persistcli/maelstrom/

txn_list_append_single.rs

// Copyright Materialize, Inc. and contributors. All rights reserved.
//
// Use of this software is governed by the Business Source License
// included in the LICENSE file.
//
// As of the Change Date specified in that file, in accordance with
// the Business Source License, use of this software will be governed
// by the Apache License, Version 2.0.

//! An implementation of the Maelstrom txn-list-append workload using a single
//! persist shard.

use std::collections::BTreeMap;
use std::sync::Arc;
use std::time::{Duration, SystemTime, UNIX_EPOCH};

use async_trait::async_trait;
use differential_dataflow::consolidation::consolidate_updates;
use differential_dataflow::lattice::Lattice;
use mz_dyncfg::ConfigUpdates;
use mz_ore::metrics::MetricsRegistry;
use mz_ore::now::SYSTEM_TIME;
use mz_persist::cfg::{BlobConfig, ConsensusConfig};
use mz_persist::location::{Blob, Consensus, ExternalError};
use mz_persist::unreliable::{UnreliableBlob, UnreliableConsensus, UnreliableHandle};
use mz_persist_client::async_runtime::IsolatedRuntime;
use mz_persist_client::cache::StateCache;
use mz_persist_client::cfg::PersistConfig;
use mz_persist_client::critical::SinceHandle;
use mz_persist_client::metrics::Metrics;
use mz_persist_client::read::{Listen, ListenEvent};
use mz_persist_client::rpc::PubSubClientConnection;
use mz_persist_client::write::WriteHandle;
use mz_persist_client::{Diagnostics, PersistClient, ShardId};
use timely::PartialOrder;
use timely::order::TotalOrder;
use timely::progress::{Antichain, Timestamp};
use tokio::sync::Mutex;
use tracing::{debug, info, trace};

use crate::maelstrom::Args;
use crate::maelstrom::api::{Body, ErrorCode, MaelstromError, NodeId, ReqTxnOp, ResTxnOp};
use crate::maelstrom::node::{Handle, Service};
use crate::maelstrom::services::{CachingBlob, MaelstromBlob, MaelstromConsensus};
use crate::maelstrom::txn_list_append_single::codec_impls::{
    MaelstromKeySchema, MaelstromValSchema,
};

/// Key of the persist shard used by [Transactor]
#[derive(Debug, Clone, Default, PartialEq, Eq, PartialOrd, Ord, Hash)]
pub struct MaelstromKey(u64);

/// Val of the persist shard used by [Transactor]
#[derive(Debug, Clone, Default, PartialEq, Eq, PartialOrd, Ord)]
pub struct MaelstromVal(Vec<u64>);

/// An implementation of read-write transactions on top of persist
///
/// This executes Maelstrom [txn-list-append] transactions. The timestamp
/// selection is modeled after Materialize's SQL implementation.
///
/// [txn-list-append]: https://github.com/jepsen-io/maelstrom/blob/v0.2.1/doc/workloads.md#workload-txn-list-append
///
/// A persist shard is maintained that directly represents a key-value map as it
/// evolves over time. (Our real SQL implementation would likely instead
/// maintain a WAL of updates to multiple tables.) Each transaction has
/// associated timestamps:
///
/// - `read_ts`: at which the contents of the map are read (inclusive)
/// - `write_ts`: at which txn mutations (if any) are written
/// - `expected_upper`: the upper bound of the previous txn
/// - `new_upper`: the upper bound after this transaction commits, if it does
///
/// To keep things simple, `write_ts` is always `read_ts+1`, `expected_upper` is
/// `antichain[write_ts]`, and `new_upper` is `antichain[write_ts+1]`.
///
/// Transactions are "committed" by using `[WriteHandle::compare_and_append]`,
/// which atomically:
/// - verifies that `expected_upper` matches
/// - writes the updates
/// - advances the upper to `new_upper`.
///
/// This guarantees that all transactions are linearized with each txn's
/// `read_ts` equal to the previous txn's `write_ts`. If two transactions race
/// by reading at the same `read_ts` and writing at the same `write_ts`, the
/// `compare_and_append` guarantees that only one of them succeeds and the other
/// must retry with new timestamps.
///
/// Upon first use, the persist shard is initialized by advancing the upper to
/// `antichain[T::minimum() + 1]`. This allows the first txn to use 0 as the
/// `read_ts` and 1 as the `write_ts`.
///
/// Similarly modeling Materialize, the since of the shard is kept following the
/// upper. To keep things simple, this is done by a fixed offset. This exercises
/// persist compaction.
///
/// To ensure that both [mz_persist_client::read::ReadHandle::snapshot] and
/// [mz_persist_client::read::ReadHandle::listen] are exercised, when a txn
/// reads the state at `read_ts`, it artificially picks an `as_of` timestamp in
/// `[since, read_ts]` and splits up the read data between snapshot and listen
/// along this timestamp.
#[derive(Debug)]
pub struct Transactor {
    cads_token: u64,
    shard_id: ShardId,
    client: PersistClient,
    since: SinceHandle<MaelstromKey, MaelstromVal, u64, i64, u64>,
    write: WriteHandle<MaelstromKey, MaelstromVal, u64, i64>,

    read_ts: u64,

    // Keep a long-lived listen, which is incrementally read as we go. Then
    // assert that it has the same data as the short-lived snapshot+listen in
    // `read`. This hopefully stresses slightly different parts of the system.
    long_lived_updates: Vec<((MaelstromKey, MaelstromVal), u64, i64)>,
    long_lived_listen: Listen<MaelstromKey, MaelstromVal, u64, i64>,
}

impl Transactor {
    pub async fn new(
        client: &PersistClient,
        node_id: NodeId,
        shard_id: ShardId,
    ) -> Result<Self, MaelstromError> {
        let cads_token = node_id
            .0
            .trim_start_matches('n')
            .parse::<u64>()
            .expect("maelstrom node_id should be n followed by an integer");

        let (mut write, mut read) = client
            .open(
                shard_id,
                Arc::new(MaelstromKeySchema),
                Arc::new(MaelstromValSchema),
                Diagnostics::from_purpose("maelstrom long-lived"),
                true,
            )
            .await?;
        // Use the CONTROLLER_CRITICAL_SINCE id for all nodes so we get coverage
        // of contending traffic.
        let since = client
            .open_critical_since(
                shard_id,
                PersistClient::CONTROLLER_CRITICAL_SINCE,
                Diagnostics::from_purpose("maelstrom since"),
            )
            .await?;
        let read_ts = Self::maybe_init_shard(&mut write).await?;

        let mut long_lived_updates = Vec::new();
        let as_of = since.since().clone();
        let mut updates = read
            .snapshot_and_fetch(as_of.clone())
            .await
            .expect("as_of unexpectedly unavailable");
        long_lived_updates.append(&mut updates);
        let long_lived_listen = read
            .listen(as_of.clone())
            .await
            .expect("as_of unexpectedly unavailable");

        Ok(Transactor {
            client: client.clone(),
            cads_token,
            shard_id,
            since,
            write,
            read_ts,
            long_lived_updates,
            long_lived_listen,
        })
    }

    /// Initializes the shard, if it hasn't been already, and returns the read
    /// timestamp.
    async fn maybe_init_shard(
        write: &mut WriteHandle<MaelstromKey, MaelstromVal, u64, i64>,
    ) -> Result<u64, MaelstromError> {
        debug!("Transactor::maybe_init");
        const EMPTY_UPDATES: &[((MaelstromKey, MaelstromVal), u64, i64)] = &[];
        let ts_min = u64::minimum();
        let initial_upper = Antichain::from_elem(ts_min);
        let new_upper = Antichain::from_elem(ts_min + 1);
        let cas_res = write
            .compare_and_append(EMPTY_UPDATES, initial_upper.clone(), new_upper.clone())
            .await;
        let read_ts = match cas_res? {
            Ok(()) => 0,
            Err(mismatch) => Self::extract_ts(&mismatch.current)? - 1,
        };
        Ok(read_ts)
    }

    pub async fn transact(
        &mut self,
        req_ops: &[ReqTxnOp],
    ) -> Result<Vec<ResTxnOp>, MaelstromError> {
        loop {
            trace!("transact req={:?}", req_ops);
            let state = self.read().await?;
            debug!("transact req={:?} state={:?}", req_ops, state);
            let (writes, res_ops) = Self::eval_txn(&state, req_ops);

            // NB: We do the CaS even if writes is empty, so that read-only txns
            // are also linearizable.
            let write_ts = self.read_ts + 1;
            let updates = writes
                .into_iter()
                .map(|(k, v, diff)| ((k, v), write_ts, diff));
            let expected_upper = Antichain::from_elem(write_ts);
            let new_upper = Antichain::from_elem(write_ts + 1);
            let cas_res = self
                .write
                .compare_and_append(updates, expected_upper.clone(), new_upper)
                .await?;
            match cas_res {
                Ok(()) => {
                    self.read_ts = write_ts;
                    self.advance_since().await?;
                    return Ok(res_ops);
                }
                // We lost the CaS race, try again.
                Err(mismatch) => {
                    info!(
                        "transact lost the CaS race, retrying: {:?} vs {:?}",
                        mismatch.expected, mismatch.current
                    );
                    self.read_ts = Self::extract_ts(&mismatch.current)? - 1;
                    continue;
                }
            }
        }
    }

    async fn read_short_lived(
        &mut self,
    ) -> Result<
        (
            Vec<((MaelstromKey, MaelstromVal), u64, i64)>,
            Antichain<u64>,
        ),
        MaelstromError,
    > {
        loop {
            // We're reading as of read_ts, but we can split the read between the
            // snapshot and listen at any ts in `[since_ts, read_ts]`. Intentionally
            // pick one that uses a combination of both to get coverage.
            let as_of = Antichain::from_elem(self.read_ts);
            let since_ts = Self::extract_ts(self.since.since())?;
            assert!(self.read_ts >= since_ts, "{} vs {}", self.read_ts, since_ts);
            let snap_ts = since_ts + (self.read_ts - since_ts) / 2;
            let snap_as_of = Antichain::from_elem(snap_ts);

            // Intentionally create this from scratch so we get a brand new copy of
            // state and exercise some more code paths.
            let mut read = self
                .client
                .open_leased_reader(
                    self.shard_id,
                    Arc::new(MaelstromKeySchema),
                    Arc::new(MaelstromValSchema),
                    Diagnostics::from_purpose("maelstrom short-lived"),
                    true,
                )
                .await
                .expect("codecs should match");

            if !PartialOrder::less_equal(read.since(), &snap_as_of) {
                // The desired snapshot as-of is before the actual critical since;
                // refresh our state and try again.
                info!(
                    "read handle since {:?} is not before as_of {}, fetching a new read_ts and trying again",
                    read.since().elements(),
                    snap_ts,
                );
                self.advance_since().await?;

                let recent_upper = self.write.fetch_recent_upper().await;
                self.read_ts = Self::extract_ts(recent_upper)? - 1;
                continue;
            }

            let updates_res = read.snapshot_and_fetch(snap_as_of.clone()).await;
            let mut updates = match updates_res {
                Ok(x) => x,
                Err(since) => {
                    // Because we artificially share the same CriticalReaderId
                    // between nodes, it doesn't quite act like a capability.
                    // Prod doesn't have this issue, because a new one will
                    // fence out old ones, but it's done here to stress edge
                    // cases.
                    //
                    // If we did succeed in this read, we'd anyway just find out
                    // from compare_and_append that our read_ts was out of date,
                    // so proceed by fetching a new one and trying again. We
                    // also re-register the since to get an updated since value
                    // for the snap_ts calculation above.
                    info!(
                        "snapshot cannot serve requested as_of {} since is {:?}, fetching a new read_ts and trying again",
                        snap_ts,
                        since.0.as_option()
                    );
                    self.advance_since().await?;

                    let recent_upper = self.write.fetch_recent_upper().await;
                    self.read_ts = Self::extract_ts(recent_upper)? - 1;
                    continue;
                }
            };

            let listen_res = read.listen(snap_as_of).await;
            let listen = match listen_res {
                Ok(x) => x,
                Err(since) => {
                    // Because we artificially share the same CriticalReaderId
                    // between nodes, it doesn't quite act like a capability.
                    // Prod doesn't have this issue, because a new one will
                    // fence out old ones, but it's done here to stress edge
                    // cases.
                    //
                    // If we did succeed in this read, we'd anyway just find out
                    // from compare_and_append that our read_ts was out of date,
                    // so proceed by fetching a new one and trying again. We
                    // also re-register the since to get an updated since value
                    // for the snap_ts calculation above.
                    info!(
                        "listen cannot serve requested as_of {} since is {:?}, fetching a new read_ts and trying again",
                        snap_ts,
                        since.0.as_option(),
                    );
                    self.advance_since().await?;
                    let recent_upper = self.write.fetch_recent_upper().await;
                    self.read_ts = Self::extract_ts(recent_upper)? - 1;
                    assert!(
                        PartialOrder::less_than(self.since.since(), recent_upper),
                        "invariant: since {:?} should be held behind the recent upper {:?}",
                        &**self.since.since(),
                        &**recent_upper
                    );
                    continue;
                }
            };

            trace!(
                "read updates from snapshot as_of {}: {:?}",
                snap_ts, updates
            );
            let listen_updates = Self::listen_through(listen, &as_of).await?;
            trace!(
                "read updates from listener as_of {} through {}: {:?}",
                snap_ts, self.read_ts, listen_updates
            );
            updates.extend(listen_updates);

            // Compute the contents of the collection as of `as_of`.
            for (_, t, _) in updates.iter_mut() {
                t.advance_by(as_of.borrow());
            }
            consolidate_updates(&mut updates);
            return Ok((updates, as_of));
        }
    }

    async fn read(&mut self) -> Result<BTreeMap<MaelstromKey, MaelstromVal>, MaelstromError> {
        let (updates, as_of) = self.read_short_lived().await?;

        let long_lived = self.read_long_lived(&as_of).await;
        assert_eq!(&updates, &long_lived);

        Self::extract_state_map(self.read_ts, updates)
    }

    async fn listen_through(
        mut listen: Listen<MaelstromKey, MaelstromVal, u64, i64>,
        frontier: &Antichain<u64>,
    ) -> Result<Vec<((MaelstromKey, MaelstromVal), u64, i64)>, ExternalError> {
        let mut ret = Vec::new();
        loop {
            for event in listen.fetch_next().await {
                match event {
                    ListenEvent::Progress(x) => {
                        // NB: Unlike the snapshot as_of, a listener frontier is
                        // not inclusive, so we have to wait until it's > our
                        // as_of to be sure we have everything.
                        if PartialOrder::less_than(frontier, &x) {
                            return Ok(ret);
                        }
                    }
                    ListenEvent::Updates(x) => {
                        // We want the collection at as_of, so skip anything
                        // past that.
                        ret.extend(x.into_iter().filter(|(_, ts, _)| !frontier.less_than(ts)));
                    }
                }
            }
        }
    }

    async fn read_long_lived(
        &mut self,
        as_of: &Antichain<u64>,
    ) -> Vec<((MaelstromKey, MaelstromVal), u64, i64)> {
        while PartialOrder::less_equal(self.long_lived_listen.frontier(), as_of) {
            for event in self.long_lived_listen.fetch_next().await {
                match event {
                    ListenEvent::Updates(mut updates) => {
                        self.long_lived_updates.append(&mut updates)
                    }
                    ListenEvent::Progress(_) => {} // No-op.
                }
            }
        }
        for (_, t, _) in self.long_lived_updates.iter_mut() {
            t.advance_by(as_of.borrow());
        }
        consolidate_updates(&mut self.long_lived_updates);

        // If as_of is less_than the frontier, we may have ended up with updates
        // that we didn't want yet. We can't remove them from
        // `self.long_lived_updates` because the long lived listener will only
        // emit them once and we'll want them later. If performance was
        // important, we could sort them to the end and return a subset, but
        // it's not, so do the easy thing and copy the Vec.
        self.long_lived_updates
            .iter()
            .filter(|(_, t, _)| !as_of.less_than(t))
            .cloned()
            .collect()
    }

    fn extract_state_map(
        read_ts: u64,
        updates: Vec<((MaelstromKey, MaelstromVal), u64, i64)>,
    ) -> Result<BTreeMap<MaelstromKey, MaelstromVal>, MaelstromError> {
        let mut ret = BTreeMap::new();
        for ((k, v), _, d) in updates {
            if d != 1 {
                return Err(MaelstromError {
                    code: ErrorCode::Crash,
                    text: format!("invalid read at time {}", read_ts),
                });
            }
            if ret.contains_key(&k) {
                return Err(MaelstromError {
                    code: ErrorCode::Crash,
                    text: format!("unexpected duplicate key {:?}", k),
                });
            }
            ret.insert(k, v);
        }
        Ok(ret)
    }

    fn eval_txn(
        state: &BTreeMap<MaelstromKey, MaelstromVal>,
        req_ops: &[ReqTxnOp],
    ) -> (Vec<(MaelstromKey, MaelstromVal, i64)>, Vec<ResTxnOp>) {
        let mut res_ops = Vec::new();
        let mut updates = Vec::new();
        let mut txn_state = BTreeMap::new();

        for req_op in req_ops.iter() {
            match req_op {
                ReqTxnOp::Read { key } => {
                    let current = txn_state
                        .get(&MaelstromKey(*key))
                        .or_else(|| state.get(&MaelstromKey(*key)));
                    let val = current.cloned().unwrap_or_default().0;
                    res_ops.push(ResTxnOp::Read { key: *key, val })
                }
                ReqTxnOp::Append { key, val } => {
                    let current = txn_state
                        .get(&MaelstromKey(*key))
                        .or_else(|| state.get(&MaelstromKey(*key)));
                    let mut vals = match current {
                        Some(val) => {
                            // Retract the value we're about to overwrite.
                            updates.push((MaelstromKey(*key), val.clone(), -1));
                            val.clone()
                        }
                        None => MaelstromVal::default(),
                    };
                    vals.0.push(*val);
                    txn_state.insert(MaelstromKey(*key), vals.clone());
                    updates.push((MaelstromKey(*key), vals, 1));
                    res_ops.push(ResTxnOp::Append {
                        key: key.clone(),
                        val: *val,
                    })
                }
            }
        }

        debug!(
            "eval_txn\n  req={:?}\n  res={:?}\n  updates={:?}\n  state={:?}\n  txn_state={:?}",
            req_ops, res_ops, updates, state, txn_state
        );
        (updates, res_ops)
    }

    async fn advance_since(&mut self) -> Result<(), MaelstromError> {
        // To keep things interesting, advance the since.
        const SINCE_LAG: u64 = 10;
        let new_since = Antichain::from_elem(self.read_ts.saturating_sub(SINCE_LAG));

        let mut expected_token = self.cads_token;
        loop {
            let res = self
                .since
                .maybe_compare_and_downgrade_since(&expected_token, (&self.cads_token, &new_since))
                .await;
            match res {
                Some(Ok(latest_since)) => {
                    // Success! Another process might have advanced past our read_ts, so
                    // forward read_ts to since_ts.
                    let since_ts = Self::extract_ts(&latest_since)?;
                    if since_ts > self.read_ts {
                        info!(
                            "since was last updated by {}, forwarding our read_ts from {} to {}",
                            expected_token, self.read_ts, since_ts
                        );
                        self.read_ts = since_ts;
                    }

                    return Ok(());
                }
                Some(Err(actual_token)) => {
                    debug!(
                        "actual downgrade_since token {} didn't match expected {}, retrying",
                        actual_token, expected_token,
                    );
                    expected_token = actual_token;
                }
                None => {
                    panic!("should not no-op `maybe_compare_and_downgrade_since` during testing");
                }
            }
        }
    }

    fn extract_ts<T: TotalOrder + Copy>(frontier: &Antichain<T>) -> Result<T, MaelstromError> {
        frontier.as_option().copied().ok_or_else(|| MaelstromError {
            code: ErrorCode::Crash,
            text: "shard unexpectedly closed".into(),
        })
    }
}

/// An adaptor to implement [Service] using [Transactor]
#[derive(Debug)]
pub struct TransactorService(pub Arc<Mutex<Transactor>>);

#[async_trait]
impl Service for TransactorService {
    async fn init(args: &Args, handle: &Handle) -> Result<Self, MaelstromError> {
        // Use the Maelstrom services to initialize a new random ShardId (so we
        // can repeatedly run tests against the same Blob and Consensus without
        // conflicting) and communicate it between processes.
        let shard_id = handle.maybe_init_shard_id().await?;

        // Make sure the seed is recomputed each time through the retry
        // closure, so we don't retry the same deterministic timeouts.
        let seed: u64 = SystemTime::now()
            .duration_since(UNIX_EPOCH)
            .unwrap_or_default()
            .subsec_nanos()
            .into();
        // It doesn't particularly matter what we set should_happen to, so we do
        // this to have a convenient single tunable param.
        let should_happen = 1.0 - args.unreliability;
        // For consensus, set should_timeout to `args.unreliability` so that once we split
        // ExternalErrors into determinate vs indeterminate, then
        // `args.unreliability` will also be the fraction of txns that it's
        // not save for Maelstrom to retry (b/c indeterminate error in
        // Consensus CaS).
        let should_timeout = args.unreliability;
        // It doesn't particularly matter what we set should_happen and
        // should_timeout to for blobs, so use the same handle for both.
        let unreliable = UnreliableHandle::new(seed, should_happen, should_timeout);

        let mut config =
            PersistConfig::new_default_configs(&mz_persist_client::BUILD_INFO, SYSTEM_TIME.clone());
        {
            // We only use the Postgres tuned queries when connected to vanilla
            // Postgres, so we always want to enable them for testing.
            let mut updates = ConfigUpdates::default();
            updates.add(&mz_persist::postgres::USE_POSTGRES_TUNED_QUERIES, true);
            config.apply_from(&updates);
        }

        let metrics = Arc::new(Metrics::new(&config, &MetricsRegistry::new()));

        // Construct requested Blob.
        let blob = match &args.blob_uri {
            Some(blob_uri) => {
                let cfg = BlobConfig::try_from(
                    blob_uri,
                    Box::new(config.clone()),
                    metrics.s3_blob.clone(),
                    Arc::clone(&config.configs),
                )
                .await
                .expect("blob_uri should be valid");
                loop {
                    match cfg.clone().open().await {
                        Ok(x) => break x,
                        Err(err) => {
                            info!("failed to open blob, trying again: {}", err);
                        }
                    }
                }
            }
            None => MaelstromBlob::new(handle.clone()),
        };
        let blob: Arc<dyn Blob> = Arc::new(UnreliableBlob::new(blob, unreliable.clone()));
        // Normal production persist usage (even including a real SQL txn impl)
        // isn't particularly benefitted by a cache, so we don't have one baked
        // into persist. In contrast, our Maelstrom transaction model
        // intentionally exercises both a new snapshot and new listener on each
        // txn. As a result, without a cache, things would be terribly slow,
        // unreliable would cause more retries than are interesting, and the
        // Lamport diagrams that Maelstrom generates would be noisy.
        let blob = CachingBlob::new(blob);
        // to simplify some downstream logic (+ a bit more stress testing),
        // always downgrade the since of critical handles when asked
        config.critical_downgrade_interval = Duration::from_secs(0);
        // set a live diff scan limit such that we'll explore both the fast and slow paths
        config.set_state_versions_recent_live_diffs_limit(5);
        let consensus = match &args.consensus_uri {
            Some(consensus_uri) => {
                let cfg = ConsensusConfig::try_from(
                    consensus_uri,
                    Box::new(config.clone()),
                    metrics.postgres_consensus.clone(),
                    Arc::clone(&config.configs),
                )
                .expect("consensus_uri should be valid");
                loop {
                    match cfg.clone().open().await {
                        Ok(x) => break x,
                        Err(err) => {
                            info!("failed to open consensus, trying again: {}", err);
                        }
                    }
                }
            }
            None => MaelstromConsensus::new(handle.clone()),
        };
        let consensus: Arc<dyn Consensus> =
            Arc::new(UnreliableConsensus::new(consensus, unreliable));

        // Wire up the TransactorService.
        let isolated_runtime = Arc::new(IsolatedRuntime::new_for_tests());
        let pubsub_sender = PubSubClientConnection::noop().sender;
        let shared_states = Arc::new(StateCache::new(
            &config,
            Arc::clone(&metrics),
            Arc::clone(&pubsub_sender),
        ));
        let client = PersistClient::new(
            config,
            blob,
            consensus,
            metrics,
            isolated_runtime,
            shared_states,
            pubsub_sender,
        )?;
        let transactor = Transactor::new(&client, handle.node_id(), shard_id).await?;
        let service = TransactorService(Arc::new(Mutex::new(transactor)));
        Ok(service)
    }

    async fn eval(&self, handle: Handle, src: NodeId, req: Body) {
        match req {
            Body::ReqTxn { msg_id, txn } => {
                let in_reply_to = msg_id;
                match self.0.lock().await.transact(&txn).await {
                    Ok(txn) => handle.send_res(src, |msg_id| Body::ResTxn {
                        msg_id,
                        in_reply_to,
                        txn,
                    }),
                    Err(MaelstromError { code, text }) => {
                        handle.send_res(src, |msg_id| Body::Error {
                            msg_id: Some(msg_id),
                            in_reply_to,
                            code,
                            text,
                        })
                    }
                }
            }
            req => unimplemented!("unsupported req: {:?}", req),
        }
    }
}

mod codec_impls {
    use arrow::array::{BinaryArray, BinaryBuilder, UInt64Array, UInt64Builder};
    use arrow::datatypes::ToByteSlice;
    use bytes::Bytes;
    use mz_persist_types::Codec;
    use mz_persist_types::codec_impls::{
        SimpleColumnarData, SimpleColumnarDecoder, SimpleColumnarEncoder,
    };
    use mz_persist_types::columnar::Schema;
    use mz_persist_types::stats::NoneStats;

    use crate::maelstrom::txn_list_append_single::{MaelstromKey, MaelstromVal};

    impl Codec for MaelstromKey {
        type Storage = ();
        type Schema = MaelstromKeySchema;

        fn codec_name() -> String {
            "MaelstromKey".into()
        }

        fn encode<B>(&self, buf: &mut B)
        where
            B: bytes::BufMut,
        {
            let bytes = serde_json::to_vec(&self.0).expect("failed to encode key");
            buf.put(bytes.as_slice());
        }

        fn decode<'a>(buf: &'a [u8], _schema: &MaelstromKeySchema) -> Result<Self, String> {
            Ok(MaelstromKey(
                serde_json::from_slice(buf).map_err(|err| err.to_string())?,
            ))
        }

        fn encode_schema(_schema: &Self::Schema) -> Bytes {
            Bytes::new()
        }

        fn decode_schema(buf: &Bytes) -> Self::Schema {
            assert_eq!(*buf, Bytes::new());
            MaelstromKeySchema
        }
    }

    impl SimpleColumnarData for MaelstromKey {
        type ArrowBuilder = UInt64Builder;
        type ArrowColumn = UInt64Array;

        fn goodbytes(builder: &Self::ArrowBuilder) -> usize {
            builder.values_slice().to_byte_slice().len()
        }

        fn push(&self, builder: &mut Self::ArrowBuilder) {
            builder.append_value(self.0);
        }
        fn push_null(builder: &mut Self::ArrowBuilder) {
            builder.append_null();
        }

        fn read(&mut self, idx: usize, column: &Self::ArrowColumn) {
            *self = MaelstromKey(column.value(idx));
        }
    }

    #[derive(Debug, PartialEq)]
    pub struct MaelstromKeySchema;

    impl Schema<MaelstromKey> for MaelstromKeySchema {
        type ArrowColumn = UInt64Array;
        type Statistics = NoneStats;

        type Decoder = SimpleColumnarDecoder<MaelstromKey>;
        type Encoder = SimpleColumnarEncoder<MaelstromKey>;

        fn encoder(&self) -> Result<Self::Encoder, anyhow::Error> {
            Ok(SimpleColumnarEncoder::default())
        }

        fn decoder(&self, col: Self::ArrowColumn) -> Result<Self::Decoder, anyhow::Error> {
            Ok(SimpleColumnarDecoder::new(col))
        }
    }

    impl Codec for MaelstromVal {
        type Storage = ();
        type Schema = MaelstromValSchema;

        fn codec_name() -> String {
            "MaelstromVal".into()
        }

        fn encode<B>(&self, buf: &mut B)
        where
            B: bytes::BufMut,
        {
            let bytes = serde_json::to_vec(&self.0).expect("failed to encode val");
            buf.put(bytes.as_slice());
        }

        fn decode<'a>(buf: &'a [u8], _schema: &MaelstromValSchema) -> Result<Self, String> {
            Ok(MaelstromVal(
                serde_json::from_slice(buf).map_err(|err| err.to_string())?,
            ))
        }

        fn encode_schema(_schema: &Self::Schema) -> Bytes {
            Bytes::new()
        }

        fn decode_schema(buf: &Bytes) -> Self::Schema {
            assert_eq!(*buf, Bytes::new());
            MaelstromValSchema
        }
    }

    impl SimpleColumnarData for MaelstromVal {
        type ArrowBuilder = BinaryBuilder;
        type ArrowColumn = BinaryArray;

        fn goodbytes(builder: &Self::ArrowBuilder) -> usize {
            builder.values_slice().to_byte_slice().len()
        }

        fn push(&self, builder: &mut Self::ArrowBuilder) {
            builder.append_value(&self.encode_to_vec());
        }
        fn push_null(builder: &mut Self::ArrowBuilder) {
            builder.append_null()
        }

        fn read(&mut self, idx: usize, column: &Self::ArrowColumn) {
            *self = MaelstromVal::decode(column.value(idx), &MaelstromValSchema)
                .expect("should be valid MaelstromVal");
        }
    }

    #[derive(Debug, PartialEq)]
    pub struct MaelstromValSchema;

    impl Schema<MaelstromVal> for MaelstromValSchema {
        type ArrowColumn = BinaryArray;
        type Statistics = NoneStats;

        type Decoder = SimpleColumnarDecoder<MaelstromVal>;
        type Encoder = SimpleColumnarEncoder<MaelstromVal>;

        fn encoder(&self) -> Result<Self::Encoder, anyhow::Error> {
            Ok(SimpleColumnarEncoder::default())
        }

        fn decoder(&self, col: Self::ArrowColumn) -> Result<Self::Decoder, anyhow::Error> {
            Ok(SimpleColumnarDecoder::new(col))
        }
    }
}