zebra_state/

service.rs

//! [`tower::Service`]s for Zebra's cached chain state.
//!
//! Zebra provides cached state access via two main services:
//! - [`StateService`]: a read-write service that writes blocks to the state,
//!   and redirects most read requests to the [`ReadStateService`].
//! - [`ReadStateService`]: a read-only service that answers from the most
//!   recent committed block.
//!
//! Most users should prefer [`ReadStateService`], unless they need to write blocks to the state.
//!
//! Zebra also provides access to the best chain tip via:
//! - [`LatestChainTip`]: a read-only channel that contains the latest committed
//!   tip.
//! - [`ChainTipChange`]: a read-only channel that can asynchronously await
//!   chain tip changes.

use std::{
    collections::HashMap,
    future::Future,
    pin::Pin,
    sync::Arc,
    task::{Context, Poll},
    time::{Duration, Instant},
};

use futures::future::FutureExt;
use tokio::sync::oneshot;
use tower::{util::BoxService, Service, ServiceExt};
use tracing::{instrument, Instrument, Span};

#[cfg(any(test, feature = "proptest-impl"))]
use tower::buffer::Buffer;

use zebra_chain::{
    block::{self, CountedHeader, HeightDiff},
    diagnostic::CodeTimer,
    parameters::{Network, NetworkUpgrade},
    serialization::ZcashSerialize,
    subtree::NoteCommitmentSubtreeIndex,
};

use crate::{
    constants::{
        MAX_FIND_BLOCK_HASHES_RESULTS, MAX_FIND_BLOCK_HEADERS_RESULTS, MAX_LEGACY_CHAIN_BLOCKS,
    },
    error::{CommitBlockError, CommitCheckpointVerifiedError, InvalidateError, ReconsiderError},
    request::TimedSpan,
    response::NonFinalizedBlocksListener,
    service::{
        block_iter::any_ancestor_blocks,
        chain_tip::{ChainTipBlock, ChainTipChange, ChainTipSender, LatestChainTip},
        finalized_state::{FinalizedState, ZebraDb},
        non_finalized_state::{Chain, NonFinalizedState},
        pending_utxos::PendingUtxos,
        queued_blocks::QueuedBlocks,
        read::find,
        watch_receiver::WatchReceiver,
    },
    AwaitUtxoError, BoxError, CheckpointVerifiedBlock, CommitSemanticallyVerifiedError, Config,
    KnownBlock, ReadRequest, ReadResponse, Request, Response, SemanticallyVerifiedBlock,
    StateInitError,
};

pub mod block_iter;
pub mod chain_tip;
pub mod watch_receiver;

pub mod check;

pub(crate) mod finalized_state;
pub(crate) mod non_finalized_state;
mod pending_utxos;
mod queued_blocks;
pub(crate) mod read;
mod traits;
mod write;

#[cfg(any(test, feature = "proptest-impl"))]
pub mod arbitrary;

#[cfg(test)]
mod tests;

pub use finalized_state::{OutputLocation, TransactionIndex, TransactionLocation};
use write::NonFinalizedWriteMessage;

use self::queued_blocks::{QueuedCheckpointVerified, QueuedSemanticallyVerified, SentHashes};

pub use self::traits::{ReadState, State};

/// A read-write service for Zebra's cached blockchain state.
///
/// This service modifies and provides access to:
/// - the non-finalized state: the most recent blocks, up to
///   [`MAX_BLOCK_REORG_HEIGHT`](crate::MAX_BLOCK_REORG_HEIGHT) of them.
///   Zebra allows chain forks in the non-finalized state,
///   stores it in memory, and re-downloads it when restarted.
/// - the finalized state: older blocks that have many confirmations.
///   Zebra stores the single best chain in the finalized state,
///   and re-loads it from disk when restarted.
///
/// Read requests to this service are buffered, then processed concurrently.
/// Block write requests are buffered, then queued, then processed in order by a separate task.
///
/// Most state users can get faster read responses using the [`ReadStateService`],
/// because its requests do not share a [`tower::buffer::Buffer`] with block write requests.
///
/// To quickly get the latest block, use [`LatestChainTip`] or [`ChainTipChange`].
/// They can read the latest block directly, without queueing any requests.
#[derive(Debug)]
pub(crate) struct StateService {
    // Configuration
    //
    /// The configured Zcash network.
    network: Network,

    /// The height that we start storing UTXOs from finalized blocks.
    ///
    /// This height should be lower than the last few checkpoints,
    /// so the full verifier can verify UTXO spends from those blocks,
    /// even if they haven't been committed to the finalized state yet.
    full_verifier_utxo_lookahead: block::Height,

    // Queued Blocks
    //
    /// Queued blocks for the [`NonFinalizedState`] that arrived out of order.
    /// These blocks are awaiting their parent blocks before they can do contextual verification.
    non_finalized_state_queued_blocks: QueuedBlocks,

    /// Queued blocks for the [`FinalizedState`] that arrived out of order.
    /// These blocks are awaiting their parent blocks before they can do contextual verification.
    ///
    /// Indexed by their parent block hash.
    finalized_state_queued_blocks: HashMap<block::Hash, QueuedCheckpointVerified>,

    /// Channels to send blocks to the block write task.
    block_write_sender: write::BlockWriteSender,

    /// The [`block::Hash`] of the most recent block sent on
    /// `finalized_block_write_sender` or `non_finalized_block_write_sender`.
    ///
    /// On startup, this is:
    /// - the finalized tip, if there are stored blocks, or
    /// - the genesis block's parent hash, if the database is empty.
    ///
    /// If `invalid_block_write_reset_receiver` gets a reset, this is:
    /// - the hash of the last valid committed block (the parent of the invalid block).
    finalized_block_write_last_sent_hash: block::Hash,

    /// A set of block hashes that have been sent to the block write task.
    /// Hashes of blocks below the finalized tip height are periodically pruned.
    non_finalized_block_write_sent_hashes: SentHashes,

    /// If an invalid block is sent on `finalized_block_write_sender`
    /// or `non_finalized_block_write_sender`,
    /// this channel gets the [`block::Hash`] of the valid tip.
    //
    // TODO: add tests for finalized and non-finalized resets (#2654)
    invalid_block_write_reset_receiver: tokio::sync::mpsc::UnboundedReceiver<block::Hash>,

    /// Receives the hash of every non-finalized block that the write task
    /// rejected, so the corresponding entry can be removed from
    /// `non_finalized_block_write_sent_hashes`.
    ///
    /// Without this, a rejected same-hash block locks out a later honest
    /// re-delivery of a block at the same hash as a "duplicate" until restart
    /// or reorg.
    non_finalized_rejected_receiver: tokio::sync::mpsc::UnboundedReceiver<block::Hash>,

    // Pending UTXO Request Tracking
    //
    /// The set of outpoints with pending requests for their associated transparent::Output.
    pending_utxos: PendingUtxos,

    /// Instant tracking the last time `pending_utxos` was pruned.
    last_prune: Instant,

    // Updating Concurrently Readable State
    //
    /// A cloneable [`ReadStateService`], used to answer concurrent read requests.
    ///
    /// TODO: move users of read [`Request`]s to [`ReadStateService`], and remove `read_service`.
    read_service: ReadStateService,

    // Metrics
    //
    /// A metric tracking the maximum height that's currently in `finalized_state_queued_blocks`
    ///
    /// Set to `f64::NAN` if `finalized_state_queued_blocks` is empty, because grafana shows NaNs
    /// as a break in the graph.
    max_finalized_queue_height: f64,
}

/// A read-only service for accessing Zebra's cached blockchain state.
///
/// This service provides read-only access to:
/// - the non-finalized state: the most recent blocks, up to
///   [`MAX_BLOCK_REORG_HEIGHT`](crate::MAX_BLOCK_REORG_HEIGHT) of them.
/// - the finalized state: older blocks that have many confirmations.
///
/// Requests to this service are processed in parallel,
/// ignoring any blocks queued by the read-write [`StateService`].
///
/// This quick response behavior is better for most state users.
/// It allows other async tasks to make progress while concurrently reading data from disk.
#[derive(Clone, Debug)]
pub struct ReadStateService {
    // Configuration
    //
    /// The configured Zcash network.
    network: Network,

    // Shared Concurrently Readable State
    //
    /// A watch channel with a cached copy of the [`NonFinalizedState`].
    ///
    /// This state is only updated between requests,
    /// so it might include some block data that is also on `disk`.
    non_finalized_state_receiver: WatchReceiver<NonFinalizedState>,

    /// The shared inner on-disk database for the finalized state.
    ///
    /// RocksDB allows reads and writes via a shared reference,
    /// but [`ZebraDb`] doesn't expose any write methods or types.
    ///
    /// This chain is updated concurrently with requests,
    /// so it might include some block data that is also in `best_mem`.
    db: ZebraDb,

    /// A shared handle to a task that writes blocks to the [`NonFinalizedState`] or [`FinalizedState`],
    /// once the queues have received all their parent blocks.
    ///
    /// Used to check for panics when writing blocks.
    block_write_task: Option<Arc<std::thread::JoinHandle<()>>>,
}

impl Drop for StateService {
    fn drop(&mut self) {
        // The state service owns the state, tasks, and channels,
        // so dropping it should shut down everything.

        // Close the channels (non-blocking)
        // This makes the block write thread exit the next time it checks the channels.
        // We want to do this here so we get any errors or panics from the block write task before it shuts down.
        self.invalid_block_write_reset_receiver.close();
        self.non_finalized_rejected_receiver.close();

        std::mem::drop(self.block_write_sender.finalized.take());
        std::mem::drop(self.block_write_sender.non_finalized.take());

        self.clear_finalized_block_queue(CommitBlockError::WriteTaskExited);
        self.clear_non_finalized_block_queue(CommitBlockError::WriteTaskExited);

        // Log database metrics before shutting down
        info!("dropping the state: logging database metrics");
        self.log_db_metrics();

        // Then drop self.read_service, which checks the block write task for panics,
        // and tries to shut down the database.
    }
}

impl Drop for ReadStateService {
    fn drop(&mut self) {
        // The read state service shares the state,
        // so dropping it should check if we can shut down.

        // TODO: move this into a try_shutdown() method
        if let Some(block_write_task) = self.block_write_task.take() {
            if let Some(block_write_task_handle) = Arc::into_inner(block_write_task) {
                // We're the last database user, so we can tell it to shut down (blocking):
                // - flushes the database to disk, and
                // - drops the database, which cleans up any database tasks correctly.
                self.db.shutdown(true);

                // We are the last state with a reference to this thread, so we can
                // wait until the block write task finishes, then check for panics (blocking).
                // (We'd also like to abort the thread, but std::thread::JoinHandle can't do that.)

                // This log is verbose during tests.
                #[cfg(not(test))]
                info!("waiting for the block write task to finish");
                #[cfg(test)]
                debug!("waiting for the block write task to finish");

                // TODO: move this into a check_for_panics() method
                if let Err(thread_panic) = block_write_task_handle.join() {
                    std::panic::resume_unwind(thread_panic);
                } else {
                    debug!("shutting down the state because the block write task has finished");
                }
            }
        } else {
            // Even if we're not the last database user, try shutting it down.
            //
            // TODO: rename this to try_shutdown()?
            self.db.shutdown(false);
        }
    }
}

impl StateService {
    const PRUNE_INTERVAL: Duration = Duration::from_secs(30);

    /// Creates a new state service for the state `config` and `network`.
    ///
    /// Uses the `max_checkpoint_height` and `checkpoint_verify_concurrency_limit`
    /// to work out when it is near the final checkpoint.
    ///
    /// Returns the read-write and read-only state services,
    /// and read-only watch channels for its best chain tip.
    pub async fn new(
        config: Config,
        network: &Network,
        max_checkpoint_height: block::Height,
        checkpoint_verify_concurrency_limit: usize,
    ) -> (Self, ReadStateService, LatestChainTip, ChainTipChange) {
        let (finalized_state, finalized_tip, timer) = {
            let config = config.clone();
            let network = network.clone();
            tokio::task::spawn_blocking(move || {
                let timer = CodeTimer::start();
                let finalized_state = FinalizedState::new(
                    &config,
                    &network,
                    #[cfg(feature = "elasticsearch")]
                    true,
                )
                .expect(
                    "opening the read-write finalized state database failed; check that the \
                     state cache directory is writable and not locked by another Zebra instance, \
                     and that there is free disk space",
                );
                timer.finish_desc("opening finalized state database");

                let timer = CodeTimer::start();
                let finalized_tip = finalized_state.db.tip_block();

                (finalized_state, finalized_tip, timer)
            })
            .await
            .expect("failed to join blocking task")
        };

        // # Correctness
        //
        // The state service must set the finalized block write sender to `None`
        // if there are blocks in the restored non-finalized state that are above
        // the max checkpoint height so that non-finalized blocks can be written, otherwise,
        // Zebra will be unable to commit semantically verified blocks, and its chain sync will stall.
        //
        // The state service must not set the finalized block write sender to `None` if there
        // aren't blocks in the restored non-finalized state that are above the max checkpoint height,
        // otherwise, unless checkpoint sync is disabled in the zebra-consensus configuration,
        // Zebra will be unable to commit checkpoint verified blocks, and its chain sync will stall.
        let finalized_tip_height = finalized_tip
            .as_ref()
            .map(|tip| tip.coinbase_height().expect("valid block must have height"));
        let is_finalized_tip_past_max_checkpoint =
            finalized_tip_height.is_some_and(|tip_height| tip_height >= max_checkpoint_height);
        let backup_dir_path = config.non_finalized_state_backup_dir(network);

        if backup_dir_path.is_some() && !is_finalized_tip_past_max_checkpoint {
            tracing::info!(
                ?finalized_tip_height,
                ?max_checkpoint_height,
                "not restoring the non-finalized state backup, because the finalized tip is absent \
                 or below the max checkpoint height: Zebra will re-download and re-verify the \
                 blocks above its finalized tip"
            );
        }
        let skip_backup_task = config.debug_skip_non_finalized_state_backup_task;
        let (non_finalized_state, non_finalized_state_sender, non_finalized_state_receiver) =
            NonFinalizedState::new(network)
                .with_backup(
                    backup_dir_path.clone(),
                    &finalized_state.db,
                    is_finalized_tip_past_max_checkpoint,
                    config.debug_skip_non_finalized_state_backup_task,
                )
                .await;

        let non_finalized_block_write_sent_hashes = SentHashes::new(&non_finalized_state);
        let initial_tip = non_finalized_state
            .best_tip_block()
            .map(|cv_block| cv_block.block.clone())
            .or(finalized_tip)
            .map(CheckpointVerifiedBlock::from)
            .map(ChainTipBlock::from);

        tracing::info!(chain_tip = ?initial_tip.as_ref().map(|tip| (tip.hash, tip.height)), "loaded Zebra state cache");

        let (chain_tip_sender, latest_chain_tip, chain_tip_change) =
            ChainTipSender::new(initial_tip, network);

        let finalized_state_for_writing = finalized_state.clone();
        let should_use_finalized_block_write_sender = non_finalized_state.is_chain_set_empty();
        let sync_backup_dir_path = backup_dir_path.filter(|_| skip_backup_task);
        let (
            block_write_sender,
            invalid_block_write_reset_receiver,
            non_finalized_rejected_receiver,
            block_write_task,
        ) = write::BlockWriteSender::spawn(
            finalized_state_for_writing,
            non_finalized_state,
            chain_tip_sender,
            non_finalized_state_sender,
            should_use_finalized_block_write_sender,
            sync_backup_dir_path,
        );

        let read_service = ReadStateService::new(
            &finalized_state,
            block_write_task,
            non_finalized_state_receiver,
        );

        let full_verifier_utxo_lookahead = max_checkpoint_height
            - HeightDiff::try_from(checkpoint_verify_concurrency_limit)
                .expect("fits in HeightDiff");
        let full_verifier_utxo_lookahead =
            full_verifier_utxo_lookahead.unwrap_or(block::Height::MIN);
        let non_finalized_state_queued_blocks = QueuedBlocks::default();
        let pending_utxos = PendingUtxos::default();

        let finalized_block_write_last_sent_hash =
            tokio::task::spawn_blocking(move || finalized_state.db.finalized_tip_hash())
                .await
                .expect("failed to join blocking task");

        let state = Self {
            network: network.clone(),
            full_verifier_utxo_lookahead,
            non_finalized_state_queued_blocks,
            finalized_state_queued_blocks: HashMap::new(),
            block_write_sender,
            finalized_block_write_last_sent_hash,
            non_finalized_block_write_sent_hashes,
            invalid_block_write_reset_receiver,
            non_finalized_rejected_receiver,
            pending_utxos,
            last_prune: Instant::now(),
            read_service: read_service.clone(),
            max_finalized_queue_height: f64::NAN,
        };
        timer.finish_desc("initializing state service");

        tracing::info!("starting legacy chain check");
        let timer = CodeTimer::start();

        if let (Some(tip), Some(nu5_activation_height)) = (
            {
                let read_state = state.read_service.clone();
                tokio::task::spawn_blocking(move || read_state.best_tip())
                    .await
                    .expect("task should not panic")
            },
            NetworkUpgrade::Nu5.activation_height(network),
        ) {
            if let Err(error) = check::legacy_chain(
                nu5_activation_height,
                any_ancestor_blocks(
                    &state.read_service.latest_non_finalized_state(),
                    &state.read_service.db,
                    tip.1,
                ),
                &state.network,
                MAX_LEGACY_CHAIN_BLOCKS,
            ) {
                let legacy_db_path = state.read_service.db.path().to_path_buf();
                panic!(
                    "Cached state contains a legacy chain.\n\
                     An outdated Zebra version did not know about a recent network upgrade,\n\
                     so it followed a legacy chain using outdated consensus branch rules.\n\
                     Hint: Delete your database, and restart Zebra to do a full sync.\n\
                     Database path: {legacy_db_path:?}\n\
                     Error: {error:?}",
                );
            }
        }

        tracing::info!("cached state consensus branch is valid: no legacy chain found");
        timer.finish_desc("legacy chain check");

        // Spawn a background task to periodically export RocksDB metrics to Prometheus
        let db_for_metrics = read_service.db.clone();
        tokio::spawn(async move {
            let mut interval = tokio::time::interval(Duration::from_secs(30));
            loop {
                interval.tick().await;
                db_for_metrics.export_metrics();
            }
        });

        (state, read_service, latest_chain_tip, chain_tip_change)
    }

    /// Call read only state service to log rocksdb database metrics.
    pub fn log_db_metrics(&self) {
        self.read_service.db.print_db_metrics();
    }

    /// Queue a checkpoint verified block for verification and storage in the finalized state.
    ///
    /// Returns a channel receiver that provides the result of the block commit.
    fn queue_and_commit_to_finalized_state(
        &mut self,
        checkpoint_verified: CheckpointVerifiedBlock,
    ) -> oneshot::Receiver<Result<block::Hash, CommitCheckpointVerifiedError>> {
        // # Correctness & Performance
        //
        // This method must not block, access the database, or perform CPU-intensive tasks,
        // because it is called directly from the tokio executor's Future threads.

        let queued_prev_hash = checkpoint_verified.block.header.previous_block_hash;
        let queued_height = checkpoint_verified.height;

        // If we're close to the final checkpoint, make the block's UTXOs available for
        // semantic block verification, even when it is in the channel.
        if self.is_close_to_final_checkpoint(queued_height) {
            self.non_finalized_block_write_sent_hashes
                .add_finalized(&checkpoint_verified)
        }

        let (rsp_tx, rsp_rx) = oneshot::channel();
        let queued = (checkpoint_verified, rsp_tx);

        if self.block_write_sender.finalized.is_some() {
            // We're still committing checkpoint verified blocks
            if let Some(duplicate_queued) = self
                .finalized_state_queued_blocks
                .insert(queued_prev_hash, queued)
            {
                Self::send_checkpoint_verified_block_error(
                    duplicate_queued,
                    CommitBlockError::new_duplicate(
                        Some(queued_prev_hash.into()),
                        KnownBlock::Queue,
                    ),
                );
            }

            self.drain_finalized_queue_and_commit();
        } else {
            // We've finished committing checkpoint verified blocks to the finalized state,
            // so drop any repeated queued blocks, and return an error.
            //
            // TODO: track the latest sent height, and drop any blocks under that height
            //       every time we send some blocks (like QueuedSemanticallyVerifiedBlocks)
            Self::send_checkpoint_verified_block_error(
                queued,
                CommitBlockError::new_duplicate(None, KnownBlock::Finalized),
            );

            self.clear_finalized_block_queue(CommitBlockError::new_duplicate(
                None,
                KnownBlock::Finalized,
            ));
        }

        if self.finalized_state_queued_blocks.is_empty() {
            self.max_finalized_queue_height = f64::NAN;
        } else if self.max_finalized_queue_height.is_nan()
            || self.max_finalized_queue_height < queued_height.0 as f64
        {
            // if there are still blocks in the queue, then either:
            //   - the new block was lower than the old maximum, and there was a gap before it,
            //     so the maximum is still the same (and we skip this code), or
            //   - the new block is higher than the old maximum, and there is at least one gap
            //     between the finalized tip and the new maximum
            self.max_finalized_queue_height = queued_height.0 as f64;
        }

        metrics::gauge!("state.checkpoint.queued.max.height").set(self.max_finalized_queue_height);
        metrics::gauge!("state.checkpoint.queued.block.count")
            .set(self.finalized_state_queued_blocks.len() as f64);

        rsp_rx
    }

    /// Finds finalized state queue blocks to be committed to the state in order,
    /// removes them from the queue, and sends them to the block commit task.
    ///
    /// After queueing a finalized block, this method checks whether the newly
    /// queued block (and any of its descendants) can be committed to the state.
    ///
    /// Returns an error if the block commit channel has been closed.
    pub fn drain_finalized_queue_and_commit(&mut self) {
        use tokio::sync::mpsc::error::{SendError, TryRecvError};

        // # Correctness & Performance
        //
        // This method must not block, access the database, or perform CPU-intensive tasks,
        // because it is called directly from the tokio executor's Future threads.

        // If a block failed, we need to start again from a valid tip.
        match self.invalid_block_write_reset_receiver.try_recv() {
            Ok(reset_tip_hash) => self.finalized_block_write_last_sent_hash = reset_tip_hash,
            Err(TryRecvError::Disconnected) => {
                info!("Block commit task closed the block reset channel. Is Zebra shutting down?");
                return;
            }
            // There are no errors, so we can just use the last block hash we sent
            Err(TryRecvError::Empty) => {}
        }

        while let Some(queued_block) = self
            .finalized_state_queued_blocks
            .remove(&self.finalized_block_write_last_sent_hash)
        {
            let last_sent_finalized_block_height = queued_block.0.height;

            self.finalized_block_write_last_sent_hash = queued_block.0.hash;

            // If we've finished sending finalized blocks, ignore any repeated blocks.
            // (Blocks can be repeated after a syncer reset.)
            if let Some(finalized_block_write_sender) = &self.block_write_sender.finalized {
                let send_result = finalized_block_write_sender.send(queued_block);

                // If the receiver is closed, we can't send any more blocks.
                if let Err(SendError(queued)) = send_result {
                    // If Zebra is shutting down, drop blocks and return an error.
                    Self::send_checkpoint_verified_block_error(
                        queued,
                        CommitBlockError::WriteTaskExited,
                    );

                    self.clear_finalized_block_queue(CommitBlockError::WriteTaskExited);
                } else {
                    metrics::gauge!("state.checkpoint.sent.block.height")
                        .set(last_sent_finalized_block_height.0 as f64);
                };
            }
        }
    }

    /// Drains every hash queued on `non_finalized_rejected_receiver` and
    /// removes it from `non_finalized_block_write_sent_hashes`.
    ///
    /// This closes the lockout window where a rejected block keeps its hash
    /// recorded as "sent", so a subsequent honest re-delivery of a block at
    /// the same hash is not short-circuited as a false "duplicate".
    ///
    /// # Correctness & Performance
    ///
    /// Like the other drain methods on `StateService`, this must not block,
    /// access the database, or perform CPU-intensive work, because it is
    /// called directly from the tokio executor's Future threads.
    fn drain_non_finalized_rejected_hashes(&mut self) {
        use tokio::sync::mpsc::error::TryRecvError;

        loop {
            match self.non_finalized_rejected_receiver.try_recv() {
                Ok(hash) => {
                    self.non_finalized_block_write_sent_hashes.remove(&hash);
                }
                Err(TryRecvError::Empty) => break,
                Err(TryRecvError::Disconnected) => {
                    info!(
                        "Block commit task closed the non-finalized rejected hash channel. \
                         Is Zebra shutting down?"
                    );
                    break;
                }
            }
        }
    }

    /// Drops all finalized state queue blocks, and sends an error on their result channels.
    fn clear_finalized_block_queue(
        &mut self,
        error: impl Into<CommitCheckpointVerifiedError> + Clone,
    ) {
        for (_hash, queued) in self.finalized_state_queued_blocks.drain() {
            Self::send_checkpoint_verified_block_error(queued, error.clone());
        }
    }

    /// Send an error on a `QueuedCheckpointVerified` block's result channel, and drop the block
    fn send_checkpoint_verified_block_error(
        queued: QueuedCheckpointVerified,
        error: impl Into<CommitCheckpointVerifiedError>,
    ) {
        let (finalized, rsp_tx) = queued;

        // The block sender might have already given up on this block,
        // so ignore any channel send errors.
        let _ = rsp_tx.send(Err(error.into()));
        std::mem::drop(finalized);
    }

    /// Drops all non-finalized state queue blocks, and sends an error on their result channels.
    fn clear_non_finalized_block_queue(
        &mut self,
        error: impl Into<CommitSemanticallyVerifiedError> + Clone,
    ) {
        for (_hash, queued) in self.non_finalized_state_queued_blocks.drain() {
            Self::send_semantically_verified_block_error(queued, error.clone());
        }
    }

    /// Send an error on a `QueuedSemanticallyVerified` block's result channel, and drop the block
    fn send_semantically_verified_block_error(
        queued: QueuedSemanticallyVerified,
        error: impl Into<CommitSemanticallyVerifiedError>,
    ) {
        let (finalized, rsp_tx) = queued;

        // The block sender might have already given up on this block,
        // so ignore any channel send errors.
        let _ = rsp_tx.send(Err(error.into()));
        std::mem::drop(finalized);
    }

    /// Queue a semantically verified block for contextual verification and check if any queued
    /// blocks are ready to be verified and committed to the state.
    ///
    /// This function encodes the logic for [committing non-finalized blocks][1]
    /// in RFC0005.
    ///
    /// [1]: https://zebra.zfnd.org/dev/rfcs/0005-state-updates.html#committing-non-finalized-blocks
    #[instrument(level = "debug", skip(self, semantically_verified))]
    fn queue_and_commit_to_non_finalized_state(
        &mut self,
        semantically_verified: SemanticallyVerifiedBlock,
    ) -> oneshot::Receiver<Result<block::Hash, CommitSemanticallyVerifiedError>> {
        tracing::debug!(block = %semantically_verified.block, "queueing block for contextual verification");
        let parent_hash = semantically_verified.block.header.previous_block_hash;

        // Drop hashes of any blocks the write task has rejected before checking
        // the SentHashes membership below. Without this, a rejected same-hash
        // block would lock out a later honest re-delivery of a block at the
        // same hash as a false "duplicate".
        self.drain_non_finalized_rejected_hashes();

        if self
            .non_finalized_block_write_sent_hashes
            .contains(&semantically_verified.hash)
        {
            let (rsp_tx, rsp_rx) = oneshot::channel();
            let _ = rsp_tx.send(Err(CommitBlockError::new_duplicate(
                Some(semantically_verified.hash.into()),
                KnownBlock::WriteChannel,
            )
            .into()));
            return rsp_rx;
        }

        if self
            .read_service
            .db
            .contains_height(semantically_verified.height)
        {
            let (rsp_tx, rsp_rx) = oneshot::channel();
            let _ = rsp_tx.send(Err(CommitBlockError::new_duplicate(
                Some(semantically_verified.height.into()),
                KnownBlock::Finalized,
            )
            .into()));
            return rsp_rx;
        }

        // [`Request::CommitSemanticallyVerifiedBlock`] contract: a request to commit a block which
        // has been queued but not yet committed to the state fails the older request and replaces
        // it with the newer request.
        let rsp_rx = if let Some((_, old_rsp_tx)) = self
            .non_finalized_state_queued_blocks
            .get_mut(&semantically_verified.hash)
        {
            tracing::debug!("replacing older queued request with new request");
            let (mut rsp_tx, rsp_rx) = oneshot::channel();
            std::mem::swap(old_rsp_tx, &mut rsp_tx);
            let _ = rsp_tx.send(Err(CommitBlockError::new_duplicate(
                Some(semantically_verified.hash.into()),
                KnownBlock::Queue,
            )
            .into()));
            rsp_rx
        } else {
            let (rsp_tx, rsp_rx) = oneshot::channel();
            self.non_finalized_state_queued_blocks
                .queue((semantically_verified, rsp_tx));
            rsp_rx
        };

        // We've finished sending checkpoint verified blocks when:
        // - we've sent the verified block for the last checkpoint, and
        // - it has been successfully written to disk.
        //
        // We detect the last checkpoint by looking for non-finalized blocks
        // that are a child of the last block we sent.
        //
        // TODO: configure the state with the last checkpoint hash instead?
        if self.block_write_sender.finalized.is_some()
            && self
                .non_finalized_state_queued_blocks
                .has_queued_children(self.finalized_block_write_last_sent_hash)
            && self.read_service.db.finalized_tip_hash()
                == self.finalized_block_write_last_sent_hash
        {
            // Tell the block write task to stop committing checkpoint verified blocks to the finalized state,
            // and move on to committing semantically verified blocks to the non-finalized state.
            std::mem::drop(self.block_write_sender.finalized.take());
            // Remove any checkpoint-verified block hashes from `non_finalized_block_write_sent_hashes`.
            self.non_finalized_block_write_sent_hashes = SentHashes::default();
            // Mark `SentHashes` as usable by the `can_fork_chain_at()` method.
            self.non_finalized_block_write_sent_hashes
                .can_fork_chain_at_hashes = true;
            // Send blocks from non-finalized queue
            self.send_ready_non_finalized_queued(self.finalized_block_write_last_sent_hash);
            // We've finished committing checkpoint verified blocks to finalized state, so drop any repeated queued blocks.
            self.clear_finalized_block_queue(CommitBlockError::new_duplicate(
                None,
                KnownBlock::Finalized,
            ));
        } else if !self.can_fork_chain_at(&parent_hash) {
            tracing::trace!("unready to verify, returning early");
        } else if self.block_write_sender.finalized.is_none() {
            // Wait until block commit task is ready to write non-finalized blocks before dequeuing them
            self.send_ready_non_finalized_queued(parent_hash);

            let finalized_tip_height = self.read_service.db.finalized_tip_height().expect(
                "Finalized state must have at least one block before committing non-finalized state",
            );

            self.non_finalized_state_queued_blocks
                .prune_by_height(finalized_tip_height);

            self.non_finalized_block_write_sent_hashes
                .prune_by_height(finalized_tip_height);
        }

        rsp_rx
    }

    /// Returns `true` if `hash` is a valid previous block hash for new non-finalized blocks.
    fn can_fork_chain_at(&self, hash: &block::Hash) -> bool {
        self.non_finalized_block_write_sent_hashes
            .can_fork_chain_at(hash)
            || &self.read_service.db.finalized_tip_hash() == hash
    }

    /// Returns `true` if `queued_height` is near the final checkpoint.
    ///
    /// The semantic block verifier needs access to UTXOs from checkpoint verified blocks
    /// near the final checkpoint, so that it can verify blocks that spend those UTXOs.
    ///
    /// If it doesn't have the required UTXOs, some blocks will time out,
    /// but succeed after a syncer restart.
    fn is_close_to_final_checkpoint(&self, queued_height: block::Height) -> bool {
        queued_height >= self.full_verifier_utxo_lookahead
    }

    /// Sends all queued blocks whose parents have recently arrived starting from `new_parent`
    /// in breadth-first ordering to the block write task which will attempt to validate and commit them
    #[tracing::instrument(level = "debug", skip(self, new_parent))]
    fn send_ready_non_finalized_queued(&mut self, new_parent: block::Hash) {
        use tokio::sync::mpsc::error::SendError;
        if let Some(non_finalized_block_write_sender) = &self.block_write_sender.non_finalized {
            let mut new_parents: Vec<block::Hash> = vec![new_parent];

            while let Some(parent_hash) = new_parents.pop() {
                let queued_children = self
                    .non_finalized_state_queued_blocks
                    .dequeue_children(parent_hash);

                for queued_child in queued_children {
                    let (SemanticallyVerifiedBlock { hash, .. }, _) = queued_child;

                    self.non_finalized_block_write_sent_hashes
                        .add(&queued_child.0);
                    let send_result = non_finalized_block_write_sender.send(queued_child.into());

                    if let Err(SendError(NonFinalizedWriteMessage::Commit(queued))) = send_result {
                        // If Zebra is shutting down, drop blocks and return an error.
                        Self::send_semantically_verified_block_error(
                            queued,
                            CommitBlockError::WriteTaskExited,
                        );

                        self.clear_non_finalized_block_queue(CommitBlockError::WriteTaskExited);

                        return;
                    };

                    new_parents.push(hash);
                }
            }

            self.non_finalized_block_write_sent_hashes.finish_batch();
        };
    }

    /// Return the tip of the current best chain.
    pub fn best_tip(&self) -> Option<(block::Height, block::Hash)> {
        self.read_service.best_tip()
    }

    fn send_invalidate_block(
        &self,
        hash: block::Hash,
    ) -> oneshot::Receiver<Result<block::Hash, InvalidateError>> {
        let (rsp_tx, rsp_rx) = oneshot::channel();

        let Some(sender) = &self.block_write_sender.non_finalized else {
            let _ = rsp_tx.send(Err(InvalidateError::ProcessingCheckpointedBlocks));
            return rsp_rx;
        };

        if let Err(tokio::sync::mpsc::error::SendError(error)) =
            sender.send(NonFinalizedWriteMessage::Invalidate { hash, rsp_tx })
        {
            let NonFinalizedWriteMessage::Invalidate { rsp_tx, .. } = error else {
                unreachable!("should return the same Invalidate message could not be sent");
            };

            let _ = rsp_tx.send(Err(InvalidateError::SendInvalidateRequestFailed));
        }

        rsp_rx
    }

    fn send_reconsider_block(
        &self,
        hash: block::Hash,
    ) -> oneshot::Receiver<Result<Vec<block::Hash>, ReconsiderError>> {
        let (rsp_tx, rsp_rx) = oneshot::channel();

        let Some(sender) = &self.block_write_sender.non_finalized else {
            let _ = rsp_tx.send(Err(ReconsiderError::CheckpointCommitInProgress));
            return rsp_rx;
        };

        if let Err(tokio::sync::mpsc::error::SendError(error)) =
            sender.send(NonFinalizedWriteMessage::Reconsider { hash, rsp_tx })
        {
            let NonFinalizedWriteMessage::Reconsider { rsp_tx, .. } = error else {
                unreachable!("should return the same Reconsider message could not be sent");
            };

            let _ = rsp_tx.send(Err(ReconsiderError::ReconsiderSendFailed));
        }

        rsp_rx
    }

    /// Assert some assumptions about the semantically verified `block` before it is queued.
    fn assert_block_can_be_validated(&self, block: &SemanticallyVerifiedBlock) {
        // required by `Request::CommitSemanticallyVerifiedBlock` call
        assert!(
            block.height > self.network.mandatory_checkpoint_height(),
            "invalid semantically verified block height: the canopy checkpoint is mandatory, pre-canopy \
            blocks, and the canopy activation block, must be committed to the state as finalized \
            blocks"
        );
    }

    fn known_sent_hash(&self, hash: &block::Hash) -> Option<KnownBlock> {
        self.non_finalized_block_write_sent_hashes
            .contains(hash)
            .then_some(KnownBlock::WriteChannel)
    }
}

impl ReadStateService {
    /// Creates a new read-only state service, using the provided finalized state and
    /// block write task handle.
    ///
    /// Returns the newly created service,
    /// and a watch channel for updating the shared recent non-finalized chain.
    pub(crate) fn new(
        finalized_state: &FinalizedState,
        block_write_task: Option<Arc<std::thread::JoinHandle<()>>>,
        non_finalized_state_receiver: WatchReceiver<NonFinalizedState>,
    ) -> Self {
        let read_service = Self {
            network: finalized_state.network(),
            db: finalized_state.db.clone(),
            non_finalized_state_receiver,
            block_write_task,
        };

        tracing::debug!("created new read-only state service");

        read_service
    }

    /// Return the tip of the current best chain.
    pub fn best_tip(&self) -> Option<(block::Height, block::Hash)> {
        read::best_tip(&self.latest_non_finalized_state(), &self.db)
    }

    /// Gets a clone of the latest non-finalized state from the `non_finalized_state_receiver`
    fn latest_non_finalized_state(&self) -> NonFinalizedState {
        self.non_finalized_state_receiver.cloned_watch_data()
    }

    /// Gets a clone of the latest, best non-finalized chain from the `non_finalized_state_receiver`
    fn latest_best_chain(&self) -> Option<Arc<Chain>> {
        self.non_finalized_state_receiver
            .borrow_mapped(|non_finalized_state| non_finalized_state.best_chain().cloned())
    }

    /// Test-only access to the inner database.
    /// Can be used to modify the database without doing any consensus checks.
    #[cfg(any(test, feature = "proptest-impl"))]
    pub fn db(&self) -> &ZebraDb {
        &self.db
    }

    /// Logs rocksdb metrics using the read only state service.
    pub fn log_db_metrics(&self) {
        self.db.print_db_metrics();
    }
}

impl Service<Request> for StateService {
    type Response = Response;
    type Error = BoxError;
    type Future =
        Pin<Box<dyn Future<Output = Result<Self::Response, Self::Error>> + Send + 'static>>;

    fn poll_ready(&mut self, cx: &mut Context<'_>) -> Poll<Result<(), Self::Error>> {
        // Check for panics in the block write task
        let poll = self.read_service.poll_ready(cx);

        // Prune outdated UTXO requests
        let now = Instant::now();

        if self.last_prune + Self::PRUNE_INTERVAL < now {
            let tip = self.best_tip();
            let old_len = self.pending_utxos.len();

            self.pending_utxos.prune();
            self.last_prune = now;

            let new_len = self.pending_utxos.len();
            let prune_count = old_len
                .checked_sub(new_len)
                .expect("prune does not add any utxo requests");
            if prune_count > 0 {
                tracing::debug!(
                    ?old_len,
                    ?new_len,
                    ?prune_count,
                    ?tip,
                    "pruned utxo requests"
                );
            } else {
                tracing::debug!(len = ?old_len, ?tip, "no utxo requests needed pruning");
            }
        }

        poll
    }

    #[instrument(name = "state", skip(self, req))]
    fn call(&mut self, req: Request) -> Self::Future {
        req.count_metric();
        let span = Span::current();

        match req {
            // Uses non_finalized_state_queued_blocks and pending_utxos in the StateService
            // Accesses shared writeable state in the StateService, NonFinalizedState, and ZebraDb.
            //
            // The expected error type for this request is `CommitSemanticallyVerifiedError`.
            Request::CommitSemanticallyVerifiedBlock(semantically_verified) => {
                let timer = CodeTimer::start();
                self.assert_block_can_be_validated(&semantically_verified);

                self.pending_utxos
                    .check_against_ordered(&semantically_verified.new_outputs);

                // # Performance
                //
                // Allow other async tasks to make progress while blocks are being verified
                // and written to disk. But wait for the blocks to finish committing,
                // so that `StateService` multi-block queries always observe a consistent state.
                //
                // Since each block is spawned into its own task,
                // there shouldn't be any other code running in the same task,
                // so we don't need to worry about blocking it:
                // https://docs.rs/tokio/latest/tokio/task/fn.block_in_place.html

                let rsp_rx = tokio::task::block_in_place(move || {
                    span.in_scope(|| {
                        self.queue_and_commit_to_non_finalized_state(semantically_verified)
                    })
                });

                // TODO:
                //   - check for panics in the block write task here,
                //     as well as in poll_ready()

                // The work is all done, the future just waits on a channel for the result
                timer.finish_desc("CommitSemanticallyVerifiedBlock");

                // Await the channel response, flatten the result, map receive errors to
                // `CommitSemanticallyVerifiedError::WriteTaskExited`.
                // Then flatten the nested Result and convert any errors to a BoxError.
                let span = Span::current();
                async move {
                    rsp_rx
                        .await
                        .map_err(|_recv_error| CommitBlockError::WriteTaskExited.into())
                        .and_then(|result| result)
                        .map_err(BoxError::from)
                        .map(Response::Committed)
                }
                .instrument(span)
                .boxed()
            }

            // Uses finalized_state_queued_blocks and pending_utxos in the StateService.
            // Accesses shared writeable state in the StateService.
            //
            // The expected error type for this request is `CommitCheckpointVerifiedError`.
            Request::CommitCheckpointVerifiedBlock(finalized) => {
                let timer = CodeTimer::start();
                // # Consensus
                //
                // A semantic block verification could have called AwaitUtxo
                // before this checkpoint verified block arrived in the state.
                // So we need to check for pending UTXO requests sent by running
                // semantic block verifications.
                //
                // This check is redundant for most checkpoint verified blocks,
                // because semantic verification can only succeed near the final
                // checkpoint, when all the UTXOs are available for the verifying block.
                //
                // (Checkpoint block UTXOs are verified using block hash checkpoints
                // and transaction merkle tree block header commitments.)
                self.pending_utxos
                    .check_against_ordered(&finalized.new_outputs);

                // # Performance
                //
                // This method doesn't block, access the database, or perform CPU-intensive tasks,
                // so we can run it directly in the tokio executor's Future threads.
                let rsp_rx = self.queue_and_commit_to_finalized_state(finalized);

                // TODO:
                //   - check for panics in the block write task here,
                //     as well as in poll_ready()

                // The work is all done, the future just waits on a channel for the result
                timer.finish_desc("CommitCheckpointVerifiedBlock");

                // Await the channel response, flatten the result, map receive errors to
                // `CommitCheckpointVerifiedError::WriteTaskExited`.
                // Then flatten the nested Result and convert any errors to a BoxError.
                async move {
                    rsp_rx
                        .await
                        .map_err(|_recv_error| CommitBlockError::WriteTaskExited.into())
                        .and_then(|result| result)
                        .map_err(BoxError::from)
                        .map(Response::Committed)
                }
                .instrument(span)
                .boxed()
            }

            // Uses pending_utxos and non_finalized_state_queued_blocks in the StateService.
            // If the UTXO isn't in the queued blocks, runs concurrently using the ReadStateService.
            //
            // The expected error type for this request is `AwaitUtxoError`.
            Request::AwaitUtxo(outpoint) => {
                let timer = CodeTimer::start();
                // Prepare the AwaitUtxo future from PendingUtxos.
                let response_fut = self.pending_utxos.queue(outpoint);
                // Only instrument `response_fut`, the ReadStateService already
                // instruments its requests with the same span.

                let response_fut = response_fut
                    .map(|result| result.map_err(BoxError::from))
                    .instrument(span)
                    .boxed();

                // Check the non-finalized block queue outside the returned future,
                // so we can access mutable state fields.
                if let Some(utxo) = self.non_finalized_state_queued_blocks.utxo(&outpoint) {
                    self.pending_utxos.respond(&outpoint, utxo);

                    // We're finished, the returned future gets the UTXO from the respond() channel.
                    timer.finish_desc("AwaitUtxo/queued-non-finalized");

                    return response_fut;
                }

                // Check the sent non-finalized blocks
                self.drain_non_finalized_rejected_hashes();

                if let Some(utxo) = self.non_finalized_block_write_sent_hashes.utxo(&outpoint) {
                    self.pending_utxos.respond(&outpoint, utxo);

                    // We're finished, the returned future gets the UTXO from the respond() channel.
                    timer.finish_desc("AwaitUtxo/sent-non-finalized");

                    return response_fut;
                }

                // We ignore any UTXOs in FinalizedState.finalized_state_queued_blocks,
                // because it is only used during checkpoint verification.
                //
                // This creates a rare race condition, but it doesn't seem to happen much in practice.
                // See #5126 for details.

                // Manually send a request to the ReadStateService,
                // to get UTXOs from any non-finalized chain or the finalized chain.
                let read_service = self.read_service.clone();

                // Run the request in an async block, so we can await the response.
                async move {
                    let req = ReadRequest::AnyChainUtxo(outpoint);

                    let rsp = read_service
                        .oneshot(req)
                        .await
                        .map_err(AwaitUtxoError::ReadStateFailed)?;

                    // Optional TODO:
                    //  - make pending_utxos.respond() async using a channel,
                    //    so we can respond to all waiting requests here
                    //
                    // This change is not required for correctness, because:
                    // - any waiting requests should have returned when the block was sent to the state
                    // - otherwise, the request returns immediately if:
                    //   - the block is in the non-finalized queue, or
                    //   - the block is in any non-finalized chain or the finalized state
                    //
                    // And if the block is in the finalized queue,
                    // that's rare enough that a retry is ok.
                    if let ReadResponse::AnyChainUtxo(Some(utxo)) = rsp {
                        // We got a UTXO, so we replace the response future with the result own.
                        timer.finish_desc("AwaitUtxo/any-chain");

                        return Ok(Response::Utxo(utxo));
                    }

                    // We're finished, but the returned future is waiting on the respond() channel.
                    timer.finish_desc("AwaitUtxo/waiting");

                    response_fut.await
                }
                .boxed()
            }

            // Used by sync, inbound, and block verifier to check if a block is already in the state
            // before downloading or validating it.
            Request::KnownBlock(hash) => {
                let timer = CodeTimer::start();

                self.drain_non_finalized_rejected_hashes();

                let known_sent_hash = self.known_sent_hash(&hash);
                let known_queued = self
                    .non_finalized_state_queued_blocks
                    .has(hash)
                    .then_some(KnownBlock::Queue);
                let read_service = self.read_service.clone();

                async move {
                    if let Some(loc) = known_sent_hash.or(known_queued) {
                        return Ok(Response::KnownBlock(Some(loc)));
                    };

                    let response = read::non_finalized_state_contains_block_hash(
                        &read_service.latest_non_finalized_state(),
                        hash,
                    )
                    // TODO: Move this to a blocking task, perhaps by moving some of this logic to the ReadStateService.
                    .or_else(|| read::finalized_state_contains_block_hash(&read_service.db, hash));

                    timer.finish_desc("Request::KnownBlock");

                    Ok(Response::KnownBlock(response))
                }
                .boxed()
            }

            // The expected error type for this request is `InvalidateError`
            Request::InvalidateBlock(block_hash) => {
                let rsp_rx = tokio::task::block_in_place(move || {
                    span.in_scope(|| self.send_invalidate_block(block_hash))
                });

                // Await the channel response, flatten the result, map receive errors to
                // `InvalidateError::InvalidateRequestDropped`.
                // Then flatten the nested Result and convert any errors to a BoxError.
                let span = Span::current();
                async move {
                    rsp_rx
                        .await
                        .map_err(|_recv_error| InvalidateError::InvalidateRequestDropped)
                        .and_then(|result| result)
                        .map_err(BoxError::from)
                        .map(Response::Invalidated)
                }
                .instrument(span)
                .boxed()
            }

            // The expected error type for this request is `ReconsiderError`
            Request::ReconsiderBlock(block_hash) => {
                let rsp_rx = tokio::task::block_in_place(move || {
                    span.in_scope(|| self.send_reconsider_block(block_hash))
                });

                // Await the channel response, flatten the result, map receive errors to
                // `ReconsiderError::ReconsiderResponseDropped`.
                // Then flatten the nested Result and convert any errors to a BoxError.
                let span = Span::current();
                async move {
                    rsp_rx
                        .await
                        .map_err(|_recv_error| ReconsiderError::ReconsiderResponseDropped)
                        .and_then(|result| result)
                        .map_err(BoxError::from)
                        .map(Response::Reconsidered)
                }
                .instrument(span)
                .boxed()
            }

            // Runs concurrently using the ReadStateService
            Request::Tip
            | Request::Depth(_)
            | Request::BestChainNextMedianTimePast
            | Request::BestChainBlockHash(_)
            | Request::BlockLocator
            | Request::Transaction(_)
            | Request::AnyChainTransaction(_)
            | Request::UnspentBestChainUtxo(_)
            | Request::Block(_)
            | Request::AnyChainBlock(_)
            | Request::BlockAndSize(_)
            | Request::BlockHeader(_)
            | Request::FindBlockHashes { .. }
            | Request::FindBlockHeaders { .. }
            | Request::CheckBestChainTipNullifiersAndAnchors(_)
            | Request::CheckBlockProposalValidity(_) => {
                // Redirect the request to the concurrent ReadStateService
                let read_service = self.read_service.clone();

                async move {
                    let req = req
                        .try_into()
                        .expect("ReadRequest conversion should not fail");

                    let rsp = read_service.oneshot(req).await?;
                    let rsp = rsp.try_into().expect("Response conversion should not fail");

                    Ok(rsp)
                }
                .boxed()
            }
        }
    }
}

impl Service<ReadRequest> for ReadStateService {
    type Response = ReadResponse;
    type Error = BoxError;
    type Future =
        Pin<Box<dyn Future<Output = Result<Self::Response, Self::Error>> + Send + 'static>>;

    fn poll_ready(&mut self, _: &mut Context<'_>) -> Poll<Result<(), Self::Error>> {
        // Check for panics in the block write task
        //
        // TODO: move into a check_for_panics() method
        let block_write_task = self.block_write_task.take();

        if let Some(block_write_task) = block_write_task {
            if block_write_task.is_finished() {
                if let Some(block_write_task) = Arc::into_inner(block_write_task) {
                    // We are the last state with a reference to this task, so we can propagate any panics
                    if let Err(thread_panic) = block_write_task.join() {
                        std::panic::resume_unwind(thread_panic);
                    }
                }
            } else {
                // It hasn't finished, so we need to put it back
                self.block_write_task = Some(block_write_task);
            }
        }

        self.db.check_for_panics();

        Poll::Ready(Ok(()))
    }

    #[instrument(name = "read_state", skip(self, req))]
    fn call(&mut self, req: ReadRequest) -> Self::Future {
        req.count_metric();
        let timer = CodeTimer::start_desc(req.variant_name());
        let span = Span::current();
        let timed_span = TimedSpan::new(timer, span);
        let state = self.clone();

        if let ReadRequest::NonFinalizedBlocksListener { known_chain_tips } = req {
            // The non-finalized blocks listener is used to notify the state service
            // about new blocks that have been added to the non-finalized state.
            let non_finalized_blocks_listener = NonFinalizedBlocksListener::spawn(
                self.non_finalized_state_receiver.clone(),
                known_chain_tips,
            );

            return async move {
                Ok(ReadResponse::NonFinalizedBlocksListener(
                    non_finalized_blocks_listener,
                ))
            }
            .boxed();
        };

        let request_handler = move || match req {
            // Used by the `getblockchaininfo` RPC.
            ReadRequest::UsageInfo => Ok(ReadResponse::UsageInfo(state.db.size())),

            // Used by the StateService.
            ReadRequest::Tip => Ok(ReadResponse::Tip(read::tip(
                state.latest_best_chain(),
                &state.db,
            ))),

            // Used by `getblockchaininfo` RPC method.
            ReadRequest::TipPoolValues => {
                let (tip_height, tip_hash, value_balance) =
                    read::tip_with_value_balance(state.latest_best_chain(), &state.db)?
                        .ok_or(BoxError::from("no chain tip available yet"))?;

                Ok(ReadResponse::TipPoolValues {
                    tip_height,
                    tip_hash,
                    value_balance,
                })
            }

            // Used by getblock
            ReadRequest::BlockInfo(hash_or_height) => Ok(ReadResponse::BlockInfo(
                read::block_info(state.latest_best_chain(), &state.db, hash_or_height),
            )),

            // Used by the StateService.
            ReadRequest::Depth(hash) => Ok(ReadResponse::Depth(read::depth(
                state.latest_best_chain(),
                &state.db,
                hash,
            ))),

            // Used by the StateService.
            ReadRequest::BestChainNextMedianTimePast => {
                Ok(ReadResponse::BestChainNextMedianTimePast(
                    read::next_median_time_past(&state.latest_non_finalized_state(), &state.db)?,
                ))
            }

            // Used by the get_block (raw) RPC and the StateService.
            ReadRequest::Block(hash_or_height) => Ok(ReadResponse::Block(read::block(
                state.latest_best_chain(),
                &state.db,
                hash_or_height,
            ))),

            ReadRequest::AnyChainBlock(hash_or_height) => Ok(ReadResponse::Block(read::any_block(
                state.latest_non_finalized_state().chain_iter(),
                &state.db,
                hash_or_height,
            ))),

            // Used by the get_block (raw) RPC and the StateService.
            ReadRequest::BlockAndSize(hash_or_height) => Ok(ReadResponse::BlockAndSize(
                read::block_and_size(state.latest_best_chain(), &state.db, hash_or_height),
            )),

            // Used by the get_block (verbosity 3) RPC.
            ReadRequest::SpentOutputs(hash_or_height) => Ok(ReadResponse::SpentOutputs(
                read::spent_outputs_for_block(state.latest_best_chain(), &state.db, hash_or_height),
            )),

            // Used by the get_block (verbose) RPC and the StateService.
            ReadRequest::BlockHeader(hash_or_height) => {
                let best_chain = state.latest_best_chain();

                let height = hash_or_height
                    .height_or_else(|hash| {
                        read::find::height_by_hash(best_chain.clone(), &state.db, hash)
                    })
                    .ok_or_else(|| BoxError::from("block hash or height not found"))?;

                let hash = hash_or_height
                    .hash_or_else(|height| {
                        read::find::hash_by_height(best_chain.clone(), &state.db, height)
                    })
                    .ok_or_else(|| BoxError::from("block hash or height not found"))?;

                let next_height = height.next()?;
                let next_block_hash =
                    read::find::hash_by_height(best_chain.clone(), &state.db, next_height);

                let header = read::block_header(best_chain, &state.db, height.into())
                    .ok_or_else(|| BoxError::from("block hash or height not found"))?;

                Ok(ReadResponse::BlockHeader {
                    header,
                    hash,
                    height,
                    next_block_hash,
                })
            }

            // For the get_raw_transaction RPC and the StateService.
            ReadRequest::Transaction(hash) => Ok(ReadResponse::Transaction(
                read::mined_transaction(state.latest_best_chain(), &state.db, hash),
            )),

            ReadRequest::AnyChainTransaction(hash) => {
                Ok(ReadResponse::AnyChainTransaction(read::any_transaction(
                    state.latest_non_finalized_state().chain_iter(),
                    &state.db,
                    hash,
                )))
            }

            // Used by the getblock (verbose) RPC.
            ReadRequest::TransactionIdsForBlock(hash_or_height) => Ok(
                ReadResponse::TransactionIdsForBlock(read::transaction_hashes_for_block(
                    state.latest_best_chain(),
                    &state.db,
                    hash_or_height,
                )),
            ),

            ReadRequest::AnyChainTransactionIdsForBlock(hash_or_height) => {
                Ok(ReadResponse::AnyChainTransactionIdsForBlock(
                    read::transaction_hashes_for_any_block(
                        state.latest_non_finalized_state().chain_iter(),
                        &state.db,
                        hash_or_height,
                    ),
                ))
            }

            #[cfg(feature = "indexer")]
            ReadRequest::SpendingTransactionId(spend) => Ok(ReadResponse::TransactionId(
                read::spending_transaction_hash(state.latest_best_chain(), &state.db, spend),
            )),

            ReadRequest::UnspentBestChainUtxo(outpoint) => Ok(ReadResponse::UnspentBestChainUtxo(
                read::unspent_utxo(state.latest_best_chain(), &state.db, outpoint),
            )),

            // Manually used by the StateService to implement part of AwaitUtxo.
            ReadRequest::AnyChainUtxo(outpoint) => Ok(ReadResponse::AnyChainUtxo(read::any_utxo(
                state.latest_non_finalized_state(),
                &state.db,
                outpoint,
            ))),

            // Used by the StateService.
            ReadRequest::BlockLocator => Ok(ReadResponse::BlockLocator(
                read::block_locator(state.latest_best_chain(), &state.db).unwrap_or_default(),
            )),

            // Used by the StateService.
            ReadRequest::FindBlockHashes { known_blocks, stop } => {
                Ok(ReadResponse::BlockHashes(read::find_chain_hashes(
                    state.latest_best_chain(),
                    &state.db,
                    known_blocks,
                    stop,
                    MAX_FIND_BLOCK_HASHES_RESULTS,
                )))
            }

            // Used by the StateService.
            ReadRequest::FindBlockHeaders { known_blocks, stop } => Ok(ReadResponse::BlockHeaders(
                read::find_chain_headers(
                    state.latest_best_chain(),
                    &state.db,
                    known_blocks,
                    stop,
                    MAX_FIND_BLOCK_HEADERS_RESULTS,
                )
                .into_iter()
                .map(|header| CountedHeader { header })
                .collect(),
            )),

            ReadRequest::FindForkPoint { known_blocks } => {
                // Reject over-long locators before doing any work, so an untrusted
                // caller can't force unbounded lookups.
                let locator_len: u64 = known_blocks
                    .len()
                    .try_into()
                    .expect("usize always fits in u64 on supported (<=64-bit) platforms");
                if locator_len > block::MAX_BLOCK_LOCATOR_LENGTH {
                    return Err(BoxError::from(format!(
                        "FindForkPoint locator length {locator_len} exceeds \
                         MAX_BLOCK_LOCATOR_LENGTH ({})",
                        block::MAX_BLOCK_LOCATOR_LENGTH,
                    )));
                }

                Ok(ReadResponse::ForkPoint(read::find_fork_point(
                    state.latest_best_chain(),
                    &state.db,
                    known_blocks,
                )))
            }

            ReadRequest::SaplingTree(hash_or_height) => Ok(ReadResponse::SaplingTree(
                read::sapling_tree(state.latest_best_chain(), &state.db, hash_or_height),
            )),

            ReadRequest::OrchardTree(hash_or_height) => Ok(ReadResponse::OrchardTree(
                read::orchard_tree(state.latest_best_chain(), &state.db, hash_or_height),
            )),

            ReadRequest::AnyChainSaplingTree(hash) => {
                Ok(ReadResponse::SaplingTree(read::any_sapling_tree(
                    state.latest_non_finalized_state().chain_iter(),
                    &state.db,
                    hash,
                )))
            }

            ReadRequest::AnyChainOrchardTree(hash) => {
                Ok(ReadResponse::OrchardTree(read::any_orchard_tree(
                    state.latest_non_finalized_state().chain_iter(),
                    &state.db,
                    hash,
                )))
            }

            ReadRequest::IronwoodTree(hash_or_height) => Ok(ReadResponse::IronwoodTree(
                read::ironwood_tree(state.latest_best_chain(), &state.db, hash_or_height),
            )),

            ReadRequest::AnyChainIronwoodTree(hash) => {
                Ok(ReadResponse::IronwoodTree(read::any_ironwood_tree(
                    state.latest_non_finalized_state().chain_iter(),
                    &state.db,
                    hash,
                )))
            }

            ReadRequest::SaplingSubtrees { start_index, limit } => {
                let end_index = limit
                    .and_then(|limit| start_index.0.checked_add(limit.0))
                    .map(NoteCommitmentSubtreeIndex);

                let best_chain = state.latest_best_chain();
                let sapling_subtrees = if let Some(end_index) = end_index {
                    read::sapling_subtrees(best_chain, &state.db, start_index..end_index)
                } else {
                    // If there is no end bound, just return all the trees.
                    // If the end bound would overflow, just returns all the trees, because that's what
                    // `zcashd` does. (It never calculates an end bound, so it just keeps iterating until
                    // the trees run out.)
                    read::sapling_subtrees(best_chain, &state.db, start_index..)
                };

                Ok(ReadResponse::SaplingSubtrees(sapling_subtrees))
            }

            ReadRequest::OrchardSubtrees { start_index, limit } => {
                let end_index = limit
                    .and_then(|limit| start_index.0.checked_add(limit.0))
                    .map(NoteCommitmentSubtreeIndex);

                let best_chain = state.latest_best_chain();
                let orchard_subtrees = if let Some(end_index) = end_index {
                    read::orchard_subtrees(best_chain, &state.db, start_index..end_index)
                } else {
                    // If there is no end bound, just return all the trees.
                    // If the end bound would overflow, just returns all the trees, because that's what
                    // `zcashd` does. (It never calculates an end bound, so it just keeps iterating until
                    // the trees run out.)
                    read::orchard_subtrees(best_chain, &state.db, start_index..)
                };

                Ok(ReadResponse::OrchardSubtrees(orchard_subtrees))
            }

            ReadRequest::IronwoodSubtrees { start_index, limit } => {
                let end_index = limit
                    .and_then(|limit| start_index.0.checked_add(limit.0))
                    .map(NoteCommitmentSubtreeIndex);

                let best_chain = state.latest_best_chain();
                let ironwood_subtrees = if let Some(end_index) = end_index {
                    read::ironwood_subtrees(best_chain, &state.db, start_index..end_index)
                } else {
                    // If there is no end bound, just return all the trees.
                    // If the end bound would overflow, just returns all the trees, because that's what
                    // `zcashd` does. (It never calculates an end bound, so it just keeps iterating until
                    // the trees run out.)
                    read::ironwood_subtrees(best_chain, &state.db, start_index..)
                };

                Ok(ReadResponse::IronwoodSubtrees(ironwood_subtrees))
            }

            // For the get_address_balance RPC.
            ReadRequest::AddressBalance(addresses) => {
                let (balance, received) =
                    read::transparent_balance(state.latest_best_chain(), &state.db, addresses)?;
                Ok(ReadResponse::AddressBalance { balance, received })
            }

            // For the get_address_tx_ids RPC.
            ReadRequest::TransactionIdsByAddresses {
                addresses,
                height_range,
                max_entries,
            } => read::transparent_tx_ids(
                state.latest_best_chain(),
                &state.db,
                addresses,
                height_range,
                max_entries,
            )
            .map(ReadResponse::AddressesTransactionIds),

            // For the get_address_utxos RPC.
            ReadRequest::UtxosByAddresses {
                addresses,
                height_range,
                max_entries,
            } => read::address_utxos(
                &state.network,
                state.latest_best_chain(),
                &state.db,
                addresses,
                height_range,
                max_entries,
            )
            .map(ReadResponse::AddressUtxos),

            ReadRequest::CheckBestChainTipNullifiersAndAnchors(unmined_tx) => {
                let latest_non_finalized_best_chain = state.latest_best_chain();

                check::nullifier::tx_no_duplicates_in_chain(
                    &state.db,
                    latest_non_finalized_best_chain.as_ref(),
                    &unmined_tx.transaction,
                )?;

                check::anchors::tx_anchors_refer_to_final_treestates(
                    &state.db,
                    latest_non_finalized_best_chain.as_ref(),
                    &unmined_tx,
                )?;

                Ok(ReadResponse::ValidBestChainTipNullifiersAndAnchors)
            }

            // Used by the get_block and get_block_hash RPCs.
            ReadRequest::BestChainBlockHash(height) => Ok(ReadResponse::BlockHash(
                read::hash_by_height(state.latest_best_chain(), &state.db, height),
            )),

            // Used by get_block_template and getblockchaininfo RPCs.
            ReadRequest::ChainInfo => {
                // # Correctness
                //
                // It is ok to do these lookups using multiple database calls. Finalized state updates
                // can only add overlapping blocks, and block hashes are unique across all chain forks.
                //
                // If there is a large overlap between the non-finalized and finalized states,
                // where the finalized tip is above the non-finalized tip,
                // Zebra is receiving a lot of blocks, or this request has been delayed for a long time.
                //
                // In that case, the `getblocktemplate` RPC will return an error because Zebra
                // is not synced to the tip. That check happens before the RPC makes this request.
                read::difficulty::get_block_template_chain_info(
                    &state.latest_non_finalized_state(),
                    &state.db,
                    &state.network,
                )
                .map(ReadResponse::ChainInfo)
            }

            // Used by getmininginfo, getnetworksolps, and getnetworkhashps RPCs.
            ReadRequest::SolutionRate { num_blocks, height } => {
                let latest_non_finalized_state = state.latest_non_finalized_state();
                // # Correctness
                //
                // It is ok to do these lookups using multiple database calls. Finalized state updates
                // can only add overlapping blocks, and block hashes are unique across all chain forks.
                //
                // The worst that can happen here is that the default `start_hash` will be below
                // the chain tip.
                let (tip_height, tip_hash) =
                    match read::tip(latest_non_finalized_state.best_chain(), &state.db) {
                        Some(tip_hash) => tip_hash,
                        None => return Ok(ReadResponse::SolutionRate(None)),
                    };

                let start_height = height.map_or(tip_height, |height| height.min(tip_height));
                let start_hash = if start_height < tip_height {
                    read::hash_by_height(
                        latest_non_finalized_state.best_chain(),
                        &state.db,
                        start_height,
                    )
                } else {
                    Some(tip_hash)
                };
                let num_blocks = num_blocks.unwrap_or_else(|| {
                    NetworkUpgrade::averaging_window_for_height(&state.network, start_height)
                });

                let solution_rate = start_hash.and_then(|start_hash| {
                    read::difficulty::solution_rate(
                        &latest_non_finalized_state,
                        &state.db,
                        num_blocks,
                        start_hash,
                    )
                });

                Ok(ReadResponse::SolutionRate(solution_rate))
            }

            ReadRequest::CheckBlockProposalValidity(semantically_verified) => {
                tracing::debug!(
                    "attempting to validate and commit block proposal \
                         onto a cloned non-finalized state"
                );
                let mut latest_non_finalized_state = state.latest_non_finalized_state();

                // The previous block of a valid proposal must be on the best chain tip.
                let Some((_best_tip_height, best_tip_hash)) =
                    read::best_tip(&latest_non_finalized_state, &state.db)
                else {
                    return Err(
                        "state is empty: wait for Zebra to sync before submitting a proposal"
                            .into(),
                    );
                };

                if semantically_verified.block.header.previous_block_hash != best_tip_hash {
                    return Err("proposal is not based on the current best chain tip: \
                                    previous block hash must be the best chain tip"
                        .into());
                }

                // This clone of the non-finalized state is dropped when this closure returns.
                // The non-finalized state that's used in the rest of the state (including finalizing
                // blocks into the db) is not mutated here.
                //
                // TODO: Convert `CommitSemanticallyVerifiedError` to a new `ValidateProposalError`?
                latest_non_finalized_state.disable_metrics();

                write::validate_and_commit_non_finalized(
                    &state.db,
                    &mut latest_non_finalized_state,
                    semantically_verified,
                )?;

                Ok(ReadResponse::ValidBlockProposal)
            }

            ReadRequest::TipBlockSize => {
                // Respond with the length of the obtained block if any.
                Ok(ReadResponse::TipBlockSize(
                    state
                        .best_tip()
                        .and_then(|(tip_height, _)| {
                            read::block_info(
                                state.latest_best_chain(),
                                &state.db,
                                tip_height.into(),
                            )
                        })
                        .map(|info| info.size().try_into().expect("u32 should fit in usize"))
                        .or_else(|| {
                            find::tip_block(state.latest_best_chain(), &state.db)
                                .map(|b| b.zcash_serialized_size())
                        }),
                ))
            }

            ReadRequest::NonFinalizedBlocksListener { .. } => {
                unreachable!("should return early");
            }

            // Used by `gettxout` RPC method.
            ReadRequest::IsTransparentOutputSpent(outpoint) => {
                let is_spent = read::unspent_utxo(state.latest_best_chain(), &state.db, outpoint);
                Ok(ReadResponse::IsTransparentOutputSpent(is_spent.is_none()))
            }
        };

        timed_span.spawn_blocking(request_handler)
    }
}

/// Initialize a state service from the provided [`Config`].
/// Returns a boxed state service, a read-only state service,
/// and receivers for state chain tip updates.
///
/// Each `network` has its own separate on-disk database.
///
/// The state uses the `max_checkpoint_height` and `checkpoint_verify_concurrency_limit`
/// to work out when it is near the final checkpoint.
///
/// To share access to the state, wrap the returned service in a `Buffer`,
/// or clone the returned [`ReadStateService`].
///
/// It's possible to construct multiple state services in the same application (as
/// long as they, e.g., use different storage locations), but doing so is
/// probably not what you want.
pub async fn init(
    config: Config,
    network: &Network,
    max_checkpoint_height: block::Height,
    checkpoint_verify_concurrency_limit: usize,
) -> (
    BoxService<Request, Response, BoxError>,
    ReadStateService,
    LatestChainTip,
    ChainTipChange,
) {
    let (state_service, read_only_state_service, latest_chain_tip, chain_tip_change) =
        StateService::new(
            config,
            network,
            max_checkpoint_height,
            checkpoint_verify_concurrency_limit,
        )
        .await;

    (
        BoxService::new(state_service),
        read_only_state_service,
        latest_chain_tip,
        chain_tip_change,
    )
}

/// Initialize a read state service from the provided [`Config`].
/// Returns a read-only state service,
///
/// Each `network` has its own separate on-disk database.
///
/// To share access to the state, clone the returned [`ReadStateService`].
pub fn init_read_only(
    config: Config,
    network: &Network,
) -> Result<
    (
        ReadStateService,
        ZebraDb,
        tokio::sync::watch::Sender<NonFinalizedState>,
    ),
    StateInitError,
> {
    let finalized_state = FinalizedState::new_with_debug(
        &config,
        network,
        true,
        #[cfg(feature = "elasticsearch")]
        false,
        true,
    )?;
    let (non_finalized_state_sender, non_finalized_state_receiver) =
        tokio::sync::watch::channel(NonFinalizedState::new(network));

    Ok((
        ReadStateService::new(
            &finalized_state,
            None,
            WatchReceiver::new(non_finalized_state_receiver),
        ),
        finalized_state.db.clone(),
        non_finalized_state_sender,
    ))
}

/// Calls [`init_read_only`] with the provided [`Config`] and [`Network`] from a blocking task.
///
/// Returns a [`tokio::task::JoinHandle`] whose output is a [`Result`]: awaiting it yields a
/// [`JoinError`](tokio::task::JoinError) if the blocking task panicked or was cancelled, and
/// otherwise an `Err(`[`StateInitError`]`)` if the read-only state could not be opened (for
/// example, a missing read-only database).
pub fn spawn_init_read_only(
    config: Config,
    network: &Network,
) -> tokio::task::JoinHandle<
    Result<
        (
            ReadStateService,
            ZebraDb,
            tokio::sync::watch::Sender<NonFinalizedState>,
        ),
        StateInitError,
    >,
> {
    let network = network.clone();
    tokio::task::spawn_blocking(move || init_read_only(config, &network))
}

/// Returns a [`StateService`] with an ephemeral [`Config`] and a buffer with a single slot.
///
/// This can be used to create a state service for testing. See also [`init`].
#[cfg(any(test, feature = "proptest-impl"))]
pub async fn init_test(
    network: &Network,
) -> Buffer<BoxService<Request, Response, BoxError>, Request> {
    // TODO: pass max_checkpoint_height and checkpoint_verify_concurrency limit
    //       if we ever need to test final checkpoint sent UTXO queries
    let (state_service, _, _, _) =
        StateService::new(Config::ephemeral(), network, block::Height::MAX, 0).await;

    Buffer::new(BoxService::new(state_service), 1)
}

/// Initializes a state service with an ephemeral [`Config`] and a buffer with a single slot,
/// then returns the read-write service, read-only service, and tip watch channels.
///
/// This can be used to create a state service for testing. See also [`init`].
#[cfg(any(test, feature = "proptest-impl"))]
pub async fn init_test_services(
    network: &Network,
) -> (
    Buffer<BoxService<Request, Response, BoxError>, Request>,
    ReadStateService,
    LatestChainTip,
    ChainTipChange,
) {
    // TODO: pass max_checkpoint_height and checkpoint_verify_concurrency limit
    //       if we ever need to test final checkpoint sent UTXO queries
    let (state_service, read_state_service, latest_chain_tip, chain_tip_change) =
        StateService::new(Config::ephemeral(), network, block::Height::MAX, 0).await;

    let state_service = Buffer::new(BoxService::new(state_service), 1);

    (
        state_service,
        read_state_service,
        latest_chain_tip,
        chain_tip_change,
    )
}