for

  • consensus-specific parameters (network magics, genesis block, pow parameters, etc) that determine the network consensus
  • consensus logic to decide which block is the current block
  • block and transaction verification
    • context-free validation, e.g., signature, proof verification, etc.
    • context-dependent validation, e.g., determining whether a transaction is accepted in a particular chain state context.
    • verifying mempool (unconfirmed) transactions
  • block checkpoints
    • mandatory checkpoints (genesis block, canopy activation)
    • optional regular checkpoints (every Nth block)
  • modifying the chain state
    • adding new blocks to ZebraState, including chain reorganisation
    • adding new transactions to ZebraMempoolState
  • storing the transaction mempool state
    • mempool transactions can be accessed via their hash
  • providing tower::Service interfaces for all of the above to support backpressure and batch validation.

Exported types

  • block::init() -> impl Service, the main entry-point for block verification.
  • ZebraMempoolState
    • all state management (adding transactions, getting transactions by hash) is completely encapsulated.
  • mempool::init() -> impl Service, the main entry-point for mempool transaction verification.

The init entrypoints return Services that can be used to verify blocks or transactions, and add them to the relevant state.

zebra-rpc

Internal Dependencies

  • zebra-chain for data structure definitions
  • zebra-network possibly? for definitions of network messages?

Responsible for

  • rpc interface

Exported types

  • [...]

zebra-client

Internal Dependencies

  • zebra-chain for structure definitions
  • zebra-state for transaction queries and client/wallet state storage
  • zebra-script possibly? for constructing transactions

Responsible for

  • implementation of some event a user might trigger
  • would be used to implement a full wallet
  • create transactions, monitors shielded wallet state, etc.

Notes

Communication between the client code and the rest of the node should be done by a tower service interface. Since the Service trait can abstract from a function call to RPC, this means that it will be possible for us to isolate all client code to a subprocess.

Exported types

  • [...]

zebrad

Abscissa-based application which loads configs, all application components, and connects them to each other.

Responsible for

  • actually running the server
  • connecting functionality in dependencies

Internal Dependencies

  • zebra-chain
  • zebra-network
  • zebra-state
  • zebra-consensus
  • zebra-client
  • zebra-rpc

Unassigned functionality

Responsibility for this functionality needs to be assigned to one of the modules above (subject to discussion):

  • [ ... add to this list ... ]

Zebra RFCs

We are experimenting with using a process similar to the Rust RFC process to document design decisions for Zebra.

Summary

The Bitcoin network protocol used by Zcash allows nodes to download blocks from other peers. This RFC describes how we find and download this data asynchronously.

Motivation

To sync the chain, we need to find out which blocks to download and then download them. Downloaded blocks can then be fed into the verification system and (assuming they verify correctly) into the state system. In zcashd, blocks are processed one at a time. In Zebra, however, we want to be able to pipeline block download and verification operations, using futures to explicitly specify logical dependencies between sub-tasks, which we execute concurrently and potentially out-of-order on a threadpool. This means that the procedure we use to determine which blocks to download must look somewhat different than zcashd.

Block fetching in Bitcoin

Zcash inherits its network protocol from Bitcoin. Bitcoin block fetching works roughly as follows. A node can request block information from peers using either a getblocks or getheaders message. Both of these messages contain a block locator object consisting of a sequence of block hashes. The block hashes are ordered from highest to lowest, and represent checkpoints along the path from the node's current tip back to genesis. The remote peer computes the intersection between its chain and the node's chain by scanning through the block locator for the first hash in its chain. Then, it sends (up to) 500 subsequent block hashes in an inv message (in the case of getblocks) or (up to) 2000 block headers in a headers message (in the case of getheaders). Note: zcashd reduces the getheaders count to 160, because Zcash headers are much larger than Bitcoin headers, as noted below.

The headers message sent after getheaders contains the actual block headers, while the inv message sent after getblocks contains only hashes, which have to be fetched with a getdata message. In Bitcoin, the block headers are small relative to the size of the full block, but this is not always the case for Zcash, where the block headers are much larger due to the use of Equihash and many blocks have only a few transactions. Also, getblocks allows parallelizing block downloads, while getheaders doesn't. For these reasons and because we know we need full blocks anyways, we should probably use getblocks.

The getblocks Bitcoin message corresponds to our zebra_network::Request::FindBlocksByHash, and the getdata message is generated by zebra_network::Request::Blocks.

Pipelining block verification

As mentioned above, our goal is to be able to pipeline block download and verification. This means that the process for block lookup should ideally attempt to fetch and begin verification of future blocks without blocking on complete verification of all earlier blocks. To do this, we split the chain state into the verified block chain (held by the state component) and the prospective block chain (held only by the syncer), and use the following algorithm to pursue prospective chain tips.

ObtainTips

  1. Query the current state to construct the sequence of hashes
[tip, tip-1, tip-2, ..., tip-9, tip-20, tip-40, tip-80, tip-160 ]

The precise construction is unimportant, but this should have a Bitcoin-style dense-first, then-sparse hash structure.

The initial state should contain the genesis block for the relevant network. So the sequence of hashes will only contain the genesis block

[genesis ]

The network will respond with a list of hashes, starting at the child of the genesis block.

  1. Make a FindBlocksByHash request to the network F times, where F is a fanout parameter, to get resp1, ..., respF.

  2. For each response, starting from the beginning of the list, prune any block hashes already included in the state, stopping at the first unknown hash to get resp1', ..., respF'. (These lists may be empty).

  3. Combine the last elements of each list into a set; this is the set of prospective tips.

  4. Combine all elements of each list into a set, and queue download and verification of those blocks.

  5. If there are any prospective tips, call ExtendTips, which returns a new set of prospective tips. Continue calling ExtendTips with this new set, until there are no more prospective tips.

  6. Restart after some delay, say 15 seconds.

ExtendTips

  1. Remove all prospective tips from the set of prospective tips, then iterate through them. For each removed tip:

  2. Create a FindBlocksByHash request consisting of just the prospective tip. Send this request to the network F times.

  3. For each response, check whether the first hash in the response is a genesis block (for either the main or test network). If so, discard the response. It indicates that the remote peer does not have any blocks following the prospective tip. (Or that the remote peer is on the wrong network.)

  4. Combine the last elements of the remaining responses into a set, and add this set to the set of prospective tips.

  5. Combine all elements of the remaining responses into a set, and queue download and verification of those blocks.

DoS resistance

Because this strategy aggressively downloads any available blocks, it could be vulnerable to a DoS attack, where a malicious peer feeds us bogus chain tips, causing us to waste network and CPU on blocks that will never be valid. However, because we separate block finding from block downloading, and because of the design of our network stack, this attack is probably not feasible. The primary reason is that zebra_network randomly loadbalances outbound requests over all available peers.

Consider a malicious peer who responds to block discovery with a bogus list of hashes. We will eagerly attempt to download all of those bogus blocks, but our requests to do so will be randomly load-balanced to other peers, who are unlikely to know about the bogus blocks. When we try to extend a bogus tip, the extension request will also be randomly load-balanced, so it will likely be routed to a peer that doesn't know about it and can't extend it. And because we perform multiple block discovery queries, which will also be randomly load balanced, we're unlikely to get stuck on a false chain tip.

Fork-finding

When starting from a verified chain tip, the choice of block locator can find forks at least up to the reorg limit (99 blocks). When extending a prospective tip, forks are ignored, but this is fine, since unless we are prefetching the longest chain, we won't be able to keep extending the tip prospectively.

Retries and Fanout

We should consider the fanout parameter F and the retry policy for the different requests. I'm not sure whether we need to retry requests to discover new block hashes, since the fanout may already provide redundancy. For the block requests themselves, we should have a retry policy with a limited number of attempts, enough to insulate against network failures but not so many that we would retry a bogus block indefinitely. Maybe fanout 4 and 3 retries?

Parallel Verification

Summary

Zebra verifies blocks in several stages, most of which can be executed in parallel.

We use several different design patterns to enable this parallelism:

  • We download blocks and start verifying them in parallel,
  • We batch signature and proof verification using verification services, and
  • We defer data dependencies until just before the block is committed to the state (see the detailed design RFCs).

Motivation

Zcash (and Bitcoin) are designed to verify each block in sequence, starting from the genesis block. But during the initial sync, and when restarting with an older state, this process can be quite slow.

By deferring data dependencies, we can partially verify multiple blocks in parallel.

By parallelising block and transaction verification, we can use multithreading and batch verification for signatures, proofs, scripts, and hashes.

Definitions

Blockchain:

  • chain fork: Zcash is implemented using a tree of blocks. Each block has a single previous block, and zero to many next blocks. A chain fork consists of a tip and all its previous blocks, back to the genesis block.
  • genesis: The root of the tree of blocks is called the genesis block. It has no previous block.
  • tip: A block which has no next block is called a tip. Each chain fork can be identified using its tip.

Data:

  • consensus rule: A protocol rule which all nodes must apply consistently, so they can converge on the same chain fork.
  • context-free: Consensus rules which do not have a data dependency on previous blocks.
  • data dependency: Information contained in the previous block and its chain fork, which is required to verify the current block.
  • state: The set of verified blocks. The state might also cache some dependent data, so that we can efficiently verify subsequent blocks.

Verification Stages:

  • structural verification: Parsing raw bytes into the data structures defined by the protocol.
  • semantic verification: Verifying the consensus rules on the data structures defined by the protocol.
  • contextual verification: Verifying the current block, once its data dependencies have been satisfied by a verified previous block. This verification might also use the cached state corresponding to the previous block.

Guide-level explanation

In Zebra, we want to verify blocks in parallel. Some fields can be verified straight away, because they don't depend on the output of previous blocks. But other fields have data dependencies, which means that we need previous blocks before we can fully validate them.

If we delay checking some of these data dependencies, then we can do more of the verification in parallel.

Example: BlockHeight

Here's how Zebra can verify the different Block Height consensus rules in parallel:

Structural Verification:

  1. Parse the Block into a BlockHeader and a list of transactions.

Semantic Verification: No Data Dependencies:

  1. Check that the first input of the first transaction in the block is a coinbase input with a valid block height in its data field.

Semantic Verification: Deferring a Data Dependency:

  1. Verify other consensus rules that depend on Block Height, assuming that the Block Height is correct. For example, many consensus rules depend on the current Network Upgrade, which is determined by the Block Height. We verify these consensus rules, assuming the Block Height and Network Upgrade are correct.

Contextual Verification:

  1. Submit the block to the state for contextual verification. When it is ready to be committed (it may arrive before the previous block), check all deferred constraints, including the constraint that the block height of this block is one more than the block height of its parent block. If all constraints are satisfied, commit the block to the state. Otherwise, reject the block as invalid.

Zebra Design

Design Patterns

When designing changes to Zebra verification, use these design patterns:

  • perform context-free verification as soon as possible, (that is, verification which has no data dependencies on previous blocks),
  • defer data dependencies as long as possible, then
  • check the data dependencies.

Minimise Deferred Data

Keep the data dependencies and checks as simple as possible.

For example, Zebra could defer checking both the Block Height and Network Upgrade.

But since the Network Upgrade depends on the Block Height, we only need to defer the Block Height check. Then we can use all the fields that depend on the Block Height, as if it is correct. If the final Block Height check fails, we will reject the entire block, including all the verification we performed using the assumed Network Upgrade.

Implementation Strategy

When implementing these designs, perform as much verification as possible, await any dependencies, then perform the necessary checks.

Reference-level explanation

Verification Stages

In Zebra, verification occurs in the following stages:

  • Structural Verification: Raw block data is parsed into a block header and transactions. Invalid data is not representable in these structures: deserialization (parsing) can fail, but serialization always succeeds.
  • Semantic Verification: Parsed block fields are verified, based on their data dependencies:
    • Context-free fields have no data dependencies, so they can be verified as needed.
    • Fields with simple data dependencies defer that dependency as long as possible, so they can perform more verification in parallel. Then they await the required data, which is typically the previous block. (And potentially older blocks in its chain fork.)
    • Fields with complex data dependencies require their own parallel verification designs. These designs are out of scope for this RFC.
  • Contextual Verification: After a block is verified, it is added to the state. The details of state updates, and their interaction with semantic verification, are out of scope for this RFC.

This RFC focuses on Semantic Verification, and the design patterns that enable blocks to be verified in parallel.

Verification Interfaces

Verification is implemented by the following traits and services:

  • Structural Verification:
    • zebra_chain::ZcashDeserialize: A trait for parsing consensus-critical data structures from a byte buffer.
  • Semantic Verification:
    • ChainVerifier: Provides a verifier service that accepts a Block request, performs verification on the block, and responds with a block::Hash on success.
    • Internally, the ChainVerifier selects between a CheckpointVerifier for blocks that are within the checkpoint range, and a BlockVerifier for recent blocks.
  • Contextual Verification:
    • zebra_state::init: Provides the state update service, which accepts requests to add blocks to the state.

Checkpoint Verification

The CheckpointVerifier performs rapid verification of blocks, based on a set of hard-coded checkpoints. Each checkpoint hash can be used to verify all the

previous blocks, back to the genesis block. So Zebra can skip almost all verification for blocks in the checkpoint range.

The CheckpointVerifier uses an internal queue to store pending blocks. Checkpoint verification is cheap, so it is implemented using non-async functions within the CheckpointVerifier service.

Here is how the CheckpointVerifier implements each verification stage:

  • Structural Verification:
    • As Above: the CheckpointVerifier accepts parsed Block structs.
  • Semantic Verification:
    • check_height: makes sure the block height is within the unverified checkpoint range, and adds the block to its internal queue.
    • target_checkpoint_height: Checks for a continuous range of blocks from the previous checkpoint to a subsequent checkpoint. If the chain is incomplete, returns a future, and waits for more blocks. If the chain is complete, assumes that the previous_block_hash fields of these blocks form an unbroken chain from checkpoint to checkpoint, and starts processing the checkpoint range. (This constraint is an implicit part of the CheckpointVerifier design.)
    • process_checkpoint_range: makes sure that the blocks in the checkpoint range have an unbroken chain of previous block hashes.
  • Contextual Verification:
    • As Above: the CheckpointVerifier returns success to the ChainVerifier, which sends verified Blocks to the state service.

Block Verification

The BlockVerifier performs detailed verification of recent blocks, in parallel.

Here is how the BlockVerifier implements each verification stage:

  • Structural Verification:
    • As Above: the BlockVerifier accepts parsed Block structs.
  • Semantic Verification:
    • As Above: verifies each field in the block. Defers any data dependencies as long as possible, awaits those data dependencies, then performs data dependent checks.
    • Note: Since futures are executed concurrently, we can use the same function to:
      • perform context-free verification,
      • perform verification with deferred data dependencies,
      • await data dependencies, and
      • check data dependencies. To maximise concurrency, we should write verification functions in this specific order, so the awaits are as late as possible.
  • Contextual Verification:
    • As Above: the BlockVerifier returns success to the ChainVerifier, which sends verified Blocks to the state service.

Zcash Protocol Design

When designing a change to the Zcash protocol, minimise the data dependencies between blocks.

Try to create designs that:

  • Eliminate data dependencies,
  • Make the changes depend on a version field in the block header or transaction,
  • Make the changes depend on the current Network Upgrade, or
  • Make the changes depend on a field in the current block, with an additional consensus rule to check that field against previous blocks.

When making decisions about these design tradeoffs, consider:

  • how the data dependency could be deferred, and
  • the CPU cost of the verification - if it is trivial, then it does not matter if the verification is parallelised.

Drawbacks

This design is a bit complicated, but we think it's necessary to achieve our goals.

Rationale and alternatives

  • What makes this design a good design?
    • It enables a significant amount of parallelism
    • It is simpler than some other alternatives
    • It uses existing Rust language facilities, mainly Futures and await/async
  • Is this design a good basis for later designs or implementations?
    • We have built a UTXO design on this design
    • We believe we can build "recent blocks" and "chain summary" designs on this design
    • Each specific detailed design will need to consider how the relevant data dependencies are persisted
  • What other designs have been considered and what is the rationale for not choosing them?
    • Serial verification
      • Effectively single-threaded
    • Awaiting data dependencies as soon as they are needed
      • Less parallelism
    • Providing direct access to the state
      • Might cause data races, might be prevented by Rust's ownership rules
      • Higher risk of bugs
  • What is the impact of not doing this?
    • Verification is slow, we can't batch or parallelise some parts of the verification

Prior art

TODO: expand this section

  • zcashd
    • serial block verification
    • Zebra implements the same consensus rules, but a different design
  • tower

Unresolved questions

  • Is this design good enough to use as a framework for future RFCs?
  • Does this design require any changes to the current implementation?
    • Implement block height consensus rule (check previous block hash and height)
    • Check that the BlockVerifier performs checks in the following order:
      • verification, deferring dependencies as needed,
      • await dependencies,
      • check deferred data dependencies

Out of Scope:

  • What is the most efficient design for parallel verification?

    • (Optimisations are out of scope.)

  • How is each specific field verified?

  • How do we verify fields with complex data dependencies?

  • How does verification change with different network upgrades?

  • How do multiple chains work, in detail?

  • How do state updates work, in detail?

  • Moving the verifiers into the state service

Future possibilities

  • Separate RFCs for other data dependencies
    • Recent blocks
    • Overall chain summaries (for example, total work)
    • Reorganisation limit: multiple chains to single chain transition
  • Optimisations for parallel verification

Summary

The Bitcoin network protocol used by Zcash allows nodes to advertise data (inventory items) for download by other peers. This RFC describes how we track and use this information.

Motivation

In order to participate in the network, we need to be able to fetch new data that our peers notify us about. Because our network stack abstracts away individual peer connections, and load-balances over available peers, we need a way to direct requests for new inventory only to peers that advertised to us that they have it.

Definitions

  • Inventory item: either a block or transaction.
  • Inventory hash: the hash of an inventory item, represented by the InventoryHash type.
  • Inventory advertisement: a notification from another peer that they have some inventory item.
  • Inventory request: a request to another peer for an inventory item.

Guide-level explanation

The Bitcoin network protocol used by Zcash provides a mechanism for nodes to gossip blockchain data to each other. This mechanism is used to distribute (mined) blocks and (unmined) transactions through the network. Nodes can advertise data available in their inventory by sending an inv message containing the hashes and types of those data items. After receiving an inv message advertising data, a node can determine whether to download it.

This poses a challenge for our network stack, which goes to some effort to abstract away details of individual peers and encapsulate all peer connections behind a single request/response interface representing "the network". Currently, the peer set tracks readiness of all live peers, reports readiness if at least one peer is ready, and routes requests across ready peers randomly using the "power of two choices" algorithm.

However, while this works well for data that is already distributed across the network (e.g., existing blocks) it will not work well for fetching data during distribution across the network. If a peer informs us of some new data, and we attempt to download it from a random, unrelated peer, we will likely fail. Instead, we track recent inventory advertisements, and make a best-effort attempt to route requests to peers who advertised that inventory.

Reference-level explanation

The inventory tracking system has several components:

  1. A registration hook that monitors incoming messages for inventory advertisements;
  2. An inventory registry that tracks inventory presence by peer;
  3. Routing logic that uses the inventory registry to appropriately route requests.

The first two components have fairly straightforward design decisions, but the third has considerably less obvious choices and tradeoffs.

Inventory Monitoring

Zebra uses Tokio's codec mechanism to transform a byte-oriented I/O interface into a Stream and Sink for incoming and outgoing messages. These are passed to the peer connection state machine, which is written generically over any Stream and Sink. This construction makes it easy to "tap" the sequence of incoming messages using .then and .with stream and sink combinators.

We already do this to record Prometheus metrics on message rates as well as to report message timestamps used for liveness checks and last-seen address book metadata. The message timestamp mechanism is a good example to copy. The handshake logic instruments the incoming message stream with a closure that captures a sender handle for a mpsc channel with a large buffer (currently 100 timestamp entries). The receiver handle is owned by a separate task that shares an Arc<Mutex<AddressBook>> with other parts of the application. This task waits for new timestamp entries, acquires a lock on the address book, and updates the address book. This ensures that timestamp updates are queued asynchronously, without lock contention.

Unlike the address book, we don't need to share the inventory data with other parts of the application, so it can be owned exclusively by the peer set. This means that no lock is necessary, and the peer set can process advertisements in its poll_ready implementation. This method may be called infrequently, which could cause the channel to fill. However, because inventory advertisements are time-limited, in the sense that they're only useful before some item is fully distributed across the network, it's safe to handle excess entries by dropping them. This behavior is provided by a broadcast/mpmc channel, which can be used in place of an mpsc channel.

An inventory advertisement is an (InventoryHash, SocketAddr) pair. The stream hook should check whether an incoming message is an inv message with only a small number (e.g., 1) inventory entries. If so, it should extract the hash for each item and send it through the channel. Otherwise, it should ignore the message contents. Why? Because inv messages are also sent in response to queries, such as when we request subsequent block hashes, and in that case we want to assume that the inventory is generally available rather than restricting downloads to a single peer. However, items are usually gossiped individually (or potentially in small chunks; zcashd has an internal inv buffer subject to race conditions), so choosing a small bound such as 1 is likely to work as a heuristic for when we should assume that advertised inventory is not yet generally available.

Inventory Registry

The peer set's poll_ready implementation should extract all available (InventoryHash, SocketAddr) pairs from the channel, and log a warning event if the receiver is lagging. The channel should be configured with a generous buffer size (such as 100) so that this is unlikely to happen in normal circumstances. These pairs should be fed into an InventoryRegistry structure along these lines:


#![allow(unused)]
fn main() {
struct InventoryRegistry{
    current: HashMap<InventoryHash, HashSet<SocketAddr>>,
    prev: HashMap<InventoryHash, HashSet<SocketAddr>>,
}

impl InventoryRegistry {
    pub fn register(&mut self, item: InventoryHash, addr: SocketAddr) {
        self.0.entry(item).or_insert(HashSet::new).insert(addr);
    }

    pub fn rotate(&mut self) {
        self.prev = std::mem::take(self.current)
    }

    pub fn peers(&self, item: InventoryHash) -> impl Iterator<Item=&SocketAddr> {
        self.prev.get(item).chain(self.current.get(item)).flatten()
    }
}
}

This API allows pruning the inventory registry using rotate, which implements generational pruning of registry entries. The peer set should maintain a tokio::time::Interval with some interval parameter, and check in poll_ready whether the interval stream has any items, calling rotate for each one:


#![allow(unused)]
fn main() {
while let Poll::Ready(Some(_)) = timer.poll_next(cx) {
    registry.rotate();
}
}

By rotating for each available item in the interval stream, rather than just once, we ensure that if the peer set's poll_ready is not called for a long time, rotate will be called enough times to correctly flush old entries.

Inventory advertisements live in the registry for twice the length of the timer, so it should be chosen to be half of the desired lifetime for inventory advertisements. Setting the timer to 75 seconds, the block interval, seems like a reasonable choice.

Routing Logic

At this point, the peer set has information on recent inventory advertisements. However, the Service trait only allows poll_ready to report readiness based on the service's data and the type of the request, not the content of the request. This means that we must report readiness without knowing whether the request should be routed to a specific peer, and we must handle the case where call gets a request for an item only available at an unready peer.

This RFC suggests the following routing logic. First, check whether the request fetches data by hash. If so, and peers() returns Some(ref addrs), iterate over addrs and route the request to the first ready peer if there is one. In all other cases, fall back to p2c routing. Alternatives are suggested and discussed below.

Rationale and alternatives

The rationale is described above. The alternative choices are primarily around the routing logic.

Because the Service trait does not allow applying backpressure based on the content of a request, only based on the service's internal data (via the &mut self parameter of Service::poll_ready) and on the type of the request (which determines which impl Service is used). This means that it is impossible for us to apply backpressure until a service that can process a specific inventory request is ready, because until we get the request, we can't determine which peers might be required to process it.

We could attempt to ensure that the peer set would be ready to process a specific inventory request would be to pre-emptively "reserve" a peer as soon as it advertises an inventory item. But this doesn't actually work to ensure readiness, because a peer could advertise two inventory items, and only be able to service one request at a time. It also potentially locks the peer set, since if there are only a few peers and they all advertise inventory, the service can't process any other requests. So this approach does not work.

Another alternative would be to do some kind of buffering of inventory requests that cannot immediately be processed by a peer that advertised that inventory. There are two basic sub-approaches here.

In the first case, we could maintain an unbounded queue of yet-to-be processed inventory requests in the peer set, and every time poll_ready is called, we check whether a service that could serve those inventory requests became ready, and start processing the request if we can. This would provide the lowest latency, because we can dispatch the request to the first available peer. For instance, if peer A advertises inventory I, the peer set gets an inventory request for I, peer A is busy so the request is queued, and peer B advertises inventory I, we could dispatch the queued request to B rather than waiting for A.

However, it's not clear exactly how we'd implement this, because this mechanism is driven by calls to poll_ready, and those might not happen. So we'd need some separate task that would drive processing the buffered task to completion, but this may not be able to do so by poll_ready, since that method requires owning the service, and the peer set will be owned by a Buffer worker.

In the second case, we could select an unready peer that advertised the requested inventory, clone it, and move the cloned peer into a task that would wait for that peer to become ready and then make the request. This is conceptually much cleaner than the above mechanism, but it has the downside that we don't dispatch the request to the first ready peer. In the example above, if we cloned peer A and dispatched the request to it, we'd have to wait for A to become ready, even if the second peer B advertised the same inventory just after we dispatched the request to A. However, this is not presently possible anyways, because the peer::Clients that handle requests are not clonable. They could be made clonable (they send messages to the connection state machine over a mpsc channel), but we cannot make this change without altering our liveness mechanism, which uses bounds on the time-since-last-message to determine whether a peer connection is live and to prevent immediate reconnections to recently disconnected peers.

A final alternative would be to fail inventory requests that we cannot route to a peer which advertised that inventory. This moves the failure forward in time, but preemptively fails some cases where the request might succeed -- for instance, if the peer has inventory but just didn't tell us, or received the inventory between when we dispatch the request and when it receives our message. It seems preferable to try and fail than to not try at all.

In practice, we're likely to care about the gossip protocol and inventory fetching once we've already synced close to the chain tip. In this setting, we're likely to already have peer connections, and we're unlikely to be saturating our peer set with requests (as we do during initial block sync). This suggests that the common case is one where we have many idle peers, and that therefore we are unlikely to have dispatched any recent requests to the peer that advertised inventory. So our common case should be one where all of this analysis is irrelevant.

Summary

This RFC describes an architecture for asynchronous script verification and its interaction with the state layer. This architecture imposes constraints on the ordering of operations in the state layer.

Motivation

As in the rest of Zebra, we want to express our work as a collection of work-items with explicit dependencies, then execute these items concurrently and in parallel on a thread pool.

Definitions

  • UTXO: unspent transparent transaction output. Transparent transaction outputs are modeled in zebra-chain by the transparent::Output structure.
  • outpoint: a reference to an unspent transparent transaction output, including a transaction hash and output index. Outpoints are modeled in zebra-chain by the transparent::OutPoint structure.
  • transparent input: a previous transparent output consumed by a later transaction (the one it is an input to). Modeled in zebra-chain by the transparent::Input::PrevOut enum variant.
  • coinbase transaction: the first transaction in each block, which creates new coins.
  • lock script: the script that defines the conditions under which some UTXO can be spent. Stored in the transparent::Output::lock_script field.
  • unlock script: a script satisfying the conditions of the lock script, allowing a UTXO to be spent. Stored in the transparent::Input::PrevOut::lock_script field.

Guide-level explanation

Zcash's transparent address system is inherited from Bitcoin. Transactions spend unspent transparent transaction outputs (UTXOs) from previous transactions. These UTXOs are encumbered by locking scripts that define the conditions under which they can be spent, e.g., requiring a signature from a certain key. Transactions wishing to spend UTXOs supply an unlocking script that should satisfy the conditions of the locking script for each input they wish to spend.

This means that script verification requires access to data about previous UTXOs, in order to determine the conditions under which those UTXOs can be spent. In Zebra, we aim to run operations asynchronously and out-of-order to the greatest extent possible. For instance, we may begin verification of a block before all of its ancestors have been verified or even downloaded. So we need to design a mechanism that allows script verification to declare its data dependencies and execute as soon as all required data is available.

It's not necessary for this mechanism to ensure that the transaction outputs remain unspent, only to give enough information to perform script verification. Checking that all transaction inputs are actually unspent is done later, at the point that its containing block is committed to the chain.

At a high level, this adds a new request/response pair to the state service:

  • Request::AwaitSpendableUtxo { output: OutPoint, ..conditions } requests a spendable transparent::Output, looked up using OutPoint.
  • Response::SpendableUtxo(Utxo) supplies the requested transparent::Output as part of a new Utxo type, if the output is spendable based on conditions;

Note that this request is named differently from the other requests, AwaitSpendableUtxo rather than GetUtxo or similar. This is because the request has rather different behavior:

  • the request does not complete until the state service learns about a UTXO matching the request, which could be never. For instance, if the transaction output was already spent, the service is not required to return a response.
  • the request does not complete until the output is spendable, based on the conditions in the request.

The state service does not cancel long-running UTXO requests. Instead, the caller is responsible for deciding when a request is unlikely to complete. (For example, using a timeout layer.)

This allows a script verifier to asynchronously obtain information about previous transaction outputs and start verifying scripts as soon as the data is available. For instance, if we begin parallel download and verification of 500 blocks, we should be able to begin script verification of all scripts referencing outputs from existing blocks in parallel, and begin verification of scripts referencing outputs from new blocks as soon as they are committed to the chain.

Because spending outputs from older blocks is more common than spending outputs from recent blocks, this should allow a significant amount of parallelism.

Reference-level explanation

Data structures

We add the following request and response to the state protocol:


#![allow(unused)]
fn main() {
enum Request::AwaitSpendableUtxo {
    outpoint: OutPoint,
    spend_height: Height,
    spend_restriction: SpendRestriction,
}

/// Consensus rule:
/// "A transaction with one or more transparent inputs from coinbase transactions
/// MUST have no transparent outputs (i.e.tx_out_count MUST be 0)."
enum SpendRestriction {
    /// The UTXO is spent in a transaction with transparent outputs
    SomeTransparentOutputs,
    /// The UTXO is spent in a transaction with all shielded outputs
    AllShieldedOutputs,
}
}

As described above, the request name is intended to indicate the request's behavior. The request does not resolve until:

  • the state layer learns of a UTXO described by the request, and
  • the output is spendable at height with spend_restriction.

The new Utxo type adds a coinbase flag and height to transparent::Outputs that we look up in the state, or get from newly committed blocks:


#![allow(unused)]
fn main() {
enum Response::SpendableUtxo(Utxo)

pub struct Utxo {
    /// The output itself.
    pub output: transparent::Output,

    /// The height at which the output was created.
    pub height: block::Height,

    /// Whether the output originated in a coinbase transaction.
    pub from_coinbase: bool,
}
}

Transparent coinbase consensus rules

Specifically, if the UTXO is a transparent coinbase output, the service is not required to return a response if:

  • spend_height is less than MIN_TRANSPARENT_COINBASE_MATURITY (100) blocks after the Utxo.height, or
  • spend_restriction is SomeTransparentOutputs.

This implements the following consensus rules:

A transaction MUST NOT spend a transparent output of a coinbase transaction from a block less than 100 blocks prior to the spend.

Note that transparent outputs of coinbase transactions include Founders’ Reward outputs and transparent funding stream outputs.

A transaction with one or more transparent inputs from coinbase transactions MUST have no transparent outputs (i.e.tx_out_count MUST be 0).

Inputs from coinbase transactions include Founders’ Reward outputs and funding stream outputs.

https://zips.z.cash/protocol/protocol.pdf#txnencodingandconsensus

Parallel coinbase checks

We can perform these coinbase checks asynchronously, in the presence of multiple chain forks, as long as the following conditions both hold:

  1. We don't mistakenly accept or reject spends to the transparent pool.

  2. We don't mistakenly accept or reject mature spends.

Parallel coinbase justification

There are two parts to a spend restriction:

  • the from_coinbase flag, and
  • if the from_coinbase flag is true, the coinbase height.

If a particular transaction hash h always has the same from_coinbase value, and h exists in multiple chains, then regardless of which Utxo arrives first, the outputs of h always get the same from_coinbase value during validation. So spends can not be mistakenly accepted or rejected due to a different coinbase flag.

Similarly, if a particular coinbase transaction hash h always has the same height value, and h exists in multiple chains, then regardless of which Utxo arrives first, the outputs of h always get the same height value during validation. So coinbase spends can not be mistakenly accepted or rejected due to a different height value. (The heights of non-coinbase outputs are irrelevant, because they are never checked.)

These conditions hold as long as the following multi-chain properties are satisfied:

  • from_coinbase: across all chains, the set of coinbase transaction hashes is disjoint from the set of non-coinbase transaction hashes, and
  • coinbase height: across all chains, duplicate coinbase transaction hashes can only occur at exactly the same height.

Parallel coinbase consensus rules

These multi-chain properties can be derived from the following consensus rules:

Transaction versions 1-4:

[Pre-Sapling ] If effectiveVersion = 1 or nJoinSplit = 0, then both tx_in_count and tx_out_count MUST be nonzero. ... [Sapling onward] If effectiveVersion < 5, then at least one of tx_in_count, nSpendsSapling, and nJoinSplit MUST be nonzero.

A coinbase transaction for a block at block height greater than 0 MUST have a script that, as its first item, encodes the block height height as follows.

For height in the range {1 .. 16}, the encoding is a single byte of value 0x50 + height.

Otherwise, let heightBytes be the signed little-endian representation of height, using the minimum nonzero number of bytes such that the most significant byte is < 0x80. The length of heightBytes MUST be in the range {1 .. 8}. Then the encoding is the length of heightBytes encoded as one byte, followed by heightBytes itself.

https://zips.z.cash/protocol/protocol.pdf#txnencodingandconsensus

The transaction ID of a version 4 or earlier transaction is the SHA-256d hash of the transaction encoding in the pre-v5 format described above.

https://zips.z.cash/protocol/protocol.pdf#txnidentifiers

Transaction version 5:

[NU5 onward] If effectiveVersion ≥ 5, then this condition must hold: tx_in_count > 0 or nSpendsSapling > 0 or (nActionsOrchard > 0 and enableSpendsOrchard = 1). ... [NU5 onward] The nExpiryHeight field of a coinbase transaction MUST be equal to its block height.

https://zips.z.cash/protocol/protocol.pdf#txnencodingandconsensus

non-malleable transaction identifiers ... commit to all transaction data except for attestations to transaction validity ... A new transaction digest algorithm is defined that constructs the identifier for a transaction from a tree of hashes ... A BLAKE2b-256 hash of the following values: ... T.1e: expiry_height (4-byte little-endian block height)

https://zips.z.cash/zip-0244#t-1-header-digest

Since:

  • coinbase transaction hashes commit to the block Height,
  • non-coinbase transaction hashes commit to their inputs, and
  • double-spends are not allowed;

Therefore:

  • coinbase transaction hashes are unique for distinct heights in any chain,
  • coinbase transaction hashes are unique in a single chain, and
  • non-coinbase transaction hashes are unique in a single chain, because they recursively commit to unique inputs.

So the required parallel verification conditions are satisfied.

Script verification

To verify scripts, a script verifier requests the relevant UTXOs from the state service and waits for all of them to resolve, or fails verification with a timeout error. Currently, we outsource script verification to zcash_consensus, which does FFI into the same C++ code as zcashd uses. We need to ensure this code is thread-safe.

Database implementation

Implementing the state request correctly requires considering two sets of behaviors:

  1. behaviors related to the state's external API (a Buffered tower::Service);
  2. behaviors related to the state's internal implementation (using rocksdb).

Making this distinction helps us to ensure we don't accidentally leak "internal" behaviors into "external" behaviors, which would violate encapsulation and make it more difficult to replace rocksdb.

In the first category, our state is presented to the rest of the application as a Buffered tower::Service. The Buffer wrapper allows shared access to a service using an actor model, moving the service to be shared into a worker task and passing messages to it over an multi-producer single-consumer (mpsc) channel. The worker task receives messages and makes Service::calls. The Service::call method returns a Future, and the service is allowed to decide how much work it wants to do synchronously (in call) and how much work it wants to do asynchronously (in the Future it returns).

This means that our external API ensures that the state service sees a linearized sequence of state requests, although the exact ordering is unpredictable when there are multiple senders making requests.

Because the state service has exclusive access to the rocksdb database, and the state service sees a linearized sequence of state requests, we have an easy way to opt in to asynchronous database access. We can perform rocksdb operations synchronously in the Service::call, waiting for them to complete, and be sure that all future requests will see the resulting rocksdb state. Or, we can perform rocksdb operations asynchronously in the future returned by Service::call.

If we perform all writes synchronously and allow reads to be either synchronous or asynchronous, we ensure that writes cannot race each other. Asynchronous reads are guaranteed to read at least the state present at the time the request was processed, or a later state.

Lookup states

Now, returning to the UTXO lookup problem, we can map out the possible states with this restriction in mind. This description assumes that UTXO storage is split into disjoint sets, one in-memory (e.g., blocks after the reorg limit) and the other in rocksdb (e.g., blocks after the reorg limit). The details of this storage are not important for this design, only that the two sets are disjoint.

When the state service processes a Request::AwaitSpendableUtxo referencing some UTXO u, there are three disjoint possibilities:

  1. u is already contained in an in-memory block storage;
  2. u is already contained in the rocksdb UTXO set;
  3. u is not yet known to the state service.

In case 3, we need to queue u and scan all future blocks to see whether they contain u. However, if we have a mechanism to queue u, we can perform check 2 asynchronously, because restricting to synchronous writes means that any async read will return the current or later state. If u was in the rocksdb UTXO set when the request was processed, the only way that an async read would not return u is if the UTXO were spent, in which case the service is not required to return a response.

Lookup implementation

This behavior can be encapsulated into a PendingUtxos structure described below.


#![allow(unused)]
fn main() {
// sketch
#[derive(Default, Debug)]
struct PendingUtxos(HashMap<OutPoint, oneshot::Sender<Utxo>>);

impl PendingUtxos {
    // adds the outpoint and returns (wrapped) rx end of oneshot
    // checks the spend height and restriction before sending the utxo response
    // return can be converted to `Service::Future`
    pub fn queue(
        &mut self,
        outpoint: OutPoint,
        spend_height: Height,
        spend_restriction: SpendRestriction,
    ) -> impl Future<Output=Result<Response, ...>>;

    // if outpoint is a hashmap key, remove the entry and send output on the channel
    pub fn respond(&mut self, outpoint: OutPoint, output: transparent::Output);

    /// check the list of pending UTXO requests against the supplied `utxos`
    pub fn check_against(&mut self, utxos: &HashMap<transparent::OutPoint, Utxo>);

    // scans the hashmap and removes any entries with closed senders
    pub fn prune(&mut self);
}
}

The state service should maintain an Arc<Mutex<PendingUtxos>>, used as follows:

  1. In Service::call(Request::AwaitSpendableUtxo { outpoint: u, .. }, the service should:
  • call PendingUtxos::queue(u) to get a future f to return to the caller;
  • spawn a task that does a rocksdb lookup for u, calling PendingUtxos::respond(u, output) if present;
  • check the in-memory storage for u, calling PendingUtxos::respond(u, output) if present;
  • return f to the caller (it may already be ready). The common case is that u references an old spendable UTXO, so spawning the lookup task first means that we don't wait to check in-memory storage for u before starting the rocksdb lookup.
  1. In f, the future returned by PendingUtxos::queue(u), the service should check that the Utxo is spendable before returning it:
  • if Utxo.from_coinbase is false, return the utxo;
  • if Utxo.from_coinbase is true, check that:
    • spend_restriction is AllShieldedOutputs, and
    • spend_height is greater than or equal to MIN_TRANSPARENT_COINBASE_MATURITY plus the Utxo.height,
    • if both checks pass, return the utxo.
    • if any check fails, drop the utxo, and let the request timeout.
  1. In Service::call(Request::CommitBlock(block, ..)), the service should:
  • check for double-spends of each UTXO in the block, and
  • do any other transactional checks before committing a block as normal. Because the AwaitSpendableUtxo request is informational, there's no need to do the transactional checks before matching against pending UTXO requests, and doing so upfront can run expensive verification earlier than needed.
  1. In Service::poll_ready(), the service should call PendingUtxos::prune() at least some of the time. This is required because when a consumer uses a timeout layer, the cancelled requests should be flushed from the queue to avoid a resource leak. However, doing this on every call will result in us spending a bunch of time iterating over the hashmap.

Drawbacks

One drawback of this design is that we may have to wait on a lock. However, the critical section basically amounts to a hash lookup and a channel send, so I don't think that we're likely to run into problems with long contended periods, and it's unlikely that we would get a deadlock.

Rationale and alternatives

High-level design rationale is inline with the design sketch. One low-level option would be to avoid encapsulating behavior in the PendingUtxos and just have an Arc<Hashmap<..>>, so that the lock only protects the hashmap lookup and not sending through the channel. But I think the current design is cleaner and the cost is probably not too large.

Unresolved questions

  • We need to pick a timeout for UTXO lookup. This should be long enough to account for the fact that we may start verifying blocks before all of their ancestors are downloaded.

State Updates

  • Feature Name: state_updates
  • Start Date: 2020-08-14
  • Design PR: https://github.com/ZcashFoundation/zebra/pull/902
  • Zebra Issue: https://github.com/ZcashFoundation/zebra/issues/1049

Summary

Zebra manages chain state in the zebra-state crate, which allows state queries via asynchronous RPC (in the form of a Tower service). The state system is responsible for contextual verification in the sense of RFC2, checking that new blocks are consistent with the existing chain state before committing them. This RFC describes how the state is represented internally, and how state updates are performed.

Motivation

We need to be able to access and modify the chain state, and we want to have a description of how this happens and what guarantees are provided by the state service.

Definitions

  • state data: Any data the state service uses to represent chain state.

  • structural/semantic/contextual verification: as defined in RFC2.

  • block chain: A sequence of valid blocks linked by inclusion of the previous block hash in the subsequent block. Chains are rooted at the genesis block and extend to a tip.

  • chain state: The state of the ledger after application of a particular sequence of blocks (state transitions).

  • block work: The approximate amount of work required for a miner to generate a block hash that passes the difficulty filter. The number of block header attempts and the mining time are proportional to the work value. Numerically higher work values represent longer processing times.

  • cumulative work: The sum of the block work of all blocks in a chain, from genesis to the chain tip.

  • best chain: The chain with the greatest cumulative work. This chain represents the consensus state of the Zcash network and transactions.

  • side chain: A chain which is not contained in the best chain. Side chains are pruned at the reorg limit, when they are no longer connected to the finalized state.

  • chain reorganization: Occurs when a new best chain is found and the previous best chain becomes a side chain.

  • reorg limit: The longest reorganization accepted by zcashd, 100 blocks.

  • orphaned block: A block which is no longer included in the best chain.

  • non-finalized state: State data corresponding to blocks above the reorg limit. This data can change in the event of a chain reorg.

  • finalized state: State data corresponding to blocks below the reorg limit. This data cannot change in the event of a chain reorg.

  • non-finalized tips: The highest blocks in each non-finalized chain. These tips might be at different heights.

  • finalized tip: The highest block in the finalized state. The tip of the best chain is usually 100 blocks (the reorg limit) above the finalized tip. But it can be lower during the initial sync, and after a chain reorganization, if the new best chain is at a lower height.

  • relevant chain: The relevant chain for a block starts at the previous block, and extends back to genesis.

  • relevant tip: The tip of the relevant chain.

Guide-level explanation

The zebra-state crate provides an implementation of the chain state storage logic in a Zcash consensus node. Its main responsibility is to store chain state, validating new blocks against the existing chain state in the process, and to allow later querying of said chain state. zebra-state provides this interface via a tower::Service based on the actor model with a request/response interface for passing messages back and forth between the state service and the rest of the application.

The main entry point for the zebra-state crate is the init function. This function takes a zebra_state::Config and constructs a new state service, which it returns wrapped by a tower::Buffer. This service is then interacted with via the tower::Service trait.


#![allow(unused)]
fn main() {
use tower::{Service, ServiceExt};

let state = zebra_state::on_disk::init(state_config, network);
let request = zebra_state::Request::BlockLocator;
let response = state.ready_and().await?.call(request).await?;

assert!(matches!(response, zebra_state::Response::BlockLocator(_)));
}

Note: The tower::Service API requires that ready is always called exactly once before each call. It is up to users of the zebra state service to uphold this contract.

The tower::Buffer wrapper is Cloneable, allowing shared access to a common state service. This allows different tasks to share access to the chain state.

The set of operations supported by zebra-state are encoded in its Request enum. This enum has one variant for each supported operation.


#![allow(unused)]
fn main() {
pub enum Request {
    CommitBlock {
        block: Arc<Block>,
    },
    CommitFinalizedBlock {
        block: Arc<Block>,
    },
    Depth(Hash),
    Tip,
    BlockLocator,
    Transaction(Hash),
    Block(HashOrHeight),

    // .. some variants omitted
}
}

zebra-state breaks down its requests into two categories and provides different guarantees for each category: requests that modify the state, and requests that do not. Requests that update the state are guaranteed to run sequentially and will never race against each other. Requests that read state are done asynchronously and are guaranteed to read at least the state present at the time the request was processed by the service, or a later state present at the time the request future is executed. The state service avoids race conditions between the read state and the written state by doing all contextual verification internally.

Reference-level explanation

State Components

Zcash (as implemented by zcashd) differs from Bitcoin in its treatment of transaction finality. If a new best chain is detected that does not extend the previous best chain, blocks at the end of the previous best chain become orphaned (no longer included in the best chain). Their state updates are therefore no longer included in the best chain's chain state. The process of rolling back orphaned blocks and applying new blocks is called a chain reorganization. Bitcoin allows chain reorganizations of arbitrary depth, while zcashd limits chain reorganizations to 100 blocks. (In zcashd, the new best chain must be a side-chain that forked within 100 blocks of the tip of the current best chain.)

This difference means that in Bitcoin, chain state only has probabilistic finality, while in Zcash, chain state is final once it is beyond the reorg limit. To simplify our implementation, we split the representation of the state data at the finality boundary provided by the reorg limit.

State data from blocks above the reorg limit (non-finalized state) is stored in-memory and handles multiple chains. State data from blocks below the reorg limit (finalized state) is stored persistently using rocksdb and only tracks a single chain. This allows a simplification of our state handling, because only finalized data is persistent and the logic for finalized data handles less invariants.

One downside of this design is that restarting the node loses the last 100 blocks, but node restarts are relatively infrequent and a short re-sync is cheap relative to the cost of additional implementation complexity.

Another downside of this design is that we do not achieve exactly the same behavior as zcashd in the event of a 51% attack: zcashd limits each chain reorganization to 100 blocks, but permits multiple reorgs, while Zebra limits all chain reorgs to 100 blocks. In the event of a successful 51% attack on Zcash, this could be resolved by wiping the rocksdb state and re-syncing the new chain, but in this scenario there are worse problems.

Service Interface

The state is accessed asynchronously through a Tower service interface. Determining what guarantees the state service can and should provide to the rest of the application requires considering two sets of behaviors:

  1. behaviors related to the state's external API (a Buffered tower::Service);
  2. behaviors related to the state's internal implementation (using rocksdb).

Making this distinction helps us to ensure we don't accidentally leak "internal" behaviors into "external" behaviors, which would violate encapsulation and make it more difficult to replace rocksdb.

In the first category, our state is presented to the rest of the application as a Buffered tower::Service. The Buffer wrapper allows shared access to a service using an actor model, moving the service to be shared into a worker task and passing messages to it over an multi-producer single-consumer (mpsc) channel. The worker task receives messages and makes Service::calls. The Service::call method returns a Future, and the service is allowed to decide how much work it wants to do synchronously (in call) and how much work it wants to do asynchronously (in the Future it returns).

This means that our external API ensures that the state service sees a linearized sequence of state requests, although the exact ordering is unpredictable when there are multiple senders making requests.

Because the state service has exclusive access to the rocksdb database, and the state service sees a linearized sequence of state requests, we have an easy way to opt in to asynchronous database access. We can perform rocksdb operations synchronously in the Service::call, waiting for them to complete, and be sure that all future requests will see the resulting rocksdb state. Or, we can perform rocksdb operations asynchronously in the future returned by Service::call.

If we perform all writes synchronously and allow reads to be either synchronous or asynchronous, we ensure that writes cannot race each other. Asynchronous reads are guaranteed to read at least the state present at the time the request was processed, or a later state.

Summary

  • rocksdb reads may be done synchronously (in call) or asynchronously (in the Future), depending on the context;

  • rocksdb writes must be done synchronously (in call)

In-memory data structures

At a high level, the in-memory data structures store a collection of chains, each rooted at the highest finalized block. Each chain consists of a map from heights to blocks. Chains are stored using an ordered map from cumulative work to chains, so that the map ordering is the ordering of worst to best chains.

The Chain type

The Chain type represents a chain of blocks. Each block represents an incremental state update, and the Chain type caches the cumulative state update from its root to its tip.

The Chain type is used to represent the non-finalized portion of a complete chain of blocks rooted at the genesis block. The parent block of the root of a Chain is the tip of the finalized portion of the chain. As an exception, the finalized portion of the chain is initially empty, until the genesis block has been finalized.

The Chain type supports several operations to manipulate chains, push, pop_root, and fork. push is the most fundamental operation and handles contextual validation of chains as they are extended. pop_root is provided for finalization, and is how we move blocks from the non-finalized portion of the state to the finalized portion. fork on the other hand handles creating new chains for push when new blocks arrive whose parent isn't a tip of an existing chain.

Note: The Chain type's API is only designed to handle non-finalized data. The genesis block and all pre canopy blocks are always considered to be finalized blocks and should not be handled via the Chain type through CommitBlock. They should instead be committed directly to the finalized state with CommitFinalizedBlock. This is particularly important with the genesis block since the Chain will panic if used while the finalized state is completely empty.

The Chain type is defined by the following struct and API:


#![allow(unused)]
fn main() {
#[derive(Debug, Clone)]
pub struct Chain {
    // The function `eq_internal_state` must be updated every time a field is added to [`Chain`].
    /// The configured network for this chain.
    network: Network,

    /// The contextually valid blocks which form this non-finalized partial chain, in height order.
    pub(crate) blocks: BTreeMap<block::Height, ContextuallyValidBlock>,

    /// An index of block heights for each block hash in `blocks`.
    pub height_by_hash: HashMap<block::Hash, block::Height>,

    /// An index of [`TransactionLocation`]s for each transaction hash in `blocks`.
    pub tx_loc_by_hash: HashMap<transaction::Hash, TransactionLocation>,

    /// The [`transparent::Utxo`]s created by `blocks`.
    ///
    /// Note that these UTXOs may not be unspent.
    /// Outputs can be spent by later transactions or blocks in the chain.
    //
    // TODO: replace OutPoint with OutputLocation?
    pub(crate) created_utxos: HashMap<transparent::OutPoint, transparent::OrderedUtxo>,
    /// The [`transparent::OutPoint`]s spent by `blocks`,
    /// including those created by earlier transactions or blocks in the chain.
    pub(crate) spent_utxos: HashSet<transparent::OutPoint>,

    /// The Sprout note commitment tree of the tip of this [`Chain`],
    /// including all finalized notes, and the non-finalized notes in this chain.
    pub(super) sprout_note_commitment_tree: sprout::tree::NoteCommitmentTree,
    /// The Sprout note commitment tree for each anchor.
    /// This is required for interstitial states.
    pub(crate) sprout_trees_by_anchor:
        HashMap<sprout::tree::Root, sprout::tree::NoteCommitmentTree>,
    /// The Sapling note commitment tree of the tip of this [`Chain`],
    /// including all finalized notes, and the non-finalized notes in this chain.
    pub(super) sapling_note_commitment_tree: sapling::tree::NoteCommitmentTree,
    /// The Sapling note commitment tree for each height.
    pub(crate) sapling_trees_by_height: BTreeMap<block::Height, sapling::tree::NoteCommitmentTree>,
    /// The Orchard note commitment tree of the tip of this [`Chain`],
    /// including all finalized notes, and the non-finalized notes in this chain.
    pub(super) orchard_note_commitment_tree: orchard::tree::NoteCommitmentTree,
    /// The Orchard note commitment tree for each height.
    pub(crate) orchard_trees_by_height: BTreeMap<block::Height, orchard::tree::NoteCommitmentTree>,
    /// The ZIP-221 history tree of the tip of this [`Chain`],
    /// including all finalized blocks, and the non-finalized `blocks` in this chain.
    pub(crate) history_tree: HistoryTree,

    /// The Sprout anchors created by `blocks`.
    pub(crate) sprout_anchors: MultiSet<sprout::tree::Root>,
    /// The Sprout anchors created by each block in `blocks`.
    pub(crate) sprout_anchors_by_height: BTreeMap<block::Height, sprout::tree::Root>,
    /// The Sapling anchors created by `blocks`.
    pub(crate) sapling_anchors: MultiSet<sapling::tree::Root>,
    /// The Sapling anchors created by each block in `blocks`.
    pub(crate) sapling_anchors_by_height: BTreeMap<block::Height, sapling::tree::Root>,
    /// The Orchard anchors created by `blocks`.
    pub(crate) orchard_anchors: MultiSet<orchard::tree::Root>,
    /// The Orchard anchors created by each block in `blocks`.
    pub(crate) orchard_anchors_by_height: BTreeMap<block::Height, orchard::tree::Root>,

    /// The Sprout nullifiers revealed by `blocks`.
    pub(super) sprout_nullifiers: HashSet<sprout::Nullifier>,
    /// The Sapling nullifiers revealed by `blocks`.
    pub(super) sapling_nullifiers: HashSet<sapling::Nullifier>,
    /// The Orchard nullifiers revealed by `blocks`.
    pub(super) orchard_nullifiers: HashSet<orchard::Nullifier>,

    /// Partial transparent address index data from `blocks`.
    pub(super) partial_transparent_transfers: HashMap<transparent::Address, TransparentTransfers>,

    /// The cumulative work represented by `blocks`.
    ///
    /// Since the best chain is determined by the largest cumulative work,
    /// the work represented by finalized blocks can be ignored,
    /// because they are common to all non-finalized chains.
    pub(super) partial_cumulative_work: PartialCumulativeWork,

    /// The chain value pool balances of the tip of this [`Chain`],
    /// including the block value pool changes from all finalized blocks,
    /// and the non-finalized blocks in this chain.
    ///
    /// When a new chain is created from the finalized tip,
    /// it is initialized with the finalized tip chain value pool balances.
    pub(crate) chain_value_pools: ValueBalance<NonNegative>,
}
}

pub fn push(&mut self, block: Arc<Block>)

Push a block into a chain as the new tip

  1. Update cumulative data members

    • Add the block's hash to height_by_hash
    • Add work to self.partial_cumulative_work
    • For each transaction in block
      • Add key: transaction.hash and value: (height, tx_index) to tx_loc_by_hash
      • Add created utxos to self.created_utxos
      • Add spent utxos to self.spent_utxos
      • Add nullifiers to the appropriate self.<version>_nullifiers

  2. Add block to self.blocks

pub fn pop_root(&mut self) -> Arc<Block>

Remove the lowest height block of the non-finalized portion of a chain.

  1. Remove the lowest height block from self.blocks

  2. Update cumulative data members

    • Remove the block's hash from self.height_by_hash
    • Subtract work from self.partial_cumulative_work
    • For each transaction in block
      • Remove transaction.hash from tx_loc_by_hash
      • Remove created utxos from self.created_utxos
      • Remove spent utxos from self.spent_utxos
      • Remove the nullifiers from the appropriate self.<version>_nullifiers

  3. Return the block

pub fn fork(&self, new_tip: block::Hash) -> Option<Self>

Fork a chain at the block with the given hash, if it is part of this chain.

  1. If self does not contain new_tip return None

  2. Clone self as forked

  3. While the tip of forked is not equal to new_tip

    • call forked.pop_tip() and discard the old tip

  4. Return forked

fn pop_tip(&mut self)

Remove the highest height block of the non-finalized portion of a chain.

  1. Remove the highest height block from self.blocks

  2. Update cumulative data members

    • Remove the corresponding hash from self.height_by_hash
    • Subtract work from self.partial_cumulative_work
    • for each transaction in block
      • remove transaction.hash from tx_loc_by_hash
      • Remove created utxos from self.created_utxos
      • Remove spent utxos from self.spent_utxos
      • Remove the nullifiers from the appropriate self.<version>_nullifiers

Ord

The Chain type implements Ord for reorganizing chains. First chains are compared by their partial_cumulative_work. Ties are then broken by comparing block::Hashes of the tips of each chain. (This tie-breaker means that all Chains in the NonFinalizedState must have at least one block.)

Note: Unlike zcashd, Zebra does not use block arrival times as a tie-breaker for the best tip. Since Zebra downloads blocks in parallel, download times are not guaranteed to be unique. Using the block::Hash provides a consistent tip order. (As a side-effect, the tip order is also consistent after a node restart, and between nodes.)

Default

The Chain type implements Default for constructing new chains whose parent block is the tip of the finalized state. This implementation should be handled by #[derive(Default)].

  1. initialise cumulative data members
    • Construct an empty self.blocks, height_by_hash, tx_loc_by_hash, self.created_utxos, self.spent_utxos, self.<version>_anchors, self.<version>_nullifiers
    • Zero self.partial_cumulative_work

Note: The ChainState can be empty after a restart, because the non-finalized state is empty.

NonFinalizedState Type

The NonFinalizedState type represents the set of all non-finalized state. It consists of a set of non-finalized but verified chains and a set of unverified blocks which are waiting for the full context needed to verify them to become available.

NonFinalizedState is defined by the following structure and API:


#![allow(unused)]
fn main() {
/// The state of the chains in memory, including queued blocks.
#[derive(Debug, Default)]
pub struct NonFinalizedState {
    /// Verified, non-finalized chains.
    chain_set: BTreeSet<Chain>,
    /// Blocks awaiting their parent blocks for contextual verification.
    contextual_queue: QueuedBlocks,
}
}

pub fn finalize(&mut self) -> Arc<Block>

Finalize the lowest height block in the non-finalized portion of the best chain and updates all side chains to match.

  1. Extract the best chain from self.chain_set into best_chain

  2. Extract the rest of the chains into a side_chains temporary variable, so they can be mutated

  3. Remove the lowest height block from the best chain with let finalized_block = best_chain.pop_root();

  4. Add best_chain back to self.chain_set if best_chain is not empty

  5. For each remaining chain in side_chains

    • remove the lowest height block from chain
    • If that block is equal to finalized_block and chain is not empty add chain back to self.chain_set
    • Else, drop chain

  6. Return finalized_block

fn commit_block(&mut self, block: Arc<Block>)

Commit block to the non-finalized state.

  1. If the block is a pre-Canopy block, or the canopy activation block, panic.

  2. If any chains tip hash equal block.header.previous_block_hash remove that chain from self.chain_set

  3. Else Find the first chain that contains block.parent and fork it with block.parent as the new tip

    • let fork = self.chain_set.iter().find_map(|chain| chain.fork(block.parent));

  4. Else panic, this should be unreachable because commit_block is only called when block is ready to be committed.

  5. Push block into parent_chain

  6. Insert parent_chain into self.chain_set

pub(super) fn commit_new_chain(&mut self, block: Arc<Block>)

Construct a new chain starting with block.

  1. Construct a new empty chain

  2. push block into that new chain

  3. Insert the new chain into self.chain_set

The QueuedBlocks type

The queued blocks type represents the non-finalized blocks that were committed before their parent blocks were. It is responsible for tracking which blocks are queued by their parent so they can be committed immediately after the parent is committed. It also tracks blocks by their height so they can be discarded if they ever end up below the reorg limit.

NonFinalizedState is defined by the following structure and API:


#![allow(unused)]
fn main() {
/// A queue of blocks, awaiting the arrival of parent blocks.
#[derive(Debug, Default)]
struct QueuedBlocks {
    /// Blocks awaiting their parent blocks for contextual verification.
    blocks: HashMap<block::Hash, QueuedBlock>,
    /// Hashes from `queued_blocks`, indexed by parent hash.
    by_parent: HashMap<block::Hash, Vec<block::Hash>>,
    /// Hashes from `queued_blocks`, indexed by block height.
    by_height: BTreeMap<block::Height, Vec<block::Hash>>,
}
}

pub fn queue(&mut self, new: QueuedBlock)

Add a block to the queue of blocks waiting for their requisite context to become available.

  1. extract the parent_hash, new_hash, and new_height from new.block

  2. Add new to self.blocks using new_hash as the key

  3. Add new_hash to the set of hashes in self.by_parent.entry(parent_hash).or_default()

  4. Add new_hash to the set of hashes in self.by_height.entry(new_height).or_default()

pub fn dequeue_children(&mut self, parent: block::Hash) -> Vec<QueuedBlock>

Dequeue the set of blocks waiting on parent.

  1. Remove the set of hashes waiting on parent from self.by_parent

  2. Remove and collect each block in that set of hashes from self.blocks as queued_children

  3. For each block in queued_children remove the associated block.hash from self.by_height

  4. Return queued_children

pub fn prune_by_height(&mut self, finalized_height: block::Height)

Prune all queued blocks whose height are less than or equal to finalized_height.

  1. Split the by_height list at the finalized height, removing all heights that are below finalized_height

  2. for each hash in the removed values of by_height

    • remove the corresponding block from self.blocks
    • remove the block's hash from the list of blocks waiting on block.header.previous_block_hash from self.by_parent

Summary

  • Chain represents the non-finalized portion of a single chain

  • NonFinalizedState represents the non-finalized portion of all chains

  • QueuedBlocks represents all unverified blocks that are waiting for context to be available.

The state service uses the following entry points:

  • commit_block when it receives new blocks.

  • finalize to prevent chains in NonFinalizedState from growing beyond the reorg limit.

  • FinalizedState.queue_and_commit_finalized_blocks on the blocks returned by finalize, to commit those finalized blocks to disk.

Committing non-finalized blocks

New non-finalized blocks are committed as follows:

pub(super) fn queue_and_commit_non_finalized_blocks(&mut self, new: Arc<Block>) -> tokio::sync::oneshot::Receiver<block::Hash>

  1. If a duplicate block hash exists in a non-finalized chain, or the finalized chain, it has already been successfully verified:

    • create a new oneshot channel
    • immediately send Err(DuplicateBlockHash) drop the sender
    • return the receiver

  2. If a duplicate block hash exists in the queue:

    • Find the QueuedBlock for that existing duplicate block
    • create a new channel for the new request
    • replace the old sender in queued_block with the new sender
    • send Err(DuplicateBlockHash) through the old sender channel
    • continue to use the new receiver

  3. Else create a QueuedBlock for block:

    • Create a tokio::sync::oneshot channel
    • Use that channel to create a QueuedBlock for block
    • Add block to self.queued_blocks
    • continue to use the new receiver

  4. If block.header.previous_block_hash is not present in the finalized or non-finalized state:

    • Return the receiver for the block's channel

  5. Else iteratively attempt to process queued blocks by their parent hash starting with block.header.previous_block_hash

  6. While there are recently committed parent hashes to process

    • Dequeue all blocks waiting on parent with let queued_children = self.queued_blocks.dequeue_children(parent);
    • for each queued block
      • Run contextual validation on block
        • contextual validation should check that the block height is equal to the previous block height plus 1. This check will reject blocks with invalid heights.
      • If the block fails contextual validation send the result to the associated channel
      • Else if the block's previous hash is the finalized tip add to the non-finalized state with self.mem.commit_new_chain(block)
      • Else add the new block to an existing non-finalized chain or new fork with self.mem.commit_block(block);
      • Send Ok(hash) over the associated channel to indicate the block was successfully committed
      • Add block.hash to the set of recently committed parent hashes to process

  7. While the length of the non-finalized portion of the best chain is greater than the reorg limit

    • Remove the lowest height block from the non-finalized state with self.mem.finalize();
    • Commit that block to the finalized state with self.disk.commit_finalized_direct(finalized);

  8. Prune orphaned blocks from self.queued_blocks with self.queued_blocks.prune_by_height(finalized_height);

  9. Return the receiver for the block's channel

rocksdb data structures

rocksdb provides a persistent, thread-safe BTreeMap<&[u8], &[u8]>. Each map is a distinct "tree". Keys are sorted using lex order on byte strings, so integer values should be stored using big-endian encoding (so that the lex order on byte strings is the numeric ordering).

Note that the lex order storage allows creating 1-to-many maps using keys only. For example, the tx_loc_by_transparent_addr_loc allows mapping each address to all transactions related to it, by simply storing each transaction prefixed with the address as the key, leaving the value empty. Since rocksdb allows listing all keys with a given prefix, it will allow listing all transactions related to a given address.

We use the following rocksdb column families:

Column FamilyKeysValuesChanges
Blocks
hash_by_heightblock::Heightblock::HashCreate
height_by_hashblock::Hashblock::HeightCreate
block_header_by_heightblock::Heightblock::HeaderCreate
Transactions
tx_by_locTransactionLocationTransactionCreate
hash_by_tx_locTransactionLocationtransaction::HashCreate
tx_loc_by_hashtransaction::HashTransactionLocationCreate
Transparent
balance_by_transparent_addrtransparent::AddressAmount \|\| AddressLocationUpdate
tx_loc_by_transparent_addr_locAddressTransaction()Create
utxo_by_out_locOutputLocationtransparent::OutputDelete
utxo_loc_by_transparent_addr_locAddressUnspentOutput()Delete
Sprout
sprout_nullifierssprout::Nullifier()Create
sprout_anchorssprout::tree::Rootsprout::NoteCommitmentTreeCreate
sprout_note_commitment_treeblock::Heightsprout::NoteCommitmentTreeDelete
Sapling
sapling_nullifierssapling::Nullifier()Create
sapling_anchorssapling::tree::Root()Create
sapling_note_commitment_treeblock::Heightsapling::NoteCommitmentTreeCreate
Orchard
orchard_nullifiersorchard::Nullifier()Create
orchard_anchorsorchard::tree::Root()Create
orchard_note_commitment_treeblock::Heightorchard::NoteCommitmentTreeCreate
Chain
history_treeblock::HeightNonEmptyHistoryTreeDelete
tip_chain_value_pool()ValueBalanceUpdate

Zcash structures are encoded using ZcashSerialize/ZcashDeserialize. Other structures are encoded using IntoDisk/FromDisk.

Block and Transaction Data:

  • Height: 24 bits, big-endian, unsigned (allows for ~30 years worth of blocks)
  • TransactionIndex: 16 bits, big-endian, unsigned (max ~23,000 transactions in the 2 MB block limit)
  • TransactionCount: same as TransactionIndex
  • TransactionLocation: Height \|\| TransactionIndex
  • OutputIndex: 24 bits, big-endian, unsigned (max ~223,000 transfers in the 2 MB block limit)
  • transparent and shielded input indexes, and shielded output indexes: 16 bits, big-endian, unsigned (max ~49,000 transfers in the 2 MB block limit)
  • OutputLocation: TransactionLocation \|\| OutputIndex
  • AddressLocation: the first OutputLocation used by a transparent::Address. Always has the same value for each address, even if the first output is spent.
  • Utxo: Output, derives extra fields from the OutputLocation key
  • AddressUnspentOutput: AddressLocation \|\| OutputLocation, used instead of a BTreeSet<OutputLocation> value, to improve database performance
  • AddressTransaction: AddressLocation \|\| TransactionLocation used instead of a BTreeSet<TransactionLocation> value, to improve database performance

We use big-endian encoding for keys, to allow database index prefix searches.

Amounts:

  • Amount: 64 bits, little-endian, signed
  • ValueBalance: [Amount; 4]

Derived Formats:

  • *::NoteCommitmentTree: bincode using serde
  • NonEmptyHistoryTree: bincode using serde, using zcash_history's serde implementation

The following figure helps visualizing the address index, which is the most complicated part. Numbers in brackets are array sizes; bold arrows are compositions (i.e. TransactionLocation is the concatenation of Height and TransactionIndex); dashed arrows are compositions that are also 1-to-many maps (i.e. AddressTransaction is the concatenation of AddressLocation and TransactionLocation, but also is used to map each AddressLocation to multiple TransactionLocations).

graph TD;
    Address -->|"balance_by_transparent_addr<br/>"| AddressBalance;
    AddressBalance ==> Amount;
    AddressBalance ==> AddressLocation;
    AddressLocation ==> FirstOutputLocation;
    AddressLocation -.->|"tx_loc_by_transparent_addr_loc<br/>(AddressTransaction[13])"| TransactionLocation;
    TransactionLocation ==> Height;
    TransactionLocation ==> TransactionIndex;
    OutputLocation -->|utxo_by_out_loc| Output;
    OutputLocation ==> TransactionLocation;
    OutputLocation ==> OutputIndex;
    AddressLocation -.->|"utxo_loc_by_transparent_addr_loc<br/>(AddressUnspentOutput[16])"| OutputLocation;

    AddressBalance["AddressBalance[16]"];
    Amount["Amount[8]"];
    Height["Height[3]"];
    Address["Address[21]"];
    TransactionIndex["TransactionIndex[2]"];
    TransactionLocation["TransactionLocation[5]"];
    OutputIndex["OutputIndex[3]"];
    OutputLocation["OutputLocation[8]"];
    FirstOutputLocation["First OutputLocation[8]"];
    AddressLocation["AddressLocation[8]"];

Implementing consensus rules using rocksdb

Each column family handles updates differently, based on its specific consensus rules:

  • Create:
    • Each key-value entry is created once.
    • Keys are never deleted, values are never updated.
  • Delete:
    • Each key-value entry is created once.
    • Keys can be deleted, but values are never updated.
    • Code called by ReadStateService must ignore deleted keys, or use a read lock.
    • TODO: should we prevent re-inserts of keys that have been deleted?
  • Update:
    • Each key-value entry is created once.
    • Keys are never deleted, but values can be updated.
    • Code called by ReadStateService must handle old or new values, or use a read lock.

We can't do some kinds of value updates, because they cause RocksDB performance issues:

  • Append:
    • Keys are never deleted.
    • Existing values are never updated.
    • Sets of values have additional items appended to the end of the set.
    • Code called by ReadStateService must handle shorter or longer sets, or use a read lock.
  • Up/Del:
    • Keys can be deleted.
    • Sets of values have items added or deleted (in any position).
    • Code called by ReadStateService must ignore deleted keys and values, accept shorter or longer sets, and accept old or new values. Or it should use a read lock.

Avoid using large sets of values as RocksDB keys or values.

RocksDB read locks

The read-only ReadStateService needs to handle concurrent writes and deletes of the finalized column families it reads. It must also handle overlaps between the cached non-finalized Chain, and the current finalized state database.

The StateService uses RocksDB transactions for each block write. So ReadStateService queries that only access a single key or value will always see a consistent view of the database.

If a ReadStateService query only uses column families that have keys and values appended (Never in the Updates table above), it should ignore extra appended values. Most queries do this by default.

For more complex queries, there are several options:

Reading across multiple column families:

  1. Ignore deleted values using custom Rust code
  2. Take a database snapshot - https://docs.rs/rocksdb/latest/rocksdb/struct.DBWithThreadMode.html#method.snapshot

Reading a single column family: 3. multi_get - https://docs.rs/rocksdb/latest/rocksdb/struct.DBWithThreadMode.html#method.multi_get_cf 4. iterator - https://docs.rs/rocksdb/latest/rocksdb/struct.DBWithThreadMode.html#method.iterator_cf

RocksDB also has read transactions, but they don't seem to be exposed in the Rust crate.

Low-Level Implementation Details

RocksDB ignores duplicate puts and deletes, preserving the latest values. If rejecting duplicate puts or deletes is consensus-critical, check db.get_cf(cf, key)? before putting or deleting any values in a batch.

Currently, these restrictions should be enforced by code review:

  • multiple zs_inserts are only allowed on Update column families, and
  • delete_cf is only allowed on Delete column families.

In future, we could enforce these restrictions by:

  • creating traits for Never, Delete, and Update
  • doing different checks in zs_insert depending on the trait
  • wrapping delete_cf in a trait, and only implementing that trait for types that use Delete column families.

As of June 2021, the Rust rocksdb crate ignores the delete callback, and merge operators are unreliable (or have undocumented behaviour). So they should not be used for consensus-critical checks.

Notes on rocksdb column families

  • The hash_by_height and height_by_hash column families provide a bijection between block heights and block hashes. (Since the rocksdb state only stores finalized state, they are actually a bijection).

  • Similarly, the tx_loc_by_hash and hash_by_tx_loc column families provide a bijection between transaction locations and transaction hashes.

  • The block_header_by_height column family provides a bijection between block heights and block header data. There is no corresponding height_by_block column family: instead, hash the block header, and use the hash from height_by_hash. (Since the rocksdb state only stores finalized state, they are actually a bijection). Similarly, there are no column families that go from transaction data to transaction locations: hash the transaction and use tx_loc_by_hash.

  • Block headers and transactions are stored separately in the database, so that individual transactions can be accessed efficiently. Blocks can be re-created on request using the following process:

    • Look up height in height_by_hash
    • Get the block header for height from block_header_by_height
    • Iterate from TransactionIndex 0, to get each transaction with height from tx_by_loc, stopping when there are no more transactions in the block

  • Block headers are stored by height, not by hash. This has the downside that looking up a block by hash requires an extra level of indirection. The upside is that blocks with adjacent heights are adjacent in the database, and many common access patterns, such as helping a client sync the chain or doing analysis, access blocks in (potentially sparse) height order. In addition, the fact that we commit blocks in order means we're writing only to the end of the rocksdb column family, which may help save space.

  • Similarly, transaction data is stored in chain order in tx_by_loc and utxo_by_out_loc, and chain order within each vector in utxo_loc_by_transparent_addr_loc and tx_loc_by_transparent_addr_loc.

  • TransactionLocations are stored as a (height, index) pair referencing the height of the transaction's parent block and the transaction's index in that block. This would more traditionally be a (hash, index) pair, but because we store blocks by height, storing the height saves one level of indirection. Transaction hashes can be looked up using hash_by_tx_loc.

  • Similarly, UTXOs are stored in utxo_by_out_loc by OutputLocation, rather than OutPoint. OutPoints can be looked up using tx_loc_by_hash, and reconstructed using hash_by_tx_loc.

  • The Utxo type can be constructed from the OutputLocation and Output data, height: OutputLocation.height, and is_coinbase: OutputLocation.transaction_index == 0 (coinbase transactions are always the first transaction in a block).

  • balance_by_transparent_addr is the sum of all utxo_loc_by_transparent_addr_locs that are still in utxo_by_out_loc. It is cached to improve performance for addresses with large UTXO sets. It also stores the AddressLocation for each address, which allows for efficient lookups.

  • utxo_loc_by_transparent_addr_loc stores unspent transparent output locations by address. The address location and UTXO location are stored as a RocksDB key, so they are in chain order, and get good database performance. This column family includes also includes the original address location UTXO, if it has not been spent.

  • When a block write deletes a UTXO from utxo_by_out_loc, that UTXO location should be deleted from utxo_loc_by_transparent_addr_loc. The deleted UTXO can be removed efficiently, because the UTXO location is part of the key. This is an index optimisation, which does not affect query results.

  • tx_loc_by_transparent_addr_loc stores transaction locations by address. This list includes transactions containing spent UTXOs. The address location and transaction location are stored as a RocksDB key, so they are in chain order, and get good database performance. This column family also includes the TransactionLocation of the transaction for the AddressLocation.

  • The sprout_note_commitment_tree stores the note commitment tree state at the tip of the finalized state, for the specific pool. There is always a single entry. Each tree is stored as a "Merkle tree frontier" which is basically a (logarithmic) subset of the Merkle tree nodes as required to insert new items. For each block committed, the old tree is deleted and a new one is inserted by its new height. TODO: store the sprout note commitment tree by (), to avoid ReadStateService concurrent write issues.

  • The {sapling, orchard}_note_commitment_tree stores the note commitment tree state for every height, for the specific pool. Each tree is stored as a "Merkle tree frontier" which is basically a (logarithmic) subset of the Merkle tree nodes as required to insert new items.

  • history_tree stores the ZIP-221 history tree state at the tip of the finalized state. There is always a single entry for it. The tree is stored as the set of "peaks" of the "Merkle mountain range" tree structure, which is what is required to insert new items. TODO: store the history tree by (), to avoid ReadStateService concurrent write issues.

  • Each *_anchors stores the anchor (the root of a Merkle tree) of the note commitment tree of a certain block. We only use the keys since we just need the set of anchors, regardless of where they come from. The exception is sprout_anchors which also maps the anchor to the matching note commitment tree. This is required to support interstitial treestates, which are unique to Sprout. TODO: store the Root hash in sprout_note_commitment_tree, and use it to look up the note commitment tree. This de-duplicates tree state data. But we currently only store one sprout tree by height.

  • The value pools are only stored for the finalized tip.

  • We do not store the cumulative work for the finalized chain, because the finalized work is equal for all non-finalized chains. So the additional non-finalized work can be used to calculate the relative chain order, and choose the best chain.

Committing finalized blocks

If the parent block is not committed, add the block to an internal queue for future processing. Otherwise, commit the block described below, then commit any queued children. (Although the checkpointer generates verified blocks in order when it completes a checkpoint, the blocks are committed in the response futures, so they may arrive out of order).

Committing a block to the rocksdb state should be implemented as a wrapper around a function also called by Request::CommitBlock, which should:

pub(super) fn queue_and_commit_finalized_blocks(&mut self, queued_block: QueuedBlock)

  1. Obtain the highest entry of hash_by_height as (old_height, old_tip). Check that block's parent hash is old_tip and its height is old_height+1, or panic. This check is performed as defense-in-depth to prevent database corruption, but it is the caller's responsibility (e.g. the zebra-state service's responsibility) to commit finalized blocks in order.

The genesis block does not have a parent block. For genesis blocks, check that block's parent hash is null (all zeroes) and its height is 0.

  1. Insert the block and transaction data into the relevant column families.

  2. If the block is a genesis block, skip any transaction updates.

    (Due to a bug in zcashd, genesis block anchors and transactions are ignored during validation.)

  3. Update the block anchors, history tree, and chain value pools.

  4. Iterate over the enumerated transactions in the block. For each transaction, update the relevant column families.

Note: The Sprout and Sapling anchors are the roots of the Sprout and Sapling note commitment trees that have already been calculated for the last transaction(s) in the block that have JoinSplits in the Sprout case and/or Spend/Output descriptions in the Sapling case. These should be passed as fields in the Commit*Block requests.

Due to the coinbase maturity rules, the Sprout root is the empty root for the first 100 blocks. (These rules are already implemented in contextual validation and the anchor calculations.)

Hypothetically, if Sapling were activated from genesis, the specification requires a Sapling anchor, but zcashd would ignore that anchor.

These updates can be performed in a batch or without necessarily iterating over all transactions, if the data is available by other means; they're specified this way for clarity.

Accessing previous blocks for contextual validation

The state service performs contextual validation of blocks received via the CommitBlock request. Since CommitBlock is synchronous, contextual validation must also be performed synchronously.

The relevant chain for a block starts at its previous block, and follows the chain of previous blocks back to the genesis block.

Relevant chain iterator

The relevant chain can be retrieved from the state service as follows:

  • if the previous block is the finalized tip:
    • get recent blocks from the finalized state
  • if the previous block is in the non-finalized state:
    • get recent blocks from the relevant chain, then
    • get recent blocks from the finalized state, if required

The relevant chain can start at any non-finalized block, or at the finalized tip.

Relevant chain implementation

The relevant chain is implemented as a StateService iterator, which returns Arc<Block>s.

The chain iterator implements ExactSizeIterator, so Zebra can efficiently assert that the relevant chain contains enough blocks to perform each contextual validation check.


#![allow(unused)]
fn main() {
impl StateService {
    /// Return an iterator over the relevant chain of the block identified by
    /// `hash`.
    ///
    /// The block identified by `hash` is included in the chain of blocks yielded
    /// by the iterator.
    pub fn chain(&self, hash: block::Hash) -> Iter<'_> { ... }
}

impl Iterator for Iter<'_>  {
    type Item = Arc<Block>;
    ...
}
impl ExactSizeIterator for Iter<'_> { ... }
impl FusedIterator for Iter<'_> {}
}

For further details, see PR 1271.

Request / Response API

The state API is provided by a pair of Request/Response enums. Each Request variant corresponds to particular Response variants, and it's fine (and encouraged) for caller code to unwrap the expected variants with unreachable! on the unexpected variants. This is slightly inconvenient but it means that we have a unified state interface with unified backpressure.

This API includes both write and read calls. Spotting Commit requests in code review should not be a problem, but in the future, if we need to restrict access to write calls, we could implement a wrapper service that rejects these, and export "read" and "write" frontends to the same inner service.

Request::CommitBlock


#![allow(unused)]
fn main() {
CommitBlock {
    block: Arc<Block>,
    sprout_anchor: sprout::tree::Root,
    sapling_anchor: sapling::tree::Root,
}
}

Performs contextual validation of the given block, committing it to the state if successful. Returns Response::Added(block::Hash) with the hash of the newly committed block or an error.

Request::CommitFinalizedBlock


#![allow(unused)]
fn main() {
CommitFinalizedBlock {
    block: Arc<Block>,
    sprout_anchor: sprout::tree::Root,
    sapling_anchor: sapling::tree::Root,
}
}

Commits a finalized block to the rocksdb state, skipping contextual validation. This is exposed for use in checkpointing, which produces in-order finalized blocks. Returns Response::Added(block::Hash) with the hash of the committed block if successful.

Request::Depth(block::Hash)

Computes the depth in the best chain of the block identified by the given hash, returning

  • Response::Depth(Some(depth)) if the block is in the best chain;
  • Response::Depth(None) otherwise.

Implemented by querying:

  • (non-finalized) the height_by_hash map in the best chain, and
  • (finalized) the height_by_hash tree

Request::Tip

Returns Response::Tip(block::Hash) with the current best chain tip.

Implemented by querying:

  • (non-finalized) the highest height block in the best chain
  • (finalized) the highest height block in the hash_by_height tree, if the non-finalized state is empty

Request::BlockLocator

Returns Response::BlockLocator(Vec<block::Hash>) with hashes starting from the current chain tip and reaching backwards towards the genesis block. The first hash is the best chain tip. The last hash is the tip of the finalized portion of the state. If the finalized and non-finalized states are both empty, the block locator is also empty.

This can be used by the sync component to request hashes of subsequent blocks.

Implemented by querying:

  • (non-finalized) the hash_by_height map in the best chain
  • (finalized) the hash_by_height tree.

Request::Transaction(transaction::Hash)

Returns

  • Response::Transaction(Some(Transaction)) if the transaction identified by the given hash is contained in the state;

  • Response::Transaction(None) if the transaction identified by the given hash is not contained in the state.

Implemented by querying:

  • (non-finalized) the tx_loc_by_hash map (to get the block that contains the transaction) of each chain starting with the best chain, and then find block that chain's blocks (to get the block containing the transaction data)
  • (finalized) the tx_loc_by_hash tree (to get the block that contains the transaction) and then block_header_by_height tree (to get the block containing the transaction data), if the transaction is not in any non-finalized chain

Request::Block(block::Hash)

Returns

  • Response::Block(Some(Arc<Block>)) if the block identified by the given hash is contained in the state;

  • Response::Block(None) if the block identified by the given hash is not contained in the state;

Implemented by querying:

  • (non-finalized) the height_by_hash of each chain starting with the best chain, then find block that chain's blocks (to get the block data)
  • (finalized) the height_by_hash tree (to get the block height) and then the block_header_by_height tree (to get the block data), if the block is not in any non-finalized chain

Request::AwaitSpendableUtxo { outpoint: OutPoint, spend_height: Height, spend_restriction: SpendRestriction }

Returns

  • Response::SpendableUtxo(transparent::Output)

Implemented by querying:

  • (non-finalized) if any Chains contain OutPoint in their created_utxos, return the Utxo for OutPoint;
  • (finalized) else if OutPoint is in utxos_by_outpoint, return the Utxo for OutPoint;
  • else wait for OutPoint to be created as described in RFC0004;

Then validating:

  • check the transparent coinbase spend restrictions specified in RFC0004;
  • if the restrictions are satisfied, return the response;
  • if the spend is invalid, drop the request (and the caller will time out).

Drawbacks

  • Restarts can cause zebrad to redownload up to the last one hundred blocks it verified in the best chain, and potentially some recent side-chain blocks.

  • The service interface puts some extra responsibility on callers to ensure it is used correctly and does not verify the usage is correct at compile time.

  • the service API is verbose and requires manually unwrapping enums

  • We do not handle reorgs the same way zcashd does, and could in theory need to delete our entire on disk state and resync the chain in some pathological reorg cases.

  • testnet rollbacks are infrequent, but possible, due to bugs in testnet releases. Each testnet rollback will require additional state service code.

Summary

Zcash nodes use a Proof of Work algorithm to reach consensus on the best chain. Valid blocks must reach a difficulty threshold, which is adjusted after every block. The difficulty adjustment calculations depend on the difficulties and times of recent blocks. So Zebra performs contextual validation RFC2 of difficulty adjustments as part of committing blocks to the state.

Motivation

The Zcash block difficulty adjustment is one of the core Zcash consensus rules. Zebra must implement this consensus rule to make sure that its cached chain state is consistent with the consensus of Zcash nodes.

Difficulty adjustment is also a significant part of Zcash's security guarantees. It ensures that the network continues to resist takeover attacks, even as the number of Zcash miners grows.

Difficulty adjustment also ensures that blocks are regularly spaced, which allows users to create and finalise transactions with short, consistent delays. These predictable delays contribute to Zcash's usability.

Definitions

Difficulty:

  • hash difficulty: An arbitrary ranking of blocks, based on their hashes. Defined as the hash of the block, interpreted as a big-endian 256-bit number. Numerically smaller difficulties are harder to generate.

  • difficulty threshold: The easiest valid hash difficulty for a block. Numerically lower thresholds are harder to satisfy.

  • difficulty filter: A block passes the difficulty filter if the hash difficulty is less than or equal to the difficulty threshold (based on the block's difficulty field).

  • block work: The approximate amount of work required for a miner to generate a block hash that passes the difficulty filter. The number of block header attempts and the mining time are proportional to the work value. Numerically higher work values represent longer processing times.

  • averaging window: The 17 most recent blocks in the relevant chain.

  • median block span: The 11 most recent blocks from a chosen tip, typically the relevant tip.

  • target spacing: 150 seconds per block before Blossom activation, 75 seconds per block from Blossom activation onwards.

  • adjusted difficulty: After each block is mined, the difficulty threshold of the next block is adjusted, to keep the block gap close to the target spacing.

  • mean target difficulty: The arithmetic mean of the difficulty thresholds of the blocks in the averaging window.

  • median timespan: The average number of seconds taken to generate the blocks in the averaging window. Calculated using the difference of median block spans in and after the averaging window, then damped and bounded.

  • target timespan: The target spacing for an averaging window's worth of blocks.

Consensus:

  • consensus rule: A protocol rule which all nodes must apply consistently, so they can converge on the same chain fork.

  • structural/semantic/contextual verification: as defined in RFC2.

State:

  • block chain: A sequence of valid blocks linked by inclusion of the previous block hash in the subsequent block. Chains are rooted at the genesis block and extend to a tip.

  • relevant chain: The relevant chain for a block starts at the previous block, and extends back to genesis.

  • relevant tip: The tip of the relevant chain.

  • non-finalized state: State data corresponding to blocks above the reorg limit. This data can change in the event of a chain reorg.

  • finalized state: State data corresponding to blocks below the reorg limit. This data cannot change in the event of a chain reorg.

  • non-finalized tips: The highest blocks in each non-finalized chain. These tips might be at different heights.

  • finalized tip: The highest block in the finalized state. The tip of the best chain is usually 100 blocks (the reorg limit) above the finalized tip. But it can be lower during the initial sync, and after a chain reorganization, if the new best chain is at a lower height.

Guide-level explanation

Zcash's difficulty consensus rules are similar to Bitcoin.

Each block contains a difficulty threshold in its header. The hash of the block header must be less than this difficulty threshold. (When interpreted as a 256-bit integer in big-endian byte order.) This context-free semantic verification check is performed by the BlockVerifier.

After each block, the difficulty threshold is adjusted so that the block gap is close to the target spacing. On average, harder blocks take longer to mine, and easier blocks take less time.

The adjusted difficulty for the next block is calculated using the difficulty thresholds and times of recent blocks. Zcash uses the most recent 28 blocks in the relevant chain in its difficulty adjustment calculations.

The difficulty adjustment calculations adjust the mean target difficulty, based on the difference between the median timespan and the target timespan. If the median timespan is less than the target timespan, the next block is harder to mine.

The StateService calculates the adjusted difficulty using the context from the relevant chain. The difficulty contextual verification check ensures that the difficulty threshold of the next block is equal to the adjusted difficulty for its relevant chain.

State service interface changes

Contextual validation accesses recent blocks. So we modify the internal state service interface to provide an abstraction for accessing recent blocks.

The relevant chain

The relevant chain consists of the ancestors of a block, starting with its parent block, and extending back to the genesis block.

In Zebra, recent blocks are part of the non-finalized state, which can contain multiple chains. Past the reorganization limit, Zebra commits a single chain to the finalized state.

The relevant chain can start at any block in the non-finalized state, or at the finalized tip. See RFC5 for details.

Contextual validation design

Contextual validation is performed synchronously by the state service, as soon as the state has:

  • received the semantically valid next block (via CommitBlock), and
  • committed the previous block.

The difficulty adjustment check calculates the correct adjusted difficulty threshold value for a candidate block, and ensures that the block's difficulty_threshold field is equal to that value.

This check is implemented as follows:

Difficulty adjustment

The block difficulty threshold is adjusted by scaling the mean target difficulty by the median timespan.

On Testnet, if a long time has elapsed since the previous block, the difficulty adjustment is modified to allow minimum-difficulty blocks.

Mean target difficulty

The mean target difficulty is the arithmetic mean of the difficulty thresholds of the PoWAveragingWindow (17) most recent blocks in the relevant chain.

Zcash uses block difficulty thresholds in its difficulty adjustment calculations. (Block hashes are not used for difficulty adjustment.)

Median timespan

The average number of seconds taken to generate the 17 blocks in the averaging window.

The median timespan is calculated by taking the difference of the median times for:

  • the relevant tip: the PoWMedianBlockSpan (11) most recent blocks, and
  • the 11 blocks after the 17-block PoWAveragingWindow: that is, blocks 18-28 behind the relevant tip.

The median timespan is damped by the PoWDampingFactor, and bounded by PoWMaxAdjustDown and PoWMaxAdjustUp.

Test network minimum difficulty blocks

If there is a large gap after a Testnet block, the next block becomes a minimum difficulty block. Testnet minimum difficulty blocks have their difficulty_threshold set to the minimum difficulty for Testnet.

Block difficulty threshold

The block difficulty threshold for the next block is calculated by scaling the mean target difficulty by the ratio between the median timespan and the averaging window timespan.

The result of this calculation is limited by ToCompact(PoWLimit(network)), a per-network minimum block difficulty. This minimum difficulty is also used when a Testnet block's time gap exceeds the minimum difficulty gap.

Reference-level explanation

Contextual validation

Contextual validation is implemented in StateService::check_contextual_validity, which calls a separate function for each contextual validity check.

In Zebra, contextual validation starts after Canopy activation, so we can assume that the relevant chain contains at least 28 blocks on Mainnet and Testnet. (And panic if this assumption does not hold at runtime.)

Fundamental data types

Zebra is free to implement its difficulty calculations in any way that produces equivalent results to zcashd and the Zcash specification.

Difficulty

In Zcash block headers, difficulty thresholds are stored as a "compact" nBits value, which uses a custom 32-bit floating-point encoding. Zebra calls this type CompactDifficulty.

In Zcash, difficulty threshold calculations are performed using unsigned 256-bit integers. Rust has no standard u256 type, but there are a number of crates available which implement the required operations on 256-bit integers. Zebra abstracts over the chosen u256 implementation using its ExpandedDifficulty type.

Time

In Zcash, time values are unsigned 32-bit integers. But the difficulty adjustment calculations include time subtractions which could overflow an unsigned type, so they are performed using signed 64-bit integers in zcashd.

Zebra parses the header.time field into a DateTime<Utc>. Conveniently, the chrono::DateTime<_>::timestamp() function returns i64 values. So Zebra can do its signed time calculations using i64 values internally.

Note: i32 is an unsuitable type for signed time calculations. It is theoretically possible for the time gap between blocks to be larger than i32::MAX, because those times are provided by miners. Even if the median time gap is that large, the bounds and minimum difficulty in Zcash's difficulty adjustment algorithm will preserve a reasonable difficulty threshold. So Zebra must support this edge case.

Consensus-Critical Operations

The order of operations and overflow semantics for 256-bit integers can be consensus-critical.

For example:

  • dividing before multiplying discards lower-order bits, but
  • multiplying before dividing can cause overflow.

Zebra's implementation should try to match zcashd's order of operations and overflow handling as closely as possible.

Difficulty adjustment check

The difficulty adjustment check calculates the correct difficulty threshold value for a candidate block, and ensures that the block's difficulty_threshold field is equal to that value.

Context data type

The difficulty adjustment functions use a context consisting of the difficulties and times from the previous 28 blocks in the relevant chain.

These functions also use the candidate block's height and network.

To make these functions more ergonomic, we create a AdjustedDifficulty type, and implement the difficulty adjustment calculations as methods on that type.


#![allow(unused)]
fn main() {
/// The averaging window for difficulty threshold arithmetic mean calculations.                               
///                                                                                                           
/// `PoWAveragingWindow` in the Zcash specification.                                                          
pub const POW_AVERAGING_WINDOW: usize = 17;

/// The median block span for time median calculations.                                                       
///                                                                                                           
/// `PoWMedianBlockSpan` in the Zcash specification.                                                          
pub const POW_MEDIAN_BLOCK_SPAN: usize = 11;

/// Contains the context needed to calculate the adjusted difficulty for a block. 
struct AdjustedDifficulty {
    candidate_time: DateTime<Utc>,
    candidate_height: block::Height,
    network: Network,
    relevant_difficulty_thresholds: [CompactDifficulty; POW_AVERAGING_WINDOW + POW_MEDIAN_BLOCK_SPAN],
    relevant_times: [DateTime<Utc>; POW_AVERAGING_WINDOW + POW_MEDIAN_BLOCK_SPAN],
}
}

We implement some initialiser methods on AdjustedDifficulty for convenience. We might want to validate downloaded headers in future, so we include a new_from_header initialiser.


#![allow(unused)]
fn main() {
/// Initialise and return a new `AdjustedDifficulty` using a `candidate_block`,
/// `network`, and a `context`.
///
/// The `context` contains the previous
/// `PoWAveragingWindow + PoWMedianBlockSpan` (28) `difficulty_threshold`s and
/// `time`s from the relevant chain for `candidate_block`, in reverse height
/// order, starting with the previous block.
///
/// Note that the `time`s might not be in reverse chronological order, because
/// block times are supplied by miners.
///
/// Panics:
/// If the `context` contains fewer than 28 items.
pub fn new_from_block<C>(candidate_block: &Block,
                         network: Network,
                         context: C)
                         -> AdjustedDifficulty
    where
        C: IntoIterator<Item = (CompactDifficulty, DateTime<Utc>)>,
    { ... }

/// Initialise and return a new `AdjustedDifficulty` using a
/// `candidate_header`, `previous_block_height`, `network`, and a `context`.
///
/// Designed for use when validating block headers, where the full block has not
/// been downloaded yet.
///
/// See `new_from_block` for detailed information about the `context`.
///
/// Panics:
/// If the context contains fewer than 28 items.
pub fn new_from_header<C>(candidate_header: &block::Header,
                          previous_block_height: block::Height,
                          network: Network,
                          context: C)
                          -> AdjustedDifficulty
    where
        C: IntoIterator<Item = (CompactDifficulty, DateTime<Utc>)>,
    { ... }
}

Memory usage note

Copying CompactDifficulty values into the AdjustedDifficulty struct uses less memory than borrowing those values. CompactDifficulty values are 32 bits, but pointers are 64-bit on most modern machines. (And since they all come from different blocks, we need a pointer to each individual value.)

Borrowing DateTime<Utc> values might use slightly less memory than copying them - but that depends on the exact way that Rust stores associated types derived from a generic argument.

In any case, the overall size of each AdjustedDifficulty is only a few hundred bytes. If it turns up in profiles, we can look at borrowing the block header data.

Difficulty adjustment check implementation

The difficulty adjustment check ensures that the candidate_difficulty_threshold is equal to the difficulty_threshold value calculated using AdjustedDifficulty::adjusted_difficulty_threshold.

We implement this function:


#![allow(unused)]
fn main() {
/// Validate the `difficulty_threshold` from a candidate block's header, based
/// on an `expected_difficulty` for that block.
///
/// Uses `expected_difficulty` to calculate the expected `ToCompact(Threshold())`
/// value, then compares that value to the `difficulty_threshold`. Returns
/// `Ok(())` if the values are equal.
pub fn difficulty_threshold_is_valid(difficulty_threshold: CompactDifficulty,
                                     expected_difficulty: AdjustedDifficulty)
                                     -> Result<(), ValidateContextError> { ... }
}

Mean target difficulty calculation

The mean target difficulty is the arithmetic mean of the difficulty thresholds of the PoWAveragingWindow (17) most recent blocks in the relevant chain.

We implement this method on AdjustedDifficulty:


#![allow(unused)]
fn main() {
/// Calculate the arithmetic mean of the averaging window thresholds: the
/// expanded `difficulty_threshold`s from the previous `PoWAveragingWindow` (17)
/// blocks in the relevant chain.
///
/// Implements `MeanTarget` from the Zcash specification.
fn mean_target_difficulty(&self) -> ExpandedDifficulty { ... }
}

Implementation notes

Since the PoWLimits are 2^251 − 1 for Testnet, and 2^243 − 1 for Mainnet, the sum of these difficulty thresholds will be less than or equal to (2^251 − 1)*17 = 2^255 + 2^251 - 17. Therefore, this calculation can not overflow a u256 value. So the function is infalliable.

In Zebra, contextual validation starts after Canopy activation, so we can assume that the relevant chain contains at least 17 blocks. Therefore, the PoWLimit case of MeanTarget() in the Zcash specification is unreachable.

Median timespan calculation

The median timespan is the difference of the median times for:

  • the relevant tip: the PoWMedianBlockSpan (11) most recent blocks, and
  • the 11 blocks after the 17-block PoWAveragingWindow: that is, blocks 18-28 behind the relevant tip.

(The median timespan is known as the ActualTimespan in the Zcash specification, but this terminology is confusing, because it is a difference of medians, rather than any "actual" elapsed time.)

Zebra implements the median timespan using the following methods on AdjustedDifficulty:


#![allow(unused)]
fn main() {
/// Calculate the bounded median timespan. The median timespan is the
/// difference of medians of the timespan times, which are the `time`s from
/// the previous `PoWAveragingWindow + PoWMedianBlockSpan` (28) blocks in the
/// relevant chain.
///
/// Uses the candidate block's `height' and `network` to calculate the
/// `AveragingWindowTimespan` for that block.
///
/// The median timespan is damped by the `PoWDampingFactor`, and bounded by
/// `PoWMaxAdjustDown` and `PoWMaxAdjustUp`.
///
/// Implements `ActualTimespanBounded` from the Zcash specification.
///
/// Note: This calculation only uses `PoWMedianBlockSpan` (11) times at the
/// start and end of the timespan times. timespan times `[11..=16]` are ignored.
fn median_timespan_bounded(&self) -> Duration { ... }

/// Calculate the median timespan. The median timespan is the difference of
/// medians of the timespan times, which are the `time`s from the previous
/// `PoWAveragingWindow + PoWMedianBlockSpan` (28) blocks in the relevant chain.
///
/// Implements `ActualTimespan` from the Zcash specification.
///
/// See `median_timespan_bounded` for details.
fn median_timespan(&self) -> Duration { ... }

/// Calculate the median of the `median_block_span_times`: the `time`s from a
/// slice of `PoWMedianBlockSpan` (11) blocks in the relevant chain.
///
/// Implements `MedianTime` from the Zcash specification.
fn median_time(mut median_block_span_times: [DateTime<Utc>; POW_MEDIAN_BLOCK_SPAN])
               -> DateTime<Utc> { ... }
}

Zebra implements the AveragingWindowTimespan using the following methods on NetworkUpgrade:


#![allow(unused)]
fn main() {
impl NetworkUpgrade {
    /// Returns the `AveragingWindowTimespan` for the network upgrade.
    pub fn averaging_window_timespan(&self) -> Duration { ... }

    /// Returns the `AveragingWindowTimespan` for `network` and `height`.
    pub fn averaging_window_timespan_for_height(network: Network,
                                                height: block::Height)
                                                -> Duration { ... }
}
}

Implementation notes

In Zebra, contextual validation starts after Canopy activation, so we can assume that the relevant chain contains at least 28 blocks. Therefore:

  • max(0, height − PoWMedianBlockSpan) in the MedianTime() calculation simplifies to height − PoWMedianBlockSpan, and
  • there is always an odd number of blocks in MedianTime(), so the median is always the exact middle of the sequence.

Therefore, the function is infalliable.

Test network minimum difficulty calculation

A block is a Testnet minimum difficulty block if:

  • the block is a Testnet block,
  • the block's height is 299188 or greater, and
  • the time gap from the previous block is greater than the Testnet minimum difficulty gap, which is 6 times the target spacing for the block's height. (The target spacing was halved from the Blossom network upgrade onwards.)

The difficulty adjustment is modified for Testnet minimum difficulty blocks as follows:

  • the difficulty threshold in the block header is set to the Testnet minimum difficulty threshold, ToCompact(PoWLimit(network)).

Since the new difficulty changes the block header, Testnet blocks can only satisfy one of the alternate difficulty adjustment rules:

  • if the time gap is less than or equal to the Testnet minimum difficulty gap: the difficulty threshold is calculated using the default difficulty adjustment rule,
  • if the time gap is greater than the Testnet minimum difficulty gap: the difficulty threshold is the Testnet minimum difficulty threshold.

See ZIP-208 for details.

Note: some older versions of ZIPs 205 and 208 incorrectly said that:

  • the time gap threshold uses an "at least" check (it is strictly greater than),
  • the minimum difficulty threshold value was PoWLimit (it is ToCompact(PoWLimit)),
  • the difficulty_threshold (nBits) field is not modified in Testnet minimum difficulty blocks (the field is modified), and
  • the Testnet minimum difficulty value is not used to calculate future difficulty adjustments (the modified value is used in future adjustments).

ZIP 205 and 208 were fixed on 14 November 2020, see ZIP PR 417 and ZIP commit 806076c for details.

Test network minimum difficulty implementation

The Testnet minimum difficulty calculation uses the existing NetworkUpgrade::minimum_difficulty_spacing_for_height function to calculate the minimum difficulty gap.

We implement this method on NetworkUpgrade:


#![allow(unused)]
fn main() {
/// Returns true if the gap between `block_time` and `previous_block_time` is                             
/// greater than the Testnet minimum difficulty time gap. This time gap                                   
/// depends on the `network` and `block_height`.                                                          
///                                                                                                       
/// Returns false on Mainnet, when `block_height` is less than the minimum                                
/// difficulty start height, and when the time gap is too small.                                          
///                                                                                                       
/// `block_time` can be less than, equal to, or greater than                                              
/// `previous_block_time`, because block times are provided by miners.                                    
///                                                                                                       
/// Implements the Testnet minimum difficulty adjustment from ZIPs 205 and 208.                           
///                                                                                                       
/// Spec Note: Some parts of ZIPs 205 and 208 previously specified an incorrect                           
/// check for the time gap. This function implements the correct "greater than"                           
/// check.                                                                                                
pub fn is_testnet_min_difficulty_block(
    network: Network,
    block_height: block::Height,
    block_time: DateTime<Utc>,
    previous_block_time: DateTime<Utc>,
) -> bool { ... }
}

Implementation notes

In Zcash, the Testnet minimum difficulty rule starts at block 299188, and in Zebra, contextual validation starts after Canopy activation. So we can assume that there is always a previous block.

Therefore, this function is infalliable.

Block difficulty threshold calculation

The block difficulty threshold for the next block is calculated by scaling the mean target difficulty by the ratio between the median timespan and the averaging window timespan.

The result of the scaled threshold calculation is limited by ToCompact(PoWLimit(network)), a per-network minimum block difficulty. This minimum difficulty is also used when a Testnet block's time gap exceeds the minimum difficulty gap. We use the existing ExpandedDifficulty::target_difficulty_limit function to calculate the value of ToCompact(PoWLimit(network)).

In Zebra, contextual validation starts after Canopy activation, so the genesis case of Threshold() in the Zcash specification is unreachable.

Block difficulty threshold implementation

We implement these methods on AdjustedDifficulty:


#![allow(unused)]
fn main() {
/// Calculate the expected `difficulty_threshold` for a candidate block, based
/// on the `candidate_time`, `candidate_height`, `network`, and the
/// `difficulty_threshold`s and `time`s from the previous
/// `PoWAveragingWindow + PoWMedianBlockSpan` (28) blocks in the relevant chain.
///
/// Implements `ThresholdBits` from the Zcash specification, and the Testnet
/// minimum difficulty adjustment from ZIPs 205 and 208.
pub fn expected_difficulty_threshold(&self) -> CompactDifficulty { ... }

/// Calculate the `difficulty_threshold` for a candidate block, based on the
/// `candidate_height`, `network`, and the relevant `difficulty_threshold`s and
/// `time`s.
///
/// See `expected_difficulty_threshold` for details.
///
/// Implements `ThresholdBits` from the Zcash specification. (Which excludes the
/// Testnet minimum difficulty adjustment.)
fn threshold_bits(&self) -> CompactDifficulty { ... }
}

Implementation notes

Since:

  • the PoWLimits are 2^251 − 1 for Testnet, and 2^243 − 1 for Mainnet,
  • the ActualTimespanBounded can be at most MaxActualTimespan, which is floor(PoWAveragingWindow * PoWTargetSpacing * (1 + PoWMaxAdjustDown)) or floor(17 * 150 * (1 + 32/100)) = 3366,
  • AveragingWindowTimespan is at most 17 * 150 = 2250, and
  • MeanTarget is at most PoWLimit, ...

The maximum scaled value inside the Threshold() calculation is:

  • floor(PoWLimit / 2250) * 3366, which equals
  • floor((2^251 − 1) / 2250) * 3366, which equals
  • (2^251 − 1) * 132/100,
  • which is less than 2^252.

Therefore, this calculation can not overflow a u256 value. (And even if it did overflow, it would be constrained to a valid value by the PoWLimit minimum.)

Note that the multiplication by ActualTimespanBounded must happen after the division by AveragingWindowTimespan. Performing the multiplication first could overflow.

If implemented in this way, the function is infalliable.

zcashd truncates the MeanTarget after the mean calculation, and after dividing by AveragingWindowTimespan. But as long as there is no overflow, this is equivalent to the single truncation of the final result in the Zcash specification. However, Zebra should follow the order of operations in zcashd, and use repeated divisions, because that can't overflow. See the relevant comment in the zcashd source code.

Module Structure

The structs and functions in this RFC are implemented in a new zebra_state::service::check::difficulty module.

This module has two entry points:

  • DifficultyAdjustment::new_from_block
  • difficulty_threshold_is_valid

These entry points are both called from StateService::check_contextual_validity.

Test Plan

Explain how the feature will be tested, including:

  • tests for consensus-critical functionality
  • existing test vectors, if available
  • Zcash blockchain block test vectors (specify the network upgrade, feature, or block height and network)
  • property testing or fuzzing

The tests should cover:

  • positive cases: make sure the feature accepts valid inputs
    • using block test vectors for each network upgrade provides some coverage of valid inputs
  • negative cases: make sure the feature rejects invalid inputs
    • make sure there is a test case for each error condition in the code
    • if there are lots of potential errors, prioritise:
      • consensus-critical errors
      • security-critical errors, and
      • likely errors
  • edge cases: make sure that boundary conditions are correctly handled

Drawbacks

Why should we not do this?

Alternate consensus parameters

Any alternate consensus parameters or regtest mode would have to respect the constraints set by this design.

In particular:

  • the PoWLimit must be less than or equal to (2^256 - 1) / PoWAveragingWindow (approximately 2^251) to avoid overflow,
  • the PoWAveragingWindow and PoWMedianBlockSpan are fixed by function argument types (at least until Rust gets stable const generics), and
  • the design eliminates a significant number of edge cases by assuming that difficulty adjustments aren't validated for the first PoWAveragingWindow + PoWMedianBlockSpan (28) blocks in the chain.

Rationale and alternatives

Is this design a good basis for later designs or implementations?

The design includes specific methods for a future header-only validation design.

What other designs have been considered and what is the rationale for not choosing them?

A previous version of the RFC did not have the AdjustedDifficulty struct and methods. That design was easy to misuse, because each function had a complicated argument list.

What is the impact of not doing this?

Zebra could accept invalid, low-difficulty blocks from arbitrary miners. That would be a security issue.

Prior art

  • zcashd
  • the Zcash specification
  • Bitcoin

Unresolved questions

  • What parts of the design do you expect to resolve through the implementation of this feature before stabilization?

    • Guide-level examples
    • Reference-level examples
    • Corner case examples
    • Testing

  • What related issues do you consider out of scope for this RFC that could be addressed in the future independently of the solution that comes out of this RFC?

    • Monitoring and maintenance

Future possibilities

Re-using the relevant chain API in other contextual checks

The relevant chain iterator can be re-used to implement other contextual validation checks.

For example, responding to peer requests for block locators, which means implementing relevant chain hash queries as a StateService request

Header-only difficulty adjustment validation

Implementing header-only difficulty adjustment validation as a StateService request.

Caching difficulty calculations

Difficulty calculations use u256 could be a bit expensive, particularly if we get a flood of low-difficulty blocks. To reduce the impact of this kind of DoS, we could cache the value returned by threshold_bits for each block in the non-finalized state, and for the finalized tip. This value could be used to quickly calculate the difficulties for any child blocks of these blocks.

There's no need to persist this cache, or pre-fill it. (Minimum-difficulty Testnet blocks don't call threshold_bits, and some side-chain blocks will never have a next block.)

This caching is only worth implementing if these calculations show up in zebrad profiles.

Summary

The zebra-client crate handles client functionality. Client functionality is defined as all functionality related to a particular user's private data, in contrast to the other full node functionality which handles public chain state. This includes:

  • note and key management;
  • transaction generation;
  • a client component for zebrad that handles block chain scanning, with appropriate side-channel protections;
  • an RPC endpoint for zebrad that allows access to the client component;
  • Rust library code that implements basic wallet functionality;
  • a zebra-cli binary that wraps the wallet library and RPC queries in a command-line interface.

Client functionality is restricted to transparent and Sapling shielded transactions; Sprout shielded transactions are not supported. (Users should migrate to Sapling).

Motivation

We want to allow users to efficiently and securely send and receive funds via Zebra. One challenge unique to Zcash is block chain scanning: because shielded transactions reveal no metadata about the sender or receiver, users must scan the block chain for relevant transactions using viewing keys. This means that unlike a transparent blockchain with public transactions, a full node must have online access to viewing keys to scan the chain. This creates the risk of a privacy leak, because the node should not reveal which viewing keys it has access to.

Block chain scanning requires a mechanism that allows users to manage and store key material. This mechanism should also provide basic wallet functionality, so that users can send and receive funds without requiring third-party software.

To protect user privacy, this and all secret-dependent functionality should be strongly isolated from the rest of the node implementation. Care should be taken to protect against side channels that could reveal information about viewing keys. To make this isolation easier, all secret-dependent functionality is provided only by the zebra-client crate.

Definitions

  • client functionality: all functionality related to a particular user's private data, in contrast to other full node functionality which handles public chain state.

  • block chain scanning: the process of scanning the block chain for relevant transactions using a viewing key, as described in §4.19 of the protocol specification.

  • viewing key: Sapling shielded addresses support viewing keys, which represent the capability to decrypt transactions, as described in §3.1 and §4.2.2 of the protocol specification.

  • task: In this document, task refers specifically to a Tokio task. In brief, a task is a light weight, non-blocking unit of execution (green thread), similar to a Goroutine or Erlang process. Tasks execute independently and are scheduled co-operatively using explicit yield points. Tasks are executed on the Tokio runtime, which can either be single- or multi-threaded.

Guide-level explanation

There are two main parts of this functionality. The first is a Client component running as part of zebrad, and the second is a zebra-cli command-line tool.

The Client component is responsible for blockchain scanning. It maintains its own distinct sled database, which stores the viewing keys it uses to scan as well as the results of scanning. When a new block is added to the chain state, the Client component is notified asynchronously using a channel. For each Sapling shielded transaction in the block, the component attempts to perform trial decryption of that transaction's notes using each registered viewing key, as described in §4.19. If successful, decrypted notes are saved to the database.

The PING/REJECT attack demonstrates the importance of decoupling execution of normal node operations from secret-dependent operations. Zebra's network stack already makes it immune to those particular attacks, because each peer connection is executed in a different task. However, to eliminate this entire class of vulnerability, we execute the Client component in its own task, decoupled from the rest of the node functionality. In fact, each viewing key's scanning is performed independently, as described in more detail below, with an analysis of potential side-channels.

The second part is the zebra-cli command-line tool, which provides basic wallet functionality. This tool manages spending keys and addresses, and communicates with the Client component in zebrad to provide basic wallet functionality. Specifically, zebra-cli uses a distinct RPC endpoint to load viewing keys into zebrad and to query the results of block chain scanning. zebra-cli can then use the results of those queries to generate transactions and submit them to the network using zebrad.

This design upholds the principle of least authority by separating key material required for spending funds from the key material required for block chain scanning. This allows compartmentalization. For instance, a user could in principle run zebrad on a cloud VPS with only their viewing keys and store their spending keys on a laptop, or a user could run zebrad on a local machine and store their spending keys in a hardware wallet. Both of these use cases would require some additional tooling support, but are possible with this design.

Reference-level explanation

State notifications

We want a way to subscribe to updates from the state system via a channel. For the purposes of this RFC, these changes are in-flight, but in the future, these could be used for a push-based RPC mechanism.

Subscribers can subscribe to all state change notifications as they come in.

Currently the zebra_state::init() method returns a BoxService that allows you to make requests to the chain state. Instead, we would return a (BoxService, StateNotifications) tuple, where StateNotifications is a new structure initially defined as:

#[non_exhaustive]
pub struct StateNotifications {
  pub new_blocks: tokio::sync::watch::Receiver<Arc<Block>>,
}

Instead of making repeated polling requests to a state service to look for any new blocks, this channel will push new blocks to a consumer as they come in, for the consumer to use or discard at their discretion. This will be used by the client component described below. This will also be needed for gossiping blocks to other peers, as they are validated.

Online client component

This component maintains its own Sled tree. See RFC#0005 for more details on Sled.

We use the following Sled trees:

TreeKeysValues
viewing_keysIncomingViewingKeyString
height_by_keyIncomingViewingKeyBE32(height)
received_set_by_keyIncomingViewingKey?
spend_set_by_keyIncomingViewingKey?
nullifier_map_by_keyIncomingViewingKey?

See https://zips.z.cash/protocol/protocol.pdf#saplingscan

Zcash structures are encoded using ZcashSerialize/ZcashDeserialize.

This component runs inside zebrad. After incoming viewing keys are registered, it holds onto them in order to do blockchain scanning. The component keeps track of where it’s scanned to (TODO: per key?). Runs in its own separate task, in case it crashes, it’s not noticeable, and executes independently (but in the same process) of the normal node operation.

In the case of the client component that needs to do blockchain scanning and trial decryption, every valid block with non-coinbase transactions will need to be checked and its transactions trial-decrypted with registered incoming viewing keys to see if any notes have been received by the key's owner and if any notes have already been spent elsewhere.

RPC's

A specific set of privileged RPC endpoints:

  • Allows registering of incoming viewing keys with zebrad in order to do blockchain scanning
  • Allows querying of the results of that scanning, to get wallet balance, etc
  • Not authenticated to start (see 'Future possibilities')
  • Users can control access by controlling access to the privileged endpoint (ie via a firewall)

Support for sending tx's via non-privileged RPC endpoints, or via Stolon:

  • sendTransaction: once you author a transcation you can gossip it via any Zcash node, not just a specific instance of zebrad

Wallet functionality

  • Holds on to your spending keys so you can author transactions
  • Uses RPC methods to query the online client component inside zebrad about wallet balances

CLI binary

  • zebra-cli talks to the subcomponent running in zebrad
    • (can use servo/bincode to communicate with zebrad)
    • via the privileged (and possibly the unprivileged) RPC endpoints
    • can use cap-std to restrict filesystem and network access for zebra-client. See https://github.com/ZcashFoundation/zebra/issues/2340
    • can use the tui crate to render a terminal UI

Task isolation in Tokio

  • TODO: fill in
  • cooperative multitasking is fine, IF you cooperate
  • lots of tasks

Module Structure

zebra-client ( currently and empty stub) zebra-cli (does not exist yet) zebra-rfc? (exists as an empty stub, we way have zebra-cli communicate with zebra-client inside zebrad via an RPC method any/or a private IPC layer)

Test Plan

Drawbacks

Supporting a wallet assumes risk. Effort required to implement wallet functionality.

  • need to responsibly handle secret key material;
  • currently we only handle public data.

Rationale and alternatives

  • why have a separate RPC endpoint?

    • extra endpoints are cheap
    • allows segmentation by capability
    • alternative is error-prone after-the-fact ACLs like Tor control port filters

  • What is the impact of not doing this?

    • We can't send money with zebra alone.
    • rely on third party wallet software to send funds with zebra
      • we need to provide basic functionality within zebra's trust boundary, rather than forcing users to additionally trust 3p software
      • there are great 3p wallets, we want to integrate with them, just don't want to rely on them

  • What about the light client protocol?

    • does not address this use case, has different trust model (private lookup, no scanning)
    • we want our first client that interacts with zebrad to not have a long startup time, which a lightclient implementation would require
    • zebra-cli should be within the same trust and privacy boundary as the zebrad node it is interacting with
    • light client protocol as currently implemented requires stack assumptions such as protobufs and a hardcoded lightserver to talk to

  • What about having one database per key?

    • easy to reliably delete or backup all data related to a single key
    • might use slightly more space/CPU
    • slightly harder to delete all the keys

Unresolved questions

  • wait to fill this in until doing the detailed writeup.

Future possibilities

  • BlazeSync algorithm for fast syncing, like Zecwallet

  • mandatory sweeps for legacy keys

    • blazingly fast wallet startup, to match zebrad's blazingly fast sync
    • generate unified address from a new seed phrase (or one provided by the user)
    • user can just backup seed phrase rather than a set of private keys
    • handles arbitrary keys from zcashd and other wallets, even if they weren't generated from a seed phrase
    • ~handles Sprout funds without zebra-client having to support Sprout balances~
    • startup is incredibly fast
      • sweep takes a few minutes to be confirmed
      • scanning the entire chain could take hours
      • if we know when the seed phrase was created, we can skip millions of blocks during scanning
    • sweeps can also be initiated by the user for non-linkability / performance / refresh
    • sweeps should handle the "block reward recipient" case where there are a lot of small outputs
    • initial release could support mandatory sweeps, and future releases could support legacy keys

  • split Client component into subprocess

    • this helps somewhat but the benefit is reduced by our preexisting memory safety, thanks to Rust
    • not meaningful without other isolation (need to restrict zebrad from accessing viewing keys on disk, etc)
    • could use cap-std to restrict filesystem and network access for zebra-client. See https://github.com/ZcashFoundation/zebra/issues/2340
    • instead of process isolation, maybe you actually want the Light Client Protocol, or something similar?

  • hardware wallet integration for zebra-cli

    • having zebra-cli allows us to do this
    • much higher security ROI than subprocess
    • very cool future feature

  • authenticate queries for a particular viewing key by proving knowledge of the viewing key (requires crypto). this could allow public access to the client endpoint

  • Use Unified Addresses only, no legacy addrs.

Summary

Network Upgrade number 5 (NU5) introduces a new transaction type (transaction version 5). This document is a proposed design for implementing such a transaction version.

Motivation

The Zebra software wants to be a protocol compatible Zcash implementation. One of the tasks to do this includes the support of the new version 5 transactions that will be implemented in Network Upgrade 5 (NU5).

Definitions

  • NU5 - the 5th Zcash network upgrade, counting from the Overwinter upgrade as upgrade zero.
  • Orchard - a new shielded pool introduced in NU5.
  • Sapling - a new shielded pool introduced in the 1st network upgrade. (Sapling is also the name of that network upgrade, but this RFC is focused on the Sapling shielded pool.)
  • orchard data - Data types needed to support orchard transactions.
  • sapling data - Data types needed to support sapling transactions.
  • orchard transaction version - Transactions that support orchard data. Currently only V5.
  • sapling transaction version - Transactions that support sapling data. Currently V4 and V5 but the data is implemented differently in them.

Guide-level explanation

V5 transactions are described by the protocol in the second table of Transaction Encoding and Consensus.

All of the changes proposed in this document are only to the zebra-chain crate.

To highlight changes most of the document comments from the code snippets in the reference section were removed.

Sapling Changes Overview

V4 and V5 transactions both support sapling, but the underlying data structures are different. So we need to make the sapling data types generic over the V4 and V5 structures.

In V4, anchors are per-spend, but in V5, they are per-transaction. In V5, the shared anchor is only present if there is at least one spend.

For consistency, we also move some fields into the ShieldedData type, and rename some fields and types.

Orchard Additions Overview

V5 transactions are the only ones that will support orchard transactions with Orchard data types.

Orchard uses Halo2Proofs with corresponding signature type changes. Each Orchard Action contains a spend and an output. Placeholder values are substituted for unused spends and outputs.

Other Transaction V5 Changes

V5 transactions split Spends, Outputs, and AuthorizedActions into multiple arrays, with a single CompactSize count before the first array. We add new zcash_deserialize_external_count and zcash_serialize_external_count utility functions, which make it easier to serialize and deserialize these arrays correctly.

The order of some of the fields changed from V4 to V5. For example the lock_time and expiry_height were moved above the transparent inputs and outputs.

The serialized field order and field splits are in the V5 transaction section in the NU5 spec. (Currently, the V5 spec is on a separate page after the V1-V4 specs.)

Zebra's structs sometimes use a different order from the spec. We combine fields that occur together, to make it impossible to represent structurally invalid Zcash data.

In general:

  • Zebra enums and structs put fields in serialized order.
  • Composite structs and emnum variants are ordered based on last data deserialized for the composite.

Reference-level explanation

Sapling Changes

We know by protocol (2nd table of Transaction Encoding and Consensus) that V5 transactions will support sapling data however we also know by protocol that spends (Spend Description Encoding and Consensus, See †) and outputs (Output Description Encoding and Consensus, See †) fields change from V4 to V5.

ShieldedData is currently defined and implemented in zebra-chain/src/transaction/shielded_data.rs. As this is Sapling specific we propose to move this file to zebra-chain/src/sapling/shielded_data.rs.

Changes to V4 Transactions

Here we have the proposed changes for V4 transactions:

  • make sapling_shielded_data use the PerSpendAnchor anchor variant
  • rename shielded_data to sapling_shielded_data
  • move value_balance into the sapling::ShieldedData type
  • order fields based on the last data deserialized for each field

#![allow(unused)]
fn main() {
enum Transaction::V4 {
    inputs: Vec<transparent::Input>,
    outputs: Vec<transparent::Output>,
    lock_time: LockTime,
    expiry_height: block::Height,
    joinsplit_data: Option<JoinSplitData<Groth16Proof>>,
    sapling_shielded_data: Option<sapling::ShieldedData<PerSpendAnchor>>,
}
}

The following types have ZcashSerialize and ZcashDeserialize implementations, because they can be serialized into a single byte vector:

  • transparent::Input
  • transparent::Output
  • LockTime
  • block::Height
  • Option<JoinSplitData<Groth16Proof>>

Note: Option<sapling::ShieldedData<PerSpendAnchor>> does not have serialize or deserialize implementations, because the binding signature is after the joinsplits. Its serialization and deserialization is handled as part of Transaction::V4.

Anchor Variants

We add an AnchorVariant generic type trait, because V4 transactions have a per-Spend anchor, but V5 transactions have a shared anchor. This trait can be added to sapling/shielded_data.rs:


#![allow(unused)]
fn main() {
struct PerSpendAnchor {}
struct SharedAnchor {}

/// This field is not present in this transaction version.
struct FieldNotPresent;

impl AnchorVariant for PerSpendAnchor {
    type Shared = FieldNotPresent;
    type PerSpend = sapling::tree::Root;
}

impl AnchorVariant for SharedAnchor {
    type Shared = sapling::tree::Root;
    type PerSpend = FieldNotPresent;
}

trait AnchorVariant {
    type Shared;
    type PerSpend;
}
}

Changes to Sapling ShieldedData

We use AnchorVariant in ShieldedData to model the anchor differences between V4 and V5:

  • in V4, there is a per-spend anchor
  • in V5, there is a shared anchor, which is only present when there are spends

If there are no spends and no outputs:

  • in v4, the value_balance is fixed to zero
  • in v5, the value balance field is not present
  • in both versions, the binding_sig field is not present

#![allow(unused)]
fn main() {
/// ShieldedData ensures that value_balance and binding_sig are only present when
/// there is at least one spend or output.
struct sapling::ShieldedData<AnchorV: AnchorVariant> {
    value_balance: Amount,
    transfers: sapling::TransferData<AnchorV>,
    binding_sig: redjubjub::Signature<Binding>,
}

/// TransferData ensures that:
/// * there is at least one spend or output, and
/// * the shared anchor is only present when there are spends
enum sapling::TransferData<AnchorV: AnchorVariant> {
    /// In Transaction::V5, if there are any spends,
    /// there must also be a shared spend anchor.
    SpendsAndMaybeOutputs {
        shared_anchor: AnchorV::Shared,
        spends: AtLeastOne<Spend<AnchorV>>,
        maybe_outputs: Vec<Output>,
    }

    /// If there are no spends, there must not be a shared
    /// anchor.
    JustOutputs {
        outputs: AtLeastOne<Output>,
    }
}
}

The AtLeastOne type is a vector wrapper which always contains at least one element. For more details, see its documentation.

Some of these fields are in a different order to the serialized data, see the V4 and V5 transaction specs for details.

The following types have ZcashSerialize and ZcashDeserialize implementations, because they can be serialized into a single byte vector:

  • Amount
  • sapling::tree::Root
  • redjubjub::Signature<Binding>

Adding V5 Sapling Spend

Sapling spend code is located at zebra-chain/src/sapling/spend.rs. We use AnchorVariant to model the anchor differences between V4 and V5. And we create a struct for serializing V5 transaction spends:


#![allow(unused)]
fn main() {
struct Spend<AnchorV: AnchorVariant> {
    cv: commitment::ValueCommitment,
    per_spend_anchor: AnchorV::PerSpend,
    nullifier: note::Nullifier,
    rk: redjubjub::VerificationKeyBytes<SpendAuth>,
    // This field is stored in a separate array in v5 transactions, see:
    // https://zips.z.cash/protocol/nu5.pdf#txnencodingandconsensus
    // parse using `zcash_deserialize_external_count` and `zcash_serialize_external_count`
    zkproof: Groth16Proof,
    // This fields is stored in another separate array in v5 transactions
    spend_auth_sig: redjubjub::Signature<SpendAuth>,
}

/// The serialization prefix fields of a `Spend` in Transaction V5.
///
/// In `V5` transactions, spends are split into multiple arrays, so the prefix,
/// proof, and signature must be serialised and deserialized separately.
///
/// Serialized as `SpendDescriptionV5` in [protocol specification §7.3].
struct SpendPrefixInTransactionV5 {
    cv: commitment::ValueCommitment,
    nullifier: note::Nullifier,
    rk: redjubjub::VerificationKeyBytes<SpendAuth>,
}
}

The following types have ZcashSerialize and ZcashDeserialize implementations, because they can be serialized into a single byte vector:

  • Spend<PerSpendAnchor> (moved from the pre-RFC Spend)
  • SpendPrefixInTransactionV5 (new)
  • Groth16Proof
  • redjubjub::Signature<redjubjub::SpendAuth> (new - for v5 spend auth sig arrays)

Note: Spend<SharedAnchor> does not have serialize and deserialize implementations. It must be split using into_v5_parts before serialization, and recombined using from_v5_parts after deserialization.

These convenience methods convert between Spend<SharedAnchor> and its v5 parts: SpendPrefixInTransactionV5, the spend proof, and the spend auth signature.

Changes to Sapling Output

In Zcash the Sapling output fields are the same for V4 and V5 transactions, so the Output struct is unchanged. However, V4 and V5 transactions serialize outputs differently, so we create additional structs for serializing outputs in each transaction version.

The output code is located at zebra-chain/src/sapling/output.rs:


#![allow(unused)]
fn main() {
struct Output {
    cv: commitment::ValueCommitment,
    cm_u: jubjub::Fq,
    ephemeral_key: keys::EphemeralPublicKey,
    enc_ciphertext: note::EncryptedNote,
    out_ciphertext: note::WrappedNoteKey,
    // This field is stored in a separate array in v5 transactions, see:
    // https://zips.z.cash/protocol/nu5.pdf#txnencodingandconsensus
    // parse using `zcash_deserialize_external_count` and `zcash_serialize_external_count`
    zkproof: Groth16Proof,
}

/// Wrapper for `Output` serialization in a `V4` transaction.
struct OutputInTransactionV4(pub Output);

/// The serialization prefix fields of an `Output` in Transaction V5.
///
/// In `V5` transactions, spends are split into multiple arrays, so the prefix
/// and proof must be serialised and deserialized separately.
///
/// Serialized as `OutputDescriptionV5` in [protocol specification §7.3].
struct OutputPrefixInTransactionV5 {
    cv: commitment::ValueCommitment,
    cm_u: jubjub::Fq,
    ephemeral_key: keys::EphemeralPublicKey,
    enc_ciphertext: note::EncryptedNote,
    out_ciphertext: note::WrappedNoteKey,
}
}

The following fields have ZcashSerialize and ZcashDeserialize implementations, because they can be serialized into a single byte vector:

  • OutputInTransactionV4 (moved from Output)
  • OutputPrefixInTransactionV5 (new)
  • Groth16Proof

Note: The serialize and deserialize implementations on Output are moved to OutputInTransactionV4. In v4 transactions, outputs must be wrapped using into_v4 before serialization, and unwrapped using from_v4 after deserialization. In transaction v5, outputs must be split using into_v5_parts before serialization, and recombined using from_v5_parts after deserialization.

These convenience methods convert Output to:

  • its v4 serialization wrapper OutputInTransactionV4, and
  • its v5 parts: OutputPrefixInTransactionV5 and the output proof.

Adding V5 Transactions

Now lets see how the V5 transaction is specified in the protocol, this is the second table of Transaction Encoding and Consensus and how are we going to represent it based in the above changes for Sapling fields and the new Orchard fields.

We propose the following representation for transaction V5 in Zebra:


#![allow(unused)]
fn main() {
enum Transaction::V5 {
    lock_time: LockTime,
    expiry_height: block::Height,
    inputs: Vec<transparent::Input>,
    outputs: Vec<transparent::Output>,
    sapling_shielded_data: Option<sapling::ShieldedData<SharedAnchor>>,
    orchard_shielded_data: Option<orchard::ShieldedData>,
}
}

To model the V5 anchor type, sapling_shielded_data uses the SharedAnchor variant located at zebra-chain/src/transaction/sapling/shielded_data.rs.

The following fields have ZcashSerialize and ZcashDeserialize implementations, because they can be serialized into a single byte vector:

  • LockTime
  • block::Height
  • transparent::Input
  • transparent::Output
  • Option<sapling::ShieldedData<SharedAnchor>> (new)
  • Option<orchard::ShieldedData> (new)

Orchard Additions

Adding Orchard ShieldedData

The new V5 structure will create a new orchard::ShieldedData type. This new type will be defined in a new zebra-chain/src/orchard/shielded_data.rs file:


#![allow(unused)]
fn main() {
struct orchard::ShieldedData {
    flags: Flags,
    value_balance: Amount,
    shared_anchor: orchard::tree::Root,
    proof: Halo2Proof,
    actions: AtLeastOne<AuthorizedAction>,
    binding_sig: redpallas::Signature<Binding>,
}
}

The fields are ordered based on the last data deserialized for each field.

The following types have ZcashSerialize and ZcashDeserialize implementations, because they can be serialized into a single byte vector:

  • orchard::Flags (new)
  • Amount
  • Halo2Proof (new)
  • redpallas::Signature<Binding> (new)

Adding Orchard AuthorizedAction

In V5 transactions, there is one SpendAuth signature for every Action. To ensure that this structural rule is followed, we create an AuthorizedAction type in orchard/shielded_data.rs:


#![allow(unused)]
fn main() {
/// An authorized action description.
///
/// Every authorized Orchard `Action` must have a corresponding `SpendAuth` signature.
struct orchard::AuthorizedAction {
    action: Action,
    // This field is stored in a separate array in v5 transactions, see:
    // https://zips.z.cash/protocol/nu5.pdf#txnencodingandconsensus
    // parse using `zcash_deserialize_external_count` and `zcash_serialize_external_count`
    spend_auth_sig: redpallas::Signature<SpendAuth>,
}
}

Where Action is defined as Action definition.

The following types have ZcashSerialize and ZcashDeserialize implementations, because they can be serialized into a single byte vector:

  • Action (new)
  • redpallas::Signature<SpendAuth> (new)

Note: AuthorizedAction does not have serialize and deserialize implementations. It must be split using into_parts before serialization, and recombined using from_parts after deserialization.

These convenience methods convert between AuthorizedAction and its parts: Action and the spend auth signature.

Adding Orchard Flags

Finally, in the V5 transaction we have a new orchard::Flags type. This is a bitfield type defined as:


#![allow(unused)]
fn main() {
bitflags! {
    /// Per-Transaction flags for Orchard.
    ///
    /// The spend and output flags are passed to the `Halo2Proof` verifier, which verifies
    /// the relevant note spending and creation consensus rules.
    struct orchard::Flags: u8 {
        /// Enable spending non-zero valued Orchard notes.
        ///
        /// "the `enableSpendsOrchard` flag, if present, MUST be 0 for coinbase transactions"
        const ENABLE_SPENDS = 0b00000001;
        /// Enable creating new non-zero valued Orchard notes.
        const ENABLE_OUTPUTS = 0b00000010;
        // Reserved, zeros (bits 2 .. 7)
    }
}
}

This type is also defined in orchard/shielded_data.rs.

Note: A consensus rule was added to the protocol specification stating that:

In a version 5 transaction, the reserved bits 2..7 of the flagsOrchard field MUST be zero.

Test Plan

  • All renamed, modified and new types should serialize and deserialize.
  • The full V4 and V5 transactions should serialize and deserialize.
  • Prop test strategies for V4 and V5 will be updated and created.
  • Before NU5 activation on testnet, test on the following test vectors:
    • Hand-crafted Orchard-only, Orchard/Sapling, Orchard/Transparent, and Orchard/Sapling/Transparent transactions based on the spec
    • "Fake" Sapling-only and Sapling/Transparent transactions based on the existing test vectors, converted from V4 to V5 format
      • We can write a test utility function to automatically do these conversions
    • An empty transaction, with no Orchard, Sapling, or Transparent data
      • A v5 transaction with no spends, but some outputs, to test the shared anchor serialization rule
    • Any available zcashd test vectors
  • After NU5 activation on testnet:
    • Add test vectors using the testnet activation block and 2 more post-activation blocks
  • After NU5 activation on mainnet:
    • Add test vectors using the mainnet activation block and 2 more post-activation blocks

Security

To avoid parsing memory exhaustion attacks, we will make the following changes across all Transaction, ShieldedData, Spend and Output variants, V1 through to V5:

  • Check cardinality consensus rules at parse time, before deserializing any Vecs
    • In general, Zcash requires that each transaction has at least one Transparent/Sprout/Sapling/Orchard transfer, this rule is not currently encoded in our data structures (it is only checked during semantic verification)
  • Stop parsing as soon as the first error is detected

These changes should be made in a later pull request, see #1917 for details.

Summary

Zebra programmers need to carefully write async code so it doesn't deadlock or hang. This is particularly important for poll, select, Buffer, Batch, and Mutex.

Zebra executes concurrent tasks using async Rust, with the tokio executor.

At a higher level, Zebra also uses tower::Services, tower::Buffers, and our own tower-batch implementation.

Motivation

Like all concurrent codebases, Zebra needs to obey certain constraints to avoid hangs. Unfortunately, Rust's tooling in these areas is still developing. So Zebra developers need to manually check these constraints during design, development, reviews, and testing.

Definitions

  • hang: a Zebra component stops making progress.
  • constraint: a rule that Zebra must follow to prevent hangs.
  • CORRECTNESS comment: the documentation for a constraint in Zebra's code.
  • task: an async task can execute code independently of other tasks, using cooperative multitasking.
  • contention: slower execution because multiple tasks are waiting to acquire a lock, buffer/batch slot, or readiness.
  • missed wakeup: a task hangs because it is never scheduled for wakeup.
  • lock: exclusive access to a shared resource. Locks stop other code from running until they are released. For example, a mutex, buffer slot, or service readiness.
  • critical section: code that is executed while holding a lock.
  • deadlock: a hang that stops an async task executing code, because it is waiting for a lock, slot, or task readiness. For example: a task is waiting for a service to be ready, but the service readiness depends on that task making progress.
  • starvation or livelock: a hang that executes code, but doesn't do anything useful. For example: a loop never terminates.

Guide-level explanation

If you are designing, developing, or testing concurrent Zebra code, follow the patterns in these examples to avoid hangs.

If you are reviewing concurrent Zebra designs or code, make sure that:

  • it is clear how the design or code avoids hangs
  • the design or code follows the patterns in these examples (as much as possible)
  • the concurrency constraints and risks are documented

The Reference section contains in-depth background information about Rust async concurrency in Zebra.

Here are some examples of concurrent designs and documentation in Zebra:

Registering Wakeups Before Returning Poll::Pending

To avoid missed wakeups, futures must schedule a wakeup before they return Poll::Pending. For more details, see the Poll::Pending and Wakeups section.

Zebra's unready_service.rs uses the ready! macro to correctly handle Poll::Pending from the inner service.

You can see some similar constraints in pull request #1954.


#![allow(unused)]
fn main() {
// CORRECTNESS
//
// The current task must be scheduled for wakeup every time we return
// `Poll::Pending`.
//
//`ready!` returns `Poll::Pending` when the service is unready, and
// the inner `poll_ready` schedules this task for wakeup.
//
// `cancel.poll` also schedules this task for wakeup if it is canceled.
let res = ready!(this
    .service
    .as_mut()
    .expect("poll after ready")
    .poll_ready(cx));
}

Futures-Aware Mutexes

To avoid hangs or slowdowns, prefer futures-aware types, particularly for complex waiting or locking code. But in some simple cases, std::sync::Mutex is more efficient. For more details, see the Futures-Aware Types section.

Zebra's Handshake won't block other tasks on its thread, because it uses futures::lock::Mutex:


#![allow(unused)]
fn main() {
pub async fn negotiate_version(
    peer_conn: &mut Framed<TcpStream, Codec>,
    addr: &SocketAddr,
    config: Config,
    nonces: Arc<futures::lock::Mutex<HashSet<Nonce>>>,
    user_agent: String,
    our_services: PeerServices,
    relay: bool,
) -> Result<(Version, PeerServices), HandshakeError> {
    // Create a random nonce for this connection
    let local_nonce = Nonce::default();
    // # Correctness
    //
    // It is ok to wait for the lock here, because handshakes have a short
    // timeout, and the async mutex will be released when the task times
    // out.
    nonces.lock().await.insert(local_nonce);

    ...
}
}

Zebra's Inbound service can't use an async-aware mutex for its AddressBook, because the mutex is shared with non-async code. It only holds the mutex to clone the address book, reducing the amount of time that other tasks on its thread are blocked:


#![allow(unused)]
fn main() {
// # Correctness
//
// Briefly hold the address book threaded mutex while
// cloning the address book. Then sanitize after releasing
// the lock.
let peers = address_book.lock().unwrap().clone();
let mut peers = peers.sanitized();
}

Avoiding Deadlocks when Acquiring Buffer or Service Readiness

To avoid deadlocks, readiness and locks must be acquired in a consistent order. For more details, see the Acquiring Buffer Slots, Mutexes, or Readiness section.

Zebra's ChainVerifier avoids deadlocks, contention, and errors by:

  • calling poll_ready before each call
  • acquiring buffer slots for the earlier verifier first (based on blockchain order)
  • ensuring that buffers are large enough for concurrent tasks

#![allow(unused)]
fn main() {
// We acquire checkpoint readiness before block readiness, to avoid an unlikely
// hang during the checkpoint to block verifier transition. If the checkpoint and
// block verifiers are contending for the same buffer/batch, we want the checkpoint
// verifier to win, so that checkpoint verification completes, and block verification
// can start. (Buffers and batches have multiple slots, so this contention is unlikely.)
//
// The chain verifier holds one slot in each verifier, for each concurrent task.
// Therefore, any shared buffers or batches polled by these verifiers should double
// their bounds. (For example, the state service buffer.)
ready!(self
    .checkpoint
    .poll_ready(cx)
    .map_err(VerifyChainError::Checkpoint))?;
ready!(self.block.poll_ready(cx).map_err(VerifyChainError::Block))?;
Poll::Ready(Ok(()))
}

Critical Section Compiler Errors

To avoid deadlocks or slowdowns, critical sections should be as short as possible, and they should not depend on any other tasks. For more details, see the Acquiring Buffer Slots, Mutexes, or Readiness section.

Zebra's CandidateSet must release a std::sync::Mutex lock before awaiting a tokio::time::Sleep future. This ensures that the threaded mutex lock isn't held over the await point.

If the lock isn't dropped, compilation fails, because the mutex lock can't be sent between threads.


#![allow(unused)]
fn main() {
// # Correctness
//
// In this critical section, we hold the address mutex, blocking the
// current thread, and all async tasks scheduled on that thread.
//
// To avoid deadlocks, the critical section:
// - must not acquire any other locks
// - must not await any futures
//
// To avoid hangs, any computation in the critical section should
// be kept to a minimum.
let reconnect = {
    let mut guard = self.address_book.lock().unwrap();
    ...
    let reconnect = guard.reconnection_peers().next()?;

    let reconnect = MetaAddr::new_reconnect(&reconnect.addr, &reconnect.services);
    guard.update(reconnect);
    reconnect
};

// SECURITY: rate-limit new candidate connections
sleep.await;
}

Sharing Progress between Multiple Futures

To avoid starvation and deadlocks, tasks that depend on multiple futures should make progress on all of those futures. This is particularly important for tasks that depend on their own outputs. For more details, see the Unbiased Selection section.

Zebra's peer crawler task avoids starvation and deadlocks by:

You can see a range of hang fixes in pull request #1950.


#![allow(unused)]
fn main() {
// CORRECTNESS
//
// To avoid hangs and starvation, the crawler must:
// - spawn a separate task for each handshake, so they can make progress
//   independently (and avoid deadlocking each other)
// - use the `select!` macro for all actions, because the `select` function
//   is biased towards the first ready future

loop {
    let crawler_action = tokio::select! {
        a = handshakes.next() => a,
        a = crawl_timer.next() => a,
        _ = demand_rx.next() => {
            if let Some(candidate) = candidates.next().await {
                // candidates.next has a short delay, and briefly holds the address
                // book lock, so it shouldn't hang
                DemandHandshake { candidate }
            } else {
                DemandCrawl
            }
        }
    };

    match crawler_action {
        DemandHandshake { candidate } => {
            // spawn each handshake into an independent task, so it can make
            // progress independently of the crawls
            let hs_join =
                tokio::spawn(dial(candidate, connector.clone()));
            handshakes.push(Box::pin(hs_join));
        }
        DemandCrawl => {
            // update has timeouts, and briefly holds the address book
            // lock, so it shouldn't hang
            candidates.update().await?;
        }
        // handle handshake responses and the crawl timer
    }
}
}

Prioritising Cancellation Futures

To avoid starvation, cancellation futures must take priority over other futures, if multiple futures are ready. For more details, see the Biased Selection section.

Zebra's connection.rs avoids hangs by prioritising the cancel and timer futures over the peer receiver future. Under heavy load, the peer receiver future could always be ready with a new message, starving the cancel or timer futures.

You can see a range of hang fixes in pull request #1950.


#![allow(unused)]
fn main() {
// CORRECTNESS
//
// Currently, select prefers the first future if multiple
// futures are ready.
//
// If multiple futures are ready, we want the cancellation
// to take priority, then the timeout, then peer responses.
let cancel = future::select(tx.cancellation(), timer_ref);
match future::select(cancel, peer_rx.next()) {
    ...
}
}

Atomic Shutdown Flag

As of April 2021, Zebra implements some shutdown checks using an atomic bool.

Zebra's shutdown.rs avoids data races and missed updates by using the strongest memory ordering (SeqCst).

We plan to replace this raw atomic code with a channel, see #1678.


#![allow(unused)]
fn main() {
/// A flag to indicate if Zebra is shutting down.
///
/// Initialized to `false` at startup.
pub static IS_SHUTTING_DOWN: AtomicBool = AtomicBool::new(false);

/// Returns true if the application is shutting down.
pub fn is_shutting_down() -> bool {
    // ## Correctness:
    //
    // Since we're shutting down, and this is a one-time operation,
    // performance is not important. So we use the strongest memory
    // ordering.
    // https://doc.rust-lang.org/nomicon/atomics.html#sequentially-consistent
    IS_SHUTTING_DOWN.load(Ordering::SeqCst)
}

/// Sets the Zebra shutdown flag to `true`.
pub fn set_shutting_down() {
    IS_SHUTTING_DOWN.store(true, Ordering::SeqCst);
}
}

Integration Testing Async Code

Sometimes, it is difficult to unit test async code, because it has complex dependencies. For more details, see the Testing Async Code section.

zebrad's acceptance tests run short Zebra syncs on the Zcash mainnet or testnet. These acceptance tests make sure that zebrad can:

  • sync blocks using its async block download and verification pipeline
  • cancel a sync
  • reload disk state after a restart

These tests were introduced in pull request #1193.


#![allow(unused)]
fn main() {
/// Test if `zebrad` can sync some larger checkpoints on mainnet.
#[test]
fn sync_large_checkpoints_mainnet() -> Result<()> {
    let reuse_tempdir = sync_until(
        LARGE_CHECKPOINT_TEST_HEIGHT,
        Mainnet,
        STOP_AT_HEIGHT_REGEX,
        LARGE_CHECKPOINT_TIMEOUT,
        None,
    )?;

    // if stopping corrupts the rocksdb database, zebrad might hang or crash here
    // if stopping does not write the rocksdb database to disk, Zebra will
    // sync, rather than stopping immediately at the configured height
    sync_until(
        (LARGE_CHECKPOINT_TEST_HEIGHT - 1).unwrap(),
        Mainnet,
        "previous state height is greater than the stop height",
        STOP_ON_LOAD_TIMEOUT,
        Some(reuse_tempdir),
    )?;

    Ok(())
}
}

Instrumenting Async Functions

Sometimes, it is difficult to debug async code, because there are many tasks running concurrently. For more details, see the Monitoring Async Code section.

Zebra runs instrumentation on some of its async function using tracing. Here's an instrumentation example from Zebra's sync block downloader:


#![allow(unused)]
fn main() {
/// Queue a block for download and verification.
///
/// This method waits for the network to become ready, and returns an error
/// only if the network service fails. It returns immediately after queuing
/// the request.
#[instrument(level = "debug", skip(self), fields(%hash))]
pub async fn download_and_verify(&mut self, hash: block::Hash) -> Result<(), Report> {
    ...
}
}

Tracing and Metrics in Async Functions

Sometimes, it is difficult to monitor async code, because there are many tasks running concurrently. For more details, see the Monitoring Async Code section.

Zebra's client requests are monitored via:

  • trace and debug logs using tracing crate
  • related work spans using the tracing crate
  • counters using the metrics crate

#![allow(unused)]
fn main() {
/// Handle an incoming client request, possibly generating outgoing messages to the
/// remote peer.
///
/// NOTE: the caller should use .instrument(msg.span) to instrument the function.
async fn handle_client_request(&mut self, req: InProgressClientRequest) {
    trace!(?req.request);

    let InProgressClientRequest { request, tx, span } = req;

    if tx.is_canceled() {
        metrics::counter!("peer.canceled", 1);
        tracing::debug!("ignoring canceled request");
        return;
    }
    ...
}
}

Reference-level explanation

The reference section contains in-depth information about concurrency in Zebra:

Most Zebra designs or code changes will only touch on one or two of these areas.

After an await, the rest of the Future might not be run

Futures can be "canceled" at any await point. Authors of futures must be aware that after an await, the code might not run. Futures might be polled to completion causing the code to work. But then many years later, the code is changed and the future might conditionally not be polled to completion which breaks things. The burden falls on the user of the future to poll to completion, and there is no way for the lib author to enforce this - they can only document this invariant.

https://github.com/rust-lang/wg-async-foundations/blob/master/src/vision/submitted_stories/status_quo/alan_builds_a_cache.md#-frequently-asked-questions

In particular, FutureExt::now_or_never:

  • drops the future, and
  • doesn't schedule the task for wakeups.

So even if the future or service passed to now_or_never is cloned, the task won't be awoken when it is ready again.

Task starvation

Tokio tasks are scheduled cooperatively:

a task is allowed to run until it yields, indicating to the Tokio runtime’s scheduler that it cannot currently continue executing. When a task yields, the Tokio runtime switches to executing the next task.

If a task doesn't yield during a CPU-intensive operation, or a tight loop, it can starve other tasks on the same thread. This can cause hangs or timeouts.

There are a few different ways to avoid task starvation:

Poll::Pending and Wakeups

When returning Poll::Pending, poll functions must ensure that the task will be woken up when it is ready to make progress.

In most cases, the poll function calls another poll function that schedules the task for wakeup.

Any code that generates a new Poll::Pending should either have:

  • a CORRECTNESS comment explaining how the task is scheduled for wakeup, or
  • a wakeup implementation, with tests to ensure that the wakeup functions as expected.

Note: poll functions often have a qualifier, like poll_ready or poll_next.

Futures-Aware Types

Prefer futures-aware types in complex locking or waiting code, rather than types which will block the current thread.

For example:

  • Use futures::lock::Mutex rather than std::sync::Mutex
  • Use tokio::time::{sleep, timeout} rather than std::thread::sleep

Always qualify ambiguous names like Mutex and sleep, so that it is obvious when a call will block.

If you are unable to use futures-aware types:

  • block the thread for as short a time as possible
  • document the correctness of each blocking call
  • consider re-designing the code to use tower::Services, or other futures-aware types

In some simple cases, std::sync::Mutex is correct and more efficient, when:

  • the value behind the mutex is just data, and
  • the locking behaviour is simple.

In these cases:

wrap the Arc<Mutex<...>> in a struct that provides non-async methods for performing operations on the data within, and only lock the mutex inside these methods

For more details, see the tokio documentation.

Acquiring Buffer Slots, Mutexes, or Readiness

Ideally, buffer slots, mutexes, or readiness should be:

  • acquired with one lock per critical section, and
  • held for as short a time as possible.

If multiple locks are required for a critical section, acquire them in the same order any time those locks are used. If tasks acquire multiple locks in different orders, they can deadlock, each holding a lock that the other needs.

If a buffer, mutex, future or service has complex readiness dependencies, schedule those dependencies separate tasks using tokio::spawn. Otherwise, it might deadlock due to a dependency loop within a single executor task.

Carefully read the documentation of the channel methods you call, to check if they lock. For example, tokio::sync::watch::Receiver::borrow holds a read lock, so the borrowed data should always be cloned. Use Arc for efficient clones if needed.

Never have two active watch borrow guards in the same scope, because that can cause a deadlock. The watch::Sender may start acquiring a write lock while the first borrow guard is active but the second one isn't. That means that the first read lock was acquired, but the second never will be because starting to acquire the write lock blocks any other read locks from being acquired. At the same time, the write lock will also never finish acquiring, because it waits for all read locks to be released, and the first read lock won't be released before the second read lock is acquired.

In all of these cases:

  • make critical sections as short as possible, and
  • do not depend on other tasks or locks inside the critical section.

Acquiring Service Readiness

Note: do not call poll_ready on multiple tasks, then match against the results. Use the ready! macro instead, to acquire service readiness in a consistent order.

Buffer and Batch

The constraints imposed by the tower::Buffer and tower::Batch implementations are:

  1. poll_ready must be called at least once for each call
  2. Once we've reserved a buffer slot, we always get Poll::Ready from a buffer, regardless of the current readiness of the buffer or its underlying service
  3. The Buffer/Batch capacity limits the number of concurrently waiting tasks. Once this limit is reached, further tasks will block, awaiting a free reservation.
  4. Some tasks can depend on other tasks before they resolve. (For example: block validation.) If there are task dependencies, the Buffer/Batch capacity must be larger than the maximum number of concurrently waiting tasks, or Zebra could deadlock (hang).

We also avoid hangs because:

  • the timeouts on network messages, block downloads, and block verification will restart verification if it hangs
  • Buffer and Batch release their reservations when response future is returned by the buffered/batched service, even if the returned future hangs
    • in general, we should move as much work into futures as possible, unless the design requires sequential calls
  • larger Buffer/Batch bounds

Buffered Services

A service should be provided wrapped in a Buffer if:

  • it is a complex service
  • it has multiple callers, or
  • it has a single caller that calls it multiple times concurrently.

Services might also have other reasons for using a Buffer. These reasons should be documented.

Choosing Buffer Bounds

Zebra's Buffer bounds should be set to the maximum number of concurrent requests, plus 1:

it's advisable to set bound to be at least the maximum number of concurrent requests the Buffer will see https://docs.rs/tower/0.4.3/tower/buffer/struct.Buffer.html#method.new

The extra slot protects us from future changes that add an extra caller, or extra concurrency.

As a general rule, Zebra Buffers should all have at least 5 slots, because most Zebra services can be called concurrently by:

  • the sync service,
  • the inbound service, and
  • multiple concurrent zebra-client blockchain scanning tasks.

Services might also have other reasons for a larger bound. These reasons should be documented.

We should limit Buffer lengths for services whose requests or responses contain Blocks (or other large data items, such as Transaction vectors). A long Buffer full of Blocks can significantly increase memory usage.

For example, parsing a malicious 2 MB block can take up to 12 MB of RAM. So a 5 slot buffer can use 60 MB of RAM.

Long Buffers can also increase request latency. Latency isn't a concern for Zebra's core use case as a node software, but it might be an issue if wallets, exchanges, or block explorers want to use Zebra.

Awaiting Multiple Futures

When awaiting multiple futures, Zebra can use biased or unbiased selection.

Typically, we prefer unbiased selection, so that if multiple futures are ready, they each have a chance of completing. But if one of the futures needs to take priority (for example, cancellation), you might want to use biased selection.

Unbiased Selection

The futures::select! and tokio::select! macros select ready arguments at random by default.

To poll a select! in order, pass biased; as the first argument to the macro.

Also consider the FuturesUnordered stream for unbiased selection of a large number of futures. However, this macro and stream require mapping all arguments to the same type.

Consider mapping the returned type to a custom enum with module-specific names.

Biased Selection

The futures::select is biased towards its first argument. If the first argument is always ready, the second argument will never be returned. (This behavior is not documented or guaranteed.) This bias can cause starvation or hangs. Consider edge cases where queues are full, or there are a lot of messages. If in doubt:

  • put shutdown or cancel oneshots first, then timers, then other futures
  • use the select! macro to ensure fairness

Select's bias can be useful to ensure that cancel oneshots and timers are always executed first. Consider the select_biased! macro and FuturesOrdered stream for guaranteed ordered selection of futures. (However, this macro and stream require mapping all arguments to the same type.)

The futures::select Either return type is complex, particularly when nested. This makes code hard to read and maintain. Map the Either to a custom enum.

Replacing Atomics with Channels

If you're considering using atomics, prefer a safe, tested, portable abstraction, like tokio's watch or oneshot channels.

In Zebra, we try to use safe abstractions, and write obviously correct code. It takes a lot of effort to write, test, and maintain low-level code. Almost all of our performance-critical code is in cryptographic libraries. And our biggest performance gains from those libraries come from async batch cryptography.

We are gradually replacing atomics with channels in Zebra.

Atomic Risks

Some atomic sizes and atomic operations are not available on some platforms. Others come with a performance penalty on some platforms.

It's also easy to use a memory ordering that's too weak. Future code changes might require a stronger memory ordering. But it's hard to test for these kinds of memory ordering bugs.

Some memory ordering bugs can only be discovered on non-x86 platforms. And when they do occur, they can be rare. x86 processors guarantee strong orderings, even for Relaxed accesses. Since Zebra's CI all runs on x86 (as of June 2021), our tests get AcqRel orderings, even when we specify Relaxed. But ARM processors like the Apple M1 implement weaker memory orderings, including genuinely Relaxed access. For more details, see the hardware reordering section of the Rust nomicon.

But if a Zebra feature requires atomics:

  1. use an AtomicUsize with the strongest memory ordering (SeqCst)
  2. use a weaker memory ordering, with:
  • a correctness comment,
  • multithreaded tests with a concurrency permutation harness like loom, on x86 and ARM, and
  • benchmarks to prove that the low-level code is faster.

Tokio's watch channel uses SeqCst for reads and writes to its internal "version" atomic. So Zebra should do the same.

Testing Async Code

Zebra's existing acceptance and integration tests will catch most hangs and deadlocks.

Some tests are only run after merging to main. If a recently merged PR fails on main, we revert the PR, and fix the failure.

Some concurrency bugs only happen intermittently. Zebra developers should run regular full syncs to ensure that their code doesn't cause intermittent hangs. This is particularly important for code that modifies Zebra's highly concurrent crates:

  • zebrad
  • zebra-network
  • zebra-state
  • zebra-consensus
  • tower-batch
  • tower-fallback

Monitoring Async Code

Zebra uses the following crates for monitoring and diagnostics:

These introspection tools are also useful during testing:

  • tracing logs individual events
    • spans track related work through the download and verification pipeline
  • metrics monitors overall progress and error rates
    • labels split counters or gauges into different categories (for example, by peer address)

Drawbacks

Implementing and reviewing these constraints creates extra work for developers. But concurrency bugs slow down every developer, and impact users. And diagnosing those bugs can take a lot of developer effort.

Unresolved questions

Can we catch these bugs using automated tests?

How can we diagnose these kinds of issues faster and more reliably?

Summary

This document describes how to verify the Zcash chain and transaction value pools in Zebra.

Motivation

In the Zcash protocol there are consensus rules that:

  • prohibit negative chain value pools ZIP-209, and
  • restrict the creation of new money to a specific number of coins in each coinbase transaction. Spec Section 3.4

These rules make sure that a fixed amount of Zcash is created by each block, even if there are vulnerabilities in some shielded pools.

(Checking the coins created by coinbase transactions and funding streams is out of scope for this design.)

Definitions

Transaction Value Balances

  • transaction value pool - The unspent input value in a transaction. Inputs add value, outputs remove value, and value balances modify value. The pool represents the sum of transparent and shielded inputs, minus the sum of transparent and shielded outputs.
  • value balance - The change in a transaction's value pool. There is a separate value balance for each transparent and shielded pool.
  • transparent value balance - The change in the transaction value pool, due to transparent inputs and outputs. The sum of the UTXOs spent by transparent inputs in tx_in fields, minus the sum of newly created outputs in tx_out fields.
  • sprout value balance - The change in the transaction value pool, due to sprout JoinSplits. The sum of all v_sprout_new fields, minus the sum of all v_sprout_old fields.
  • sapling value balance - The change in the transaction value pool, due to sapling Spends and Outputs. Equal to the valueBalanceSapling field.
  • orchard value balance - The change in the transaction value pool, due to orchard Actions. Equal to the valueBalanceOrchard field.
  • remaining transaction value - The unspent value in the transaction value pool. The sum of the transparent and shielded value balances in each transaction. This value is equal to the transaction value pool after we know the values of all the input UTXOs.
  • coinbase transaction - A transaction which spends newly created value (coinbase), and the remaining value of other transactions in its block (miner fees). Coinbase transactions do not have any other inputs, so they can't spend the outputs of other transactions.

Chain Value Pools

Note: chain value pools and transaction value balances have opposite signs.

  • chain value pool balance - The total value of unspent outputs in the chain, for each transparent and shielded pool. The sum of all block chain value pool changes in the chain. Each of the transparent, sprout, sapling, and orchard chain value pool balances must be non-negative.
  • block chain value pool change - The change in the chain value pools caused by a block. The negative sum of all the value balances in each block.

Guide-level explanation

Transaction Value Balances

Each transaction has an individual value pool, containing its unspent input value.

Spent transparent inputs add value to this pool, and newly created transparent outputs remove value. Similarly, Sprout JoinSplits have a field that adds value to the transaction pool, and a field that removes value. These transparent and sprout values are unsigned.

Sapling and Orchard have a single signed value balance per transaction, which modifies the transaction value pool.

We need to check that each transaction's total output value is less than or equal to the total input value. The remaining value in the transaction must not be negative. This makes sure that transactions can only spend up to their total input value. (Only coinbase transactions can create new value.)

In the spec, this is called the remaining value in the transparent transaction value pool. But in Zebra, we don't assign this value to a specific pool. We just call it the transaction value pool.

Chain Value Pools

There is one chain value pool for transparent funds, and one for each kind of shielded transfer, containing their unspent outputs.

These value pools are updated using chain value pool changes, which are the negation of transaction value balances. (Transaction value balances use unspent input value, but chain value balances use unspent outputs.)

Each of the chain value pools can change its value with every block added to the chain. This is a state feature and Zebra handle this in the zebra-state crate. We propose to store the pool values for the finalized tip height on disk.

We need to check each chain value pool as blocks are added to the chain, to make sure that chain balances never go negative.

Summary of the implementation:

  • Create a new type ValueBalance that will contain Amounts for each pool(transparent, sprout, sapling, orchard).
  • Create value_pool() methods on each relevant submodule (transparent, joinsplit, sapling and orchard).
  • Create a value_pool() method in transaction with all the above and in block with all the transaction value balances.
  • Pass the value balance of the incoming block into the state.
  • Get a previously stored value balance.
  • With both values check the consensus rules (constraint violations).
  • Update the saved values for the new tip.

Reference-level explanation

Consensus rules

Shielded Chain Value Pools

Consensus rules:

If any of the "Sprout chain value pool balance", "Sapling chain value pool balance", or "Orchard chain value pool balance" would become negative in the block chain created as a result of accepting a block, then all nodes MUST reject the block as invalid.

Nodes MAY relay transactions even if one or more of them cannot be mined due to the aforementioned restriction.

https://zips.z.cash/zip-0209#specification

Transparent Transaction Value Pool & Remaining Value

The unspent input value in a transaction: the sum of the transaction value balances.

Consensus rules:

Transparent inputs to a transaction insert value into a transparent transaction value pool associated with the transaction, and transparent outputs remove value from this pool.

As in Bitcoin, the remaining value in the transparent transaction value pool of a non-coinbase transaction is available to miners as a fee. The remaining value in the transparent transaction value pool of a coinbase transaction is destroyed.

The remaining value in the transparent transaction value pool MUST be nonnegative.

https://zips.z.cash/protocol/protocol.pdf#transactions

In Zebra, the remaining value in non-coinbase transactions is not assigned to any particular pool, until a miner spends it as part of a coinbase output.

Sprout Chain Value Pool

Consensus rules:

Each JoinSplit transfer can be seen, from the perspective of the transparent transaction value pool, as an input and an output simultaneously.

v_sprout_old takes value from the transparent transaction value pool and v_sprout_new adds value to the transparent transaction value pool . As a result, v_sprout_old is treated like an output value, whereas v_sprout_new is treated like an input value.

As defined in ZIP-209, the Sprout chain value pool balance for a given block chain is the sum of all v_sprout_old field values for transactions in the block chain, minus the sum of all v_sprout_new fields values for transactions in the block chain.

If the Sprout chain value pool balance would become negative in the block chain created as a result of accepting a block, then all nodes MUST reject the block as invalid.

https://zips.z.cash/protocol/protocol.pdf#joinsplitbalance

Sapling Chain Value Pool

Consensus rules:

A positive Sapling balancing value takes value from the Sapling transaction value pool and adds it to the transparent transaction value pool. A negative Sapling balancing value does the reverse. As a result, positive vbalanceSapling is treated like an input to the transparent transaction value pool, whereas negative vbalanceSapling is treated like an output from that pool.

As defined in ZIP-209, the Sapling chain value pool balance for a given block chain is the negation of the sum of all valueBalanceSapling field values for transactions in the block chain.

If the Sapling chain value pool balance would become negative in the block chain created as a result of accepting a block, then all nodes MUST reject the block as invalid.

https://zips.z.cash/protocol/protocol.pdf#saplingbalance

Orchard Chain Value Pool

Consensus rules:

Orchard introduces Action transfers, each of which can optionally perform a spend, and optionally perform an output. Similarly to Sapling, the net value of Orchard spends minus outputs in a transaction is called the Orchard balancing value, measured in zatoshi as a signed integer vbalanceOrchard.

vbalanceOrchard is encoded in a transaction as the field valueBalanceOrchard. If a transaction has no Action descriptions, vbalanceOrchard is implicitly zero. Transaction fields are described in § 7.1 ‘Transaction Encoding and Consensus’ on p. 116.

A positive Orchard balancing value takes value from the Orchard transaction value pool and adds it to the transparent transaction value pool. A negative Orchard balancing value does the reverse. As a result, positive vbalanceOrchard is treated like an input to the transparent transaction value pool, whereas negative vbalanceOrchard is treated like an output from that pool.

Similarly to the Sapling chain value pool balance defined in ZIP-209, the Orchard chain value pool balance for a given block chain is the negation of the sum of all valueBalanceOrchard field values for transactions in the block chain.

If the Orchard chain value pool balance would become negative in the block chain created as a result of accepting a block , then all nodes MUST reject the block as invalid.

https://zips.z.cash/protocol/protocol.pdf#orchardbalance

Transparent Chain Value Pool

Consensus rule:

Transfers of transparent value work essentially as in Bitcoin

https://zips.z.cash/protocol/protocol.pdf#overview

There is no explicit Zcash consensus rule that the transparent chain value pool balance must be non-negative. But an equivalent rule must be enforced by Zcash implementations, so that each block only creates a fixed amount of coins.

Specifically, this rule can be derived from other consensus rules:

  • a transparent output must have a non-negative value,
  • a transparent input can only spend an unspent transparent output,
  • so, there must be a non-negative remaining value in the transparent transaction value pool.

Some of these consensus rules are derived from Bitcoin, so they may not be documented in the Zcash Specification.

Coinbase Transactions

In this design, we assume that all coinbase outputs are valid, to avoid checking the newly created coinbase value, and the miner fees.

The coinbase value and miner fee rules will be checked as part of a future design.

Exceptions and Edge Cases

Value pools and value balances include the value of all unspent outputs, regardless of whether they can actually be spent.

For example:

  • transparent outputs which have unsatisfiable lock scripts
  • shielded outputs which have invalid private keys

However, some value is not part of any output:

  • if created value or miner fees are not spent in a coinbase transaction, they are destroyed
  • since coinbase transaction output values are rounded to the nearest zatoshi, any fractional part of miner-controlled or funding stream outputs is destroyed by rounding

Therefore:

  • the total of all chain value pools will always be strictly less than MAX_MONEY, and
  • the current total of all chain value pools will always be less than or equal to the number of coins created in coinbase transactions.

These properties are implied by other consensus rules, and do not need to be checked separately.

Proposed Implementation

Create a new ValueBalance type

  • Code will be located in a new file: zebra-chain/src/value_balance.rs.
  • Supported operators apply to all the Amounts inside the type: +, -, +=, -=, sum().
  • Implementation of the above operators are similar to the ones implemented for Amount<C> in zebra-chain/src/amount.rs. In particular, we want to return a Result on them so we can error when a constraint is violated.
  • We will use Default to represent a totally empty ValueBalance, this is the state of all pools at the genesis block.

#![allow(unused)]
fn main() {
#[serde(bound = "C: Constraint")]
struct ValueBalance<C = NegativeAllowed> {
    transparent: Amount<C>,
    sprout: Amount<C>,
    sapling: Amount<C>,
    orchard: Amount<C>,
}

impl ValueBalance {
    /// [Consensus rule]: The remaining value in the transparent transaction value pool MUST be nonnegative.
    ///
    /// This rule applies to Block and Mempool transactions.
    ///
    /// [Consensus rule]: https://zips.z.cash/protocol/protocol.pdf#transactions
    fn remaining_transaction_value(&self) -> Result<Amount<NonNegative>, Err> {
        // This rule checks the transparent value balance minus the sum of the sprout, sapling, and orchard
        // value balances in a transaction is nonnegative
        self.transparent - [self.sprout + self.sapling + self.orchard].sum()
    }
}

impl Add for Result<ValueBalance<C>>
where
    C: Constraint,
{

}

impl Sub for Result<ValueBalance<C>>
where
    C: Constraint,
{

}

impl AddAssign for Result<ValueBalance<C>>
where
    C: Constraint,
{

}

impl SubAssign for Result<ValueBalance<C>>
where
    C: Constraint,
{

}

impl Sum for Result<ValueBalance<C>>
where
    C: Constraint,
{

}

impl Default for ValueBalance<C>
where
    C: Constraint,
{

}
}

Create a method in Transaction that returns ValueBalance<NegativeAllowed> for the transaction

We first add value_balance() methods in all the modules we need and use them to get the value balance for the whole transaction.

Create a method in Input that returns ValueBalance<NegativeAllowed>

  • Method location is at zebra-chain/src/transparent.rs.
  • Method need utxos, this information is available in verify_transparent_inputs_and_outputs.
  • If the utxos are not available in the block or state, verification will timeout and return an error

#![allow(unused)]
fn main() {
impl Input {
    fn value_balance(&self, utxos: &HashMap<OutPoint, Utxo>) -> ValueBalance<NegativeAllowed> {

    }
}
}

Create a method in Output that returns ValueBalance<NegativeAllowed>

  • Method location is at zebra-chain/src/transparent.rs.

#![allow(unused)]
fn main() {
impl Output {
    fn value_balance(&self) -> ValueBalance<NegativeAllowed> {

    }
}
}

Create a method in JoinSplitData that returns ValueBalance<NegativeAllowed>

  • Method location is at zebra-chain/src/transaction/joinsplit.rs

#![allow(unused)]
fn main() {
pub fn value_balance(&self) -> ValueBalance<NegativeAllowed> {

}
}

Create a method in sapling::ShieldedData that returns ValueBalance<NegativeAllowed>

  • Method location is at zebra-chain/src/transaction/sapling/shielded_data.rs

#![allow(unused)]
fn main() {
pub fn value_balance(&self) -> ValueBalance<NegativeAllowed> {

}
}

Create a method in orchard::ShieldedData that returns ValueBalance<NegativeAllowed>

  • Method location is at zebra-chain/src/transaction/orchard/shielded_data.rs

#![allow(unused)]
fn main() {
pub fn value_balance(&self) -> ValueBalance<NegativeAllowed> {

}
}

Create the Transaction method

  • Method location: zebra-chain/src/transaction.rs
  • Method will use all the value_balances() we created until now.

#![allow(unused)]
fn main() {
/// utxos must contain the utxos of every input in the transaction,
/// including UTXOs created by earlier transactions in this block.
pub fn value_balance(&self, utxos: &HashMap<transparent::OutPoint, Utxo>) -> ValueBalance<NegativeAllowed> {

}
}

Create a method in Block that returns ValueBalance<NegativeAllowed> for the block

  • Method location is at zebra-chain/src/block.rs.
  • Method will make use of Transaction::value_balance method created before.

#![allow(unused)]
fn main() {
/// utxos must contain the utxos of every input in the transaction,
/// including UTXOs created by a transaction in this block,
/// then spent by a later transaction that's also in this block.
pub fn value_balance(&self, utxos: &HashMap<transparent::OutPoint, Utxo>) -> ValueBalance<NegativeAllowed> {
    self.transactions()
        .map(Transaction::value_balance)
        .sum()
        .expect("Each block should have at least one coinbase transaction")
}
}

Check the remaining transaction value consensus rule

  • Do the check in zebra-consensus/src/transaction.rs
  • Make the check part of the basic checks

#![allow(unused)]
fn main() {
..
// Check the remaining transaction value consensus rule:
tx.value_balance().remaining_transaction_value()?;
..
}

Pass the value balance for this block from the consensus into the state

  • Add a new field into PreparedBlock located at zebra-state/src/request.rs, this is the NonFinalized section of the state.

#![allow(unused)]
fn main() {
pub struct PreparedBlock {
    ..
    /// The value balances for each pool for this block.
    pub block_value_balance: ValuePool<NegativeAllowed>,
}
}
  • In zebra-consensus/src/block.rs pass the value balance to the zebra-state:

#![allow(unused)]
fn main() {
let block_value_balance = block.value_balance();
let prepared_block = zs::PreparedBlock {
    ..
    block_value_balance,
};
}

Add a value pool into the state Chain struct

  • This is the value pool for the non finalized part of the blockchain.
  • Location of the Chain structure where the pool field will be added: zebra-state/src/service/non_finalized_state/chain.rs

#![allow(unused)]
fn main() {
pub struct Chain {
    ..
    /// The chain value pool balance at the tip of this chain.
    value_pool: ValueBalance<NonNegative>,
}
}
  • Add a new argument finalized_tip_value_balance to the commit_new_chain() method located in the same file.
  • Pass the new argument to the Chain in:

#![allow(unused)]
fn main() {
let mut chain = Chain::new(finalized_tip_history_tree, finalized_tip_value_balance);
}

Note: We don't need to pass the finalized tip value balance into the commit_block() method.

Check the consensus rules when the chain is updated or reversed

  • Location: zebra-state/src/service/non_finalized_state/chain.rs

#![allow(unused)]
fn main() {
impl UpdateWith<ValueBalance<NegativeAllowed>> for Chain {
    fn update_chain_state_with(&mut self, value_balance: &ValueBalance<NegativeAllowed>) -> Result<(), Err> {
        self.value_pool = (self.value_pool + value_balance)?;
        Ok(())
    }
    fn revert_chain_state_with(&mut self, value_balance: &ValueBalance<NegativeAllowed>) -> Result<(), Err> {
        self.value_pool = (self.value_pool + value_balance)?;
        Ok(())
    }
}
}

Changes to finalized state

The state service will call commit_new_chain(). We need to pass the value pool from the disk into this function.


#![allow(unused)]
fn main() {
self.mem
    .commit_new_chain(prepared, self.disk.history_tree(), self.disk.get_pool())?;
}

We now detail what is needed in order to have the get_pool() method available.

Serialization of ValueBalance<C>

In order to save ValueBalance into the disk database we must implement IntoDisk and FromDisk for ValueBalance and for Amount:


#![allow(unused)]
fn main() {
impl IntoDisk for ValueBalance<C> {
    type Bytes = [u8; 32];

    fn as_bytes(&self) -> Self::Bytes {
        [self.transparent.to_bytes(), self.sprout.to_bytes(),
        self.sapling.to_bytes(), self.orchard.to_bytes()].concat()
    }
}

impl FromDisk for ValueBalance<C> {
    fn from_bytes(bytes: impl AsRef<[u8]>) -> Self {
        let array = bytes.as_ref().try_into().unwrap();
        ValueBalance {
            transparent: Amount::from_bytes(array[0..8]).try_into().unwrap()
            sprout: Amount::from_bytes(array[8..16]).try_into().unwrap()
            sapling: Amount::from_bytes(array[16..24]).try_into().unwrap()
            orchard: Amount::from_bytes(array[24..32]).try_into().unwrap()
        }
    }
}

impl IntoDisk for Amount {
    type Bytes = [u8; 8];

    fn as_bytes(&self) -> Self::Bytes {
        self.to_bytes()
    }
}

impl FromDisk for Amount {
    fn from_bytes(bytes: impl AsRef<[u8]>) -> Self {
        let array = bytes.as_ref().try_into().unwrap();
        Amount::from_bytes(array)
    }
}
}

The above code is going to need a Amount::from_bytes new method.

Add a from_bytes method in Amount

  • Method location is at zebra-chain/src/amount.rs
  • A to_bytes() method already exist, place from_bytes() right after it.

#![allow(unused)]
fn main() {
/// From little endian byte array
pub fn from_bytes(&self, bytes: [u8; 8]) -> Self {
    let amount = i64::from_le_bytes(bytes).try_into().unwrap();
    Self(amount, PhantomData)
}
}

Changes to zebra-state/src/request.rs

Add a new field to FinalizedState:


#![allow(unused)]
fn main() {
pub struct FinalizedBlock {
    ..
    /// The value balance for transparent, sprout, sapling and orchard
    /// inside all the transactions of this block.
    pub(crate) block_value_balance: ValueBalance<NegativeAllowed>,
}
}

Populate it when PreparedBlock is converted into FinalizedBlock:


#![allow(unused)]
fn main() {
impl From<PreparedBlock> for FinalizedBlock {
    fn from(prepared: PreparedBlock) -> Self {
        let PreparedBlock {
            ..
            block_value_balance,
        } = prepared;
        Self {
            ..
            block_value_balance,
        }
    }
}
}

Changes to zebra-state/src/service/finalized_state.rs

First we add a column of type ValueBalance that will contain Amounts for all the pools: transparent, sprout, sapling, orchard:


#![allow(unused)]
fn main() {
rocksdb::ColumnFamilyDescriptor::new("tip_chain_value_pool", db_options.clone()),
}

At block commit(commit_finalized_direct()) we create the handle for the new column:


#![allow(unused)]
fn main() {
let tip_chain_value_pool = self.db.cf_handle("tip_chain_value_pool").unwrap();
}

Next we save each tip value pool into the field for each upcoming block except for the genesis block:


#![allow(unused)]
fn main() {
// Consensus rule: The block height of the genesis block is 0
// https://zips.z.cash/protocol/protocol.pdf#blockchain
if height == block::Height(0) {
    batch.zs_insert(tip_chain_value_pool, height, ValueBalance::default());
} else {
    let current_pool = self.current_value_pool();
    batch.zs_insert(tip_chain_value_pool, height, (current_pool + finalized.block_value_balance)?);
}
}

The current_value_pool() function will get the stored value of the pool at the tip as follows:


#![allow(unused)]
fn main() {
pub fn current_value_pool(&self) -> ValuePool<NonNegative> {
    self.db.cf_handle("tip_chain_value_pool")
}
}

Test Plan

Unit tests

  • Create a transaction that has a negative remaining value.
    • Test that the transaction fails the verification in Transaction::value_balance()
    • To avoid passing the utxo we can have 0 as the amount of the transparent pool and some negative shielded pool.

Prop tests

  • Create a chain strategy that ends up with a valid value balance for all the pools (transparent, sprout, sapling, orchard)
    • Test that the amounts are all added to disk.
  • Add new blocks that will make each pool became negative.
    • Test for constraint violations in the value balances for each case.
    • Failures should be at update_chain_state_with().
  • Test consensus rules success and failures in revert_chain_state_with()
    • TODO: how?
  • serialize and deserialize ValueBalance using IntoDisk and FromDisk

Manual tests

  • Zebra must sync up to tip computing all value balances and never breaking the value pool rules.

Future Work

Add an extra state request to verify the speculative chain balance after applying a Mempool transaction. (This is out of scope for our current NU5 and mempool work.)

Note: The chain value pool balance rules apply to Block transactions, but they are optional for Mempool transactions:

Nodes MAY relay transactions even if one or more of them cannot be mined due to the aforementioned restriction.

https://zips.z.cash/zip-0209#specification

Since Zebra does chain value pool balance validation in the state, we want to skip verifying the speculative chain balance of Mempool transactions.

Diagrams

  ┌───────────┐     ┌───────────┐     ┌───────────┐     ┌───────────┐
  │PeerServer │     │PeerServer │     │PeerServer │     │PeerServer │
  │ ┌───────┐ │     │ ┌───────┐ │     │ ┌───────┐ │     │ ┌───────┐ │
  │ │┌─────┐│ │     │ │┌─────┐│ │     │ │┌─────┐│ │     │ │┌─────┐│ │
  │ ││ Tcp ││ │     │ ││ Tcp ││ │     │ ││ Tcp ││ │     │ ││ Tcp ││ │
  │ │└─────┘│ │     │ │└─────┘│ │     │ │└─────┘│ │     │ │└─────┘│ │
  │ │Framed │ │     │ │Framed │ │     │ │Framed │ │     │ │Framed │ │
  │ │Stream │ │     │ │Stream │ │     │ │Stream │ │     │ │Stream │ │
  │ └───────┘─┼─┐   │ └───────┘─┼─┐   │ └───────┘─┼─┐   │ └───────┘─┼─┐
┏▶│     ┃     │ │ ┏▶│     ┃     │ │ ┏▶│     ┃     │ │ ┏▶│     ┃     │ │
┃ │     ┃     │ │ ┃ │     ┃     │ │ ┃ │     ┃     │ │ ┃ │     ┃     │ │
┃ │     ▼     │ │ ┃ │     ▼     │ │ ┃ │     ▼     │ │ ┃ │     ▼     │ │
┃ │ ┌───────┐ │ │ ┃ │ ┌───────┐ │ │ ┃ │ ┌───────┐ │ │ ┃ │ ┌───────┐ │ │
┃ │ │ Tower │ │ │ ┃ │ │ Tower │ │ │ ┃ │ │ Tower │ │ │ ┃ │ │ Tower │ │ │
┃ │ │Buffer │ │ │ ┃ │ │Buffer │ │ │ ┃ │ │Buffer │ │ │ ┃ │ │Buffer │ │ │
┃ │ └───────┘ │ │ ┃ │ └───────┘ │ │ ┃ │ └───────┘ │ │ ┃ │ └───────┘ │ │
┃ │     ┃     │ │ ┃ │     ┃     │ │ ┃ │     ┃     │ │ ┃ │     ┃     │ │
┃ └─────╋─────┘ │ ┃ └─────╋─────┘ │ ┃ └─────╋─────┘ │ ┃ └─────╋─────┘ │
┃       ┃       └─╋───────╋───────┴─╋───────╋───────┴─╋───────╋───────┴───────┐
┃       ┃         ┃       ┃         ┃       ┃         ┃       ┃               │
┃       ┃         ┃       ┃         ┃       ┃         ┃       ┃               │
┃       ┗━━━━━━━━━╋━━━━━━━┻━━━━━━━━━╋━━━━━━━┻━━━━━━━━━╋━━━━━━━┻━━━━━━━━━┓     │
┗━━━━━━━┓         ┗━━━━━━━┓         ┗━━━━━━━┓         ┗━━━━━━━┓         ┃     │
 ┌──────╋─────────────────╋─────────────────╋─────────────────╋──────┐  ┃     │
 │      ┃                 ┃                 ┃                 ┃      │  ┃     │
 │┌───────────┐     ┌───────────┐     ┌───────────┐     ┌───────────┐│  ┃     │
 ││PeerClient │     │PeerClient │     │PeerClient │     │PeerClient ││  ┃     │
 │└───────────┘     └───────────┘     └───────────┘     └───────────┘│  ┃     │
 │                                                                   │  ┃     │
 │┌──────┐      ┌──────────────┐                                     │  ┃     │
 ││ load │      │peer discovery│                              PeerSet│  ┃     │
 ││signal│   ┏━▶│   receiver   │          req: Request, rsp: Response│  ┃     │
 │└──────┘   ┃  └──────────────┘         routes all outgoing requests│  ┃     │
 │    ┃      ┃                               adds peers via discovery│  ┃     │
 └────╋──────╋───────────────────────────────────────────────────────┘  ┃     │
      ┃      ┃                                             ▲            ┃     │
      ┃      ┣━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓             ┃            ┃     │
      ┃      ┃     ┏━━━━━━━━━━━╋━━━━━━━━━━━━━╋━━━━━━━━━━━━━┫            ┃     │
      ▼      ┃     ┃           ┃             ┃             ┃            ┃     │
  ┌────────────────╋───┐┌────────────┐┌─────────────┐      ┃            ┃     │
  │Crawler         ┃   ││  Listener  ││Initial Peers│      ┃            ┃     │
  │            ┌──────┐││            ││             │      ┃            ┃     │
  │            │Tower │││            ││             │      ┃            ┃     │
  │            │Buffer│││listens for ││ connects on │      ┃            ┃     │
  │            └──────┘││  incoming  ││  launch to  │      ┃            ┃     │
  │uses peerset to     ││connections,││ seed peers  │      ┃            ┃     │
  │crawl network,      ││   sends    ││specified in │      ┃            ┃     │
  │maintains candidate ││ handshakes ││ config file │      ┃            ┃     │
  │peer set, connects  ││  to peer   ││  to build   │      ┃            ┃     │
  │to new peers on load││ discovery  ││initial peer │      ┃            ┃     │
  │signal or timer     ││  receiver  ││     set     │      ┃            ┃     │
  └────────────────────┘└────────────┘└─────────────┘      ┃            ┃     │
             │        zebra-network internals              ┃            ┃     │
─ ─ ─ ─ ─ ─ ─│─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─┃─ ─ ─ ─ ─ ─ ╋ ─ ─ ┼
             │              exposed api                    ┃            ┃     │
             │             ┌────────────────────────┐      ┃            ┃     │
             │             │Arc<Mutex<AddressBook>> │      ┃            ┃     │
             │             │last-seen timestamps for│      ┃            ┃     │
             └─────────────│ each peer, obtained by │◀─────╋────────────╋─────┘
                           │ hooking into incoming  │      ┃            ┃
                           │    message streams     │      ┃            ┃
                           └────────────────────────┘      ┃            ▼
                                             ┌────────────────┐┌───────────────┐
                                             │Outbound Service││Inbound Service│
                                             │ req: Request,  ││ req: Request, │
                                             │ rsp: Response  ││ rsp: Response │
                                             │                ││               │
                                             │  Tower Buffer  ││  routes all   │
                                             └────────────────┘│   incoming    │
                                                               │requests, uses │
                                                               │   load-shed   │
                                                               │ middleware to │
                                                               │ remove peers  │
                                                               │ when internal │
                                                               │ services are  │
                                                               │  overloaded   │
                                                               └───────────────┘

zebra-checkpoints

zebra-checkpoints uses a local zcashd instance to generate a list of checkpoints for Zebra's checkpoint verifier.

Developers should run this tool every few months to add new checkpoints to Zebra. (By default, Zebra uses these checkpoints to sync to the chain tip.)

For more information on how to run this program visit Zebra checkpoints README

API Reference

Zebra's API documentation is generated using Rustdoc: