sui_indexer_alt_framework/pipeline/concurrent/

mod.rs

// Copyright (c) Mysten Labs, Inc.
// SPDX-License-Identifier: Apache-2.0

use std::sync::Arc;
use std::time::Duration;

use async_trait::async_trait;
use serde::Deserialize;
use serde::Serialize;
use sui_futures::service::Service;
use tokio::sync::SetOnce;
use tokio::sync::mpsc;
use tracing::info;

use crate::Task;
use crate::config::ConcurrencyConfig;
use crate::ingestion::ingestion_client::CheckpointEnvelope;
use crate::metrics::IndexerMetrics;
use crate::pipeline::CommitterConfig;
use crate::pipeline::Processor;
use crate::pipeline::WatermarkPart;
use crate::pipeline::concurrent::collector::collector;
use crate::pipeline::concurrent::commit_watermark::commit_watermark;
use crate::pipeline::concurrent::committer::committer;
use crate::pipeline::concurrent::main_reader_lo::track_main_reader_lo;
use crate::pipeline::concurrent::pruner::pruner;
use crate::pipeline::concurrent::reader_watermark::reader_watermark;
use crate::pipeline::processor::processor;
use crate::store::ConcurrentStore;
use crate::store::Store;

mod collector;
mod commit_watermark;
mod committer;
mod main_reader_lo;
mod pruner;
mod reader_watermark;

/// Status returned by `Handler::batch` to indicate whether the batch is ready to be committed.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum BatchStatus {
    /// The batch can accept more values.
    Pending,
    /// The batch is full and should be committed.
    Ready,
}

/// Handlers implement the logic for a given indexing pipeline: How to process checkpoint data (by
/// implementing [Processor]) into rows for their table, and how to write those rows to the database.
///
/// The handler is also responsible for tuning the various parameters of the pipeline (provided as
/// associated values). Reasonable defaults have been chosen to balance concurrency with memory
/// usage, but each handle may choose to override these defaults, e.g.
///
/// - Handlers that produce many small rows may wish to increase their batch/chunk/max-pending
///   sizes).
/// - Handlers that do more work during processing may wish to increase their fanout so more of it
///   can be done concurrently, to preserve throughput.
///
/// Concurrent handlers can only be used in concurrent pipelines, where checkpoint data is
/// processed and committed out-of-order and a watermark table is kept up-to-date with the latest
/// checkpoint below which all data has been committed.
///
/// Back-pressure is handled through the `MAX_PENDING_SIZE` constant -- if more than this many rows
/// build up, the collector will stop accepting new checkpoints, which will eventually propagate
/// back to the ingestion service.
#[async_trait]
pub trait Handler: Processor {
    type Store: ConcurrentStore;
    type Batch: Default + Send + Sync + 'static;

    /// If at least this many rows are pending, the committer will commit them eagerly.
    const MIN_EAGER_ROWS: usize = 50;

    /// If there are more than this many rows pending, the committer applies backpressure.
    const MAX_PENDING_ROWS: usize = 5000;

    /// The maximum number of watermarks that can show up in a single batch.
    /// This limit exists to deal with pipelines that produce no data for a majority of
    /// checkpoints -- the size of these pipeline's batches will be dominated by watermark updates.
    const MAX_WATERMARK_UPDATES: usize = 10_000;

    /// Add values from the iterator to the batch. The implementation may take all, some, or none
    /// of the values from the iterator by calling `.next()`.
    ///
    /// Returns `BatchStatus::Ready` if the batch is full and should be committed,
    /// or `BatchStatus::Pending` if the batch can accept more values.
    ///
    /// Note: The handler can signal batch readiness via `BatchStatus::Ready`, but the framework
    /// may also decide to commit a batch based on the trait parameters above.
    fn batch(
        &self,
        batch: &mut Self::Batch,
        values: &mut std::vec::IntoIter<Self::Value>,
    ) -> BatchStatus;

    /// Commit the batch to the database, returning the number of rows affected.
    async fn commit<'a>(
        &self,
        batch: &Self::Batch,
        conn: &mut <Self::Store as Store>::Connection<'a>,
    ) -> anyhow::Result<usize>;

    /// Clean up data between checkpoints `_from` and `_to_exclusive` (exclusive) in the database, returning
    /// the number of rows affected. This function is optional, and defaults to not pruning at all.
    async fn prune<'a>(
        &self,
        _from: u64,
        _to_exclusive: u64,
        _conn: &mut <Self::Store as Store>::Connection<'a>,
    ) -> anyhow::Result<usize> {
        Ok(0)
    }
}

/// Configuration for a concurrent pipeline
#[derive(Serialize, Deserialize, Debug, Clone, Default)]
pub struct ConcurrentConfig {
    /// Configuration for the writer, that makes forward progress.
    pub committer: CommitterConfig,

    /// Configuration for the pruner, that deletes old data.
    pub pruner: Option<PrunerConfig>,

    /// Processor concurrency. Defaults to adaptive scaling up to the number of CPUs.
    pub fanout: Option<ConcurrencyConfig>,

    /// Override for `Handler::MIN_EAGER_ROWS` (eager batch threshold).
    pub min_eager_rows: Option<usize>,

    /// Override for `Handler::MAX_PENDING_ROWS` (backpressure threshold).
    pub max_pending_rows: Option<usize>,

    /// Override for `Handler::MAX_WATERMARK_UPDATES` (watermarks per batch cap).
    pub max_watermark_updates: Option<usize>,

    /// Size of the channel between the processor and collector.
    pub processor_channel_size: Option<usize>,

    /// Size of the channel between the collector and committer.
    pub collector_channel_size: Option<usize>,

    /// Size of the channel between the committer and the watermark updater.
    pub committer_channel_size: Option<usize>,
}

#[derive(Serialize, Deserialize, Debug, Clone)]
pub struct PrunerConfig {
    /// How often the pruner should check whether there is any data to prune, in milliseconds.
    pub interval_ms: u64,

    /// How long to wait after the reader low watermark was set, until it is safe to prune up until
    /// this new watermark, in milliseconds.
    pub delay_ms: u64,

    /// How much data to keep, this is measured in checkpoints.
    pub retention: u64,

    /// The maximum range to try and prune in one request, measured in checkpoints.
    pub max_chunk_size: u64,

    /// The max number of tasks to run in parallel for pruning.
    pub prune_concurrency: u64,
}

/// Values ready to be written to the database. This is an internal type used to communicate
/// between the collector and the committer parts of the pipeline.
///
/// Values inside each batch may or may not be from the same checkpoint. Values in the same
/// checkpoint can also be split across multiple batches.
struct BatchedRows<H: Handler> {
    /// The batch to write
    batch: H::Batch,
    /// Number of rows in the batch
    batch_len: usize,
    /// Proportions of all the watermarks that are represented in this chunk
    watermark: Vec<WatermarkPart>,
}

impl<H, V> BatchedRows<H>
where
    H: Handler<Batch = Vec<V>, Value = V>,
{
    #[cfg(test)]
    pub fn from_vec(batch: Vec<V>, watermark: Vec<WatermarkPart>) -> Self {
        let batch_len = batch.len();
        Self {
            batch,
            batch_len,
            watermark,
        }
    }
}

impl PrunerConfig {
    pub fn interval(&self) -> Duration {
        Duration::from_millis(self.interval_ms)
    }

    pub fn delay(&self) -> Duration {
        Duration::from_millis(self.delay_ms)
    }
}

impl Default for PrunerConfig {
    fn default() -> Self {
        Self {
            interval_ms: 300_000,
            delay_ms: 120_000,
            retention: 4_000_000,
            max_chunk_size: 2_000,
            prune_concurrency: 1,
        }
    }
}

/// Start a new concurrent (out-of-order) indexing pipeline served by the handler, `H`. Starting
/// strictly after the `watermark` (or from the beginning if no watermark was provided).
///
/// Each pipeline consists of a processor task which takes checkpoint data and breaks it down into
/// rows, ready for insertion, a collector which batches those rows into an appropriate size for
/// the database, a committer which writes the rows out concurrently, and a watermark task to
/// update the high watermark.
///
/// Committing is performed out-of-order: the pipeline may write out checkpoints out-of-order,
/// either because it received the checkpoints out-of-order or because of variance in processing
/// time.
///
/// The pipeline also maintains a row in the `watermarks` table for the pipeline which tracks the
/// watermark below which all data has been committed (modulo pruning).
///
/// Checkpoint data is fed into the pipeline through the `checkpoint_rx` channel, and internal
/// channels are created to communicate between its various components. The pipeline will shutdown
/// if any of its input or output channels close, any of its independent tasks fail, or if it is
/// signalled to shutdown through the returned service handle.
pub(crate) fn pipeline<H: Handler>(
    handler: H,
    next_checkpoint: u64,
    config: ConcurrentConfig,
    store: H::Store,
    task: Option<Task>,
    checkpoint_rx: mpsc::Receiver<Arc<CheckpointEnvelope>>,
    metrics: Arc<IndexerMetrics>,
) -> Service {
    info!(
        pipeline = H::NAME,
        "Starting pipeline with config: {config:#?}",
    );

    let ConcurrentConfig {
        committer: committer_config,
        pruner: pruner_config,
        fanout,
        min_eager_rows,
        max_pending_rows,
        max_watermark_updates,
        processor_channel_size,
        collector_channel_size,
        committer_channel_size,
    } = config;

    let concurrency = fanout.unwrap_or(ConcurrencyConfig::Adaptive {
        initial: 1,
        min: 1,
        max: num_cpus::get(),
        dead_band: None,
    });
    let min_eager_rows = min_eager_rows.unwrap_or(H::MIN_EAGER_ROWS);
    let max_pending_rows = max_pending_rows.unwrap_or(H::MAX_PENDING_ROWS);
    let max_watermark_updates = max_watermark_updates.unwrap_or(H::MAX_WATERMARK_UPDATES);

    let processor_channel_size = processor_channel_size.unwrap_or(num_cpus::get() / 2);
    let (processor_tx, collector_rx) = mpsc::channel(processor_channel_size);

    let collector_channel_size = collector_channel_size.unwrap_or(num_cpus::get() / 2);
    //docs::#buff (see docs/content/guides/developer/advanced/custom-indexer.mdx)
    let (collector_tx, committer_rx) = mpsc::channel(collector_channel_size);
    //docs::/#buff
    let committer_channel_size = committer_channel_size.unwrap_or(num_cpus::get());
    let (committer_tx, watermark_rx) = mpsc::channel(committer_channel_size);
    let main_reader_lo = Arc::new(SetOnce::new());

    let handler = Arc::new(handler);

    let s_processor = processor(
        handler.clone(),
        checkpoint_rx,
        processor_tx,
        metrics.clone(),
        concurrency,
        store.clone(),
    );

    let s_collector = collector::<H>(
        handler.clone(),
        committer_config.clone(),
        collector_rx,
        collector_tx,
        main_reader_lo.clone(),
        metrics.clone(),
        min_eager_rows,
        max_pending_rows,
        max_watermark_updates,
    );

    let s_committer = committer::<H>(
        handler.clone(),
        committer_config.clone(),
        committer_rx,
        committer_tx,
        store.clone(),
        metrics.clone(),
    );

    let s_commit_watermark = commit_watermark::<H>(
        next_checkpoint,
        committer_config,
        watermark_rx,
        store.clone(),
        task.as_ref().map(|t| t.task.clone()),
        metrics.clone(),
    );

    let s_track_reader_lo = track_main_reader_lo::<H>(
        main_reader_lo.clone(),
        task.as_ref().map(|t| t.reader_interval),
        store.clone(),
    );

    let s_reader_watermark =
        reader_watermark::<H>(pruner_config.clone(), store.clone(), metrics.clone());

    let s_pruner = pruner(handler, pruner_config, store, metrics);

    s_processor
        .merge(s_collector)
        .merge(s_committer)
        .merge(s_commit_watermark)
        .attach(s_track_reader_lo)
        .attach(s_reader_watermark)
        .attach(s_pruner)
}

#[cfg(test)]
mod tests {
    use std::sync::Arc;
    use std::time::Duration;

    use prometheus::Registry;
    use sui_types::digests::CheckpointDigest;
    use tokio::sync::mpsc;
    use tokio::time::timeout;

    use crate::FieldCount;
    use crate::metrics::IndexerMetrics;
    use crate::mocks::store::MockConnection;
    use crate::mocks::store::MockStore;
    use crate::pipeline::Processor;
    use crate::types::full_checkpoint_content::Checkpoint;
    use crate::types::test_checkpoint_data_builder::TestCheckpointBuilder;

    use super::*;

    const TEST_TIMEOUT: Duration = Duration::from_secs(60);
    const TEST_CHECKPOINT_BUFFER_SIZE: usize = 3; // Critical for back-pressure testing calculations

    #[derive(Clone, Debug, FieldCount)]
    struct TestValue {
        checkpoint: u64,
        data: u64,
    }

    struct DataPipeline;

    #[async_trait]
    impl Processor for DataPipeline {
        const NAME: &'static str = "test_handler";
        type Value = TestValue;

        async fn process(&self, checkpoint: &Arc<Checkpoint>) -> anyhow::Result<Vec<Self::Value>> {
            let cp_num = checkpoint.summary.sequence_number;

            // Every checkpoint will come with 2 processed values
            Ok(vec![
                TestValue {
                    checkpoint: cp_num,
                    data: cp_num * 10 + 1,
                },
                TestValue {
                    checkpoint: cp_num,
                    data: cp_num * 10 + 2,
                },
            ])
        }
    }

    #[async_trait]
    impl Handler for DataPipeline {
        type Store = MockStore;
        type Batch = Vec<TestValue>;

        const MIN_EAGER_ROWS: usize = 1000; // High value to disable eager batching
        const MAX_PENDING_ROWS: usize = 4; // Small value to trigger back pressure quickly
        const MAX_WATERMARK_UPDATES: usize = 1; // Each batch will have 1 checkpoint for an ease of testing.

        fn batch(
            &self,
            batch: &mut Self::Batch,
            values: &mut std::vec::IntoIter<Self::Value>,
        ) -> BatchStatus {
            // Take all values
            batch.extend(values);
            BatchStatus::Pending
        }

        async fn commit<'a>(
            &self,
            batch: &Self::Batch,
            conn: &mut MockConnection<'a>,
        ) -> anyhow::Result<usize> {
            // Group values by checkpoint
            let mut grouped: std::collections::HashMap<u64, Vec<u64>> =
                std::collections::HashMap::new();
            for value in batch {
                grouped
                    .entry(value.checkpoint)
                    .or_default()
                    .push(value.data);
            }

            // Commit all data at once
            conn.0.commit_bulk_data(DataPipeline::NAME, grouped).await
        }

        async fn prune<'a>(
            &self,
            from: u64,
            to_exclusive: u64,
            conn: &mut MockConnection<'a>,
        ) -> anyhow::Result<usize> {
            conn.0.prune_data(DataPipeline::NAME, from, to_exclusive)
        }
    }

    struct TestSetup {
        store: MockStore,
        checkpoint_tx: mpsc::Sender<Arc<CheckpointEnvelope>>,
        #[allow(unused)]
        pipeline: Service,
    }

    impl TestSetup {
        async fn new(config: ConcurrentConfig, store: MockStore, next_checkpoint: u64) -> Self {
            let (checkpoint_tx, checkpoint_rx) = mpsc::channel(TEST_CHECKPOINT_BUFFER_SIZE);
            let metrics = IndexerMetrics::new(None, &Registry::default());

            let pipeline = pipeline(
                DataPipeline,
                next_checkpoint,
                config,
                store.clone(),
                None,
                checkpoint_rx,
                metrics,
            );

            Self {
                store,
                checkpoint_tx,
                pipeline,
            }
        }

        async fn send_checkpoint(&self, checkpoint: u64) -> anyhow::Result<()> {
            let checkpoint_envelope = Arc::new(CheckpointEnvelope {
                checkpoint: Arc::new(
                    TestCheckpointBuilder::new(checkpoint)
                        .with_epoch(1)
                        .with_network_total_transactions(checkpoint * 2)
                        .with_timestamp_ms(1000000000 + checkpoint * 1000)
                        .build_checkpoint(),
                ),
                chain_id: CheckpointDigest::new([1; 32]).into(),
            });
            self.checkpoint_tx.send(checkpoint_envelope).await?;
            Ok(())
        }

        async fn send_checkpoint_with_timeout(
            &self,
            checkpoint: u64,
            timeout_duration: Duration,
        ) -> anyhow::Result<()> {
            timeout(timeout_duration, self.send_checkpoint(checkpoint)).await?
        }

        async fn send_checkpoint_expect_timeout(
            &self,
            checkpoint: u64,
            timeout_duration: Duration,
        ) {
            timeout(timeout_duration, self.send_checkpoint(checkpoint))
                .await
                .unwrap_err(); // Panics if send succeeds
        }
    }

    #[tokio::test]
    async fn test_e2e_pipeline() {
        let config = ConcurrentConfig {
            pruner: Some(PrunerConfig {
                interval_ms: 5_000, // Long interval to test states before pruning
                delay_ms: 100,      // Short delay for faster tests
                retention: 3,       // Keep only 3 checkpoints
                ..Default::default()
            }),
            ..Default::default()
        };
        let store = MockStore::default();
        let setup = TestSetup::new(config, store, 0).await;

        // Send initial checkpoints
        for i in 0..3 {
            setup
                .send_checkpoint_with_timeout(i, Duration::from_millis(200))
                .await
                .unwrap();
        }

        // Verify all initial data is available (before any pruning)
        for i in 0..3 {
            let data = setup
                .store
                .wait_for_data(DataPipeline::NAME, i, TEST_TIMEOUT)
                .await;
            assert_eq!(data, vec![i * 10 + 1, i * 10 + 2]);
        }

        // Add more checkpoints to trigger pruning
        for i in 3..6 {
            setup
                .send_checkpoint_with_timeout(i, Duration::from_millis(200))
                .await
                .unwrap();
        }

        // Verify data is still available BEFORE pruning kicks in
        // With long pruning interval (5s), we can safely verify data without race conditions
        for i in 0..6 {
            let data = setup
                .store
                .wait_for_data(DataPipeline::NAME, i, Duration::from_secs(1))
                .await;
            assert_eq!(data, vec![i * 10 + 1, i * 10 + 2]);
        }

        // Wait for pruning to occur. The pruner and reader_watermark tasks both run on
        // the same interval, so poll until the pruner has caught up rather instead of using a
        // fixed sleep.
        let pruning_deadline = Duration::from_secs(15);
        let start = tokio::time::Instant::now();
        loop {
            let pruned = {
                let data = setup.store.data.get(DataPipeline::NAME).unwrap();
                !data.contains_key(&0) && !data.contains_key(&1) && !data.contains_key(&2)
            };
            if pruned {
                break;
            }
            assert!(
                start.elapsed() < pruning_deadline,
                "Timed out waiting for pruning to occur"
            );
            tokio::time::sleep(Duration::from_millis(100)).await;
        }

        // Verify recent checkpoints are still available
        {
            let data = setup.store.data.get(DataPipeline::NAME).unwrap();
            assert!(data.contains_key(&3));
            assert!(data.contains_key(&4));
            assert!(data.contains_key(&5));
        };
    }

    #[tokio::test]
    async fn test_e2e_pipeline_without_pruning() {
        let config = ConcurrentConfig {
            pruner: None,
            ..Default::default()
        };
        let store = MockStore::default();
        let setup = TestSetup::new(config, store, 0).await;

        // Send several checkpoints
        for i in 0..10 {
            setup
                .send_checkpoint_with_timeout(i, Duration::from_millis(200))
                .await
                .unwrap();
        }

        // Wait for all data to be processed and committed
        let watermark = setup
            .store
            .wait_for_watermark(DataPipeline::NAME, 9, TEST_TIMEOUT)
            .await;

        // Verify ALL data was processed correctly (no pruning)
        for i in 0..10 {
            let data = setup
                .store
                .wait_for_data(DataPipeline::NAME, i, Duration::from_secs(1))
                .await;
            assert_eq!(data, vec![i * 10 + 1, i * 10 + 2]);
        }

        // Verify watermark progression
        assert_eq!(watermark.checkpoint_hi_inclusive, Some(9));
        assert_eq!(watermark.tx_hi, 18); // 9 * 2
        assert_eq!(watermark.timestamp_ms_hi_inclusive, 1000009000); // 1000000000 + 9 * 1000

        // Verify no data was pruned - all 10 checkpoints should still exist
        let total_checkpoints = {
            let data = setup.store.data.get(DataPipeline::NAME).unwrap();
            data.len()
        };
        assert_eq!(total_checkpoints, 10);
    }

    #[tokio::test]
    async fn test_out_of_order_processing() {
        let config = ConcurrentConfig::default();
        let store = MockStore::default();
        let setup = TestSetup::new(config, store, 0).await;

        // Send checkpoints out of order
        let checkpoints = vec![2, 0, 4, 1, 3];
        for cp in checkpoints {
            setup
                .send_checkpoint_with_timeout(cp, Duration::from_millis(200))
                .await
                .unwrap();
        }

        // Wait for all data to be processed
        setup
            .store
            .wait_for_watermark(DataPipeline::NAME, 4, Duration::from_secs(5))
            .await;

        // Verify all checkpoints were processed correctly despite out-of-order arrival
        for i in 0..5 {
            let data = setup
                .store
                .wait_for_data(DataPipeline::NAME, i, Duration::from_secs(1))
                .await;
            assert_eq!(data, vec![i * 10 + 1, i * 10 + 2]);
        }
    }

    #[tokio::test]
    async fn test_watermark_progression_with_gaps() {
        let config = ConcurrentConfig::default();
        let store = MockStore::default();
        let setup = TestSetup::new(config, store, 0).await;

        // Send checkpoints with a gap (0, 1, 3, 4) - missing checkpoint 2
        for cp in [0, 1, 3, 4] {
            setup
                .send_checkpoint_with_timeout(cp, Duration::from_millis(200))
                .await
                .unwrap();
        }

        // Wait for processing
        tokio::time::sleep(Duration::from_secs(1)).await;

        // Watermark should only progress to 1 (can't progress past the gap)
        let watermark = setup.store.watermark(DataPipeline::NAME).unwrap();
        assert_eq!(watermark.checkpoint_hi_inclusive, Some(1));

        // Now send the missing checkpoint 2
        setup
            .send_checkpoint_with_timeout(2, Duration::from_millis(200))
            .await
            .unwrap();

        // Now watermark should progress to 4
        let watermark = setup
            .store
            .wait_for_watermark(DataPipeline::NAME, 4, TEST_TIMEOUT)
            .await;
        assert_eq!(watermark.checkpoint_hi_inclusive, Some(4));
    }

    // ==================== BACK-PRESSURE TESTING ====================

    #[tokio::test]
    async fn test_back_pressure_collector_max_pending_rows() {
        // Pipeline Diagram - Collector Back Pressure via MAX_PENDING_ROWS:
        //
        // ┌────────────┐    ┌────────────┐    ┌────────────┐    ┌────────────┐
        // │ Checkpoint │ ─► │ Processor  │ ─► │ Collector  │ ─► │ Committer  │
        // │   Input    │    │ (fanout=2) │    │            │    │            │
        // └────────────┘    └────────────┘    └[BOTTLENECK]┘    └────────────┘
        //                │                 │                 │
        //              [●●●]           [●●●●●●●]         [●●●●●●]
        //            buffer: 3         buffer: 7         buffer: 6
        //
        // BOTTLENECK: Collector stops accepting when pending rows ≥ MAX_PENDING_ROWS (4)

        let config = ConcurrentConfig {
            committer: CommitterConfig {
                collect_interval_ms: 5_000, // Long interval to prevent timer-driven collection
                write_concurrency: 1,
                ..Default::default()
            },
            fanout: Some(ConcurrencyConfig::Fixed { value: 2 }),
            processor_channel_size: Some(7),
            collector_channel_size: Some(6),
            ..Default::default()
        };
        let store = MockStore::default();
        let setup = TestSetup::new(config, store, 0).await;

        // Wait for initial setup
        tokio::time::sleep(Duration::from_millis(200)).await;

        // Pipeline capacity analysis with collector back pressure:
        // Configuration: MAX_PENDING_ROWS=4, fanout=2
        //
        // Channel and task breakdown:
        // - Checkpoint->Processor channel: 3 slots (TEST_CHECKPOINT_BUFFER_SIZE)
        // - Processor tasks: 2 tasks (fanout=2)
        // - Processor->Collector channel: 7 slots (processor_channel_size=7)
        // - Collector pending: 2 checkpoints × 2 values = 4 values (hits MAX_PENDING_ROWS=4)
        //
        // Total capacity: 3 + 2 + 7 + 2 = 14 checkpoints

        // Fill pipeline to capacity - these should all succeed
        for i in 0..14 {
            setup
                .send_checkpoint_with_timeout(i, Duration::from_millis(200))
                .await
                .unwrap();
        }

        // Checkpoint 14 should block due to MAX_PENDING_ROWS back pressure
        setup
            .send_checkpoint_expect_timeout(14, Duration::from_millis(200))
            .await;

        // Allow pipeline to drain by sending the blocked checkpoint with longer timeout
        setup
            .send_checkpoint_with_timeout(14, TEST_TIMEOUT)
            .await
            .unwrap();

        // Verify data was processed correctly
        let data = setup
            .store
            .wait_for_data(DataPipeline::NAME, 0, TEST_TIMEOUT)
            .await;
        assert_eq!(data, vec![1, 2]);
    }

    #[tokio::test]
    async fn test_back_pressure_committer_slow_commits() {
        // Pipeline Diagram - Committer Back Pressure via Slow Database Commits:
        //
        // ┌────────────┐    ┌────────────┐    ┌────────────┐    ┌────────────┐
        // │ Checkpoint │ ─► │ Processor  │ ─► │ Collector  │ ─► │ Committer  │
        // │   Input    │    │ (fanout=2) │    │            │    │🐌 10s Delay│
        // └────────────┘    └────────────┘    └────────────┘    └[BOTTLENECK]┘
        //                │                 │                 │
        //              [●●●]           [●●●●●●●]          [●●●●●●]
        //            buffer: 3    proc_chan: 7       coll_chan: 6
        //
        // BOTTLENECK: Committer with 10s delay blocks entire pipeline

        let config = ConcurrentConfig {
            committer: CommitterConfig {
                write_concurrency: 1, // Single committer for deterministic blocking
                // MIN_EAGER_ROWS is 1000 and MAX_PENDING_ROWS is 4, so
                // this test relies on the collect interval tip to force
                // batch flushing.
                collect_interval_ms: 10,
                ..Default::default()
            },
            fanout: Some(ConcurrencyConfig::Fixed { value: 2 }),
            processor_channel_size: Some(7),
            collector_channel_size: Some(6),
            ..Default::default()
        };
        let store = MockStore::default().with_commit_delay(10_000); // 10 seconds delay
        let setup = TestSetup::new(config, store, 0).await;

        // Pipeline capacity analysis with slow commits:
        // Configuration: fanout=2, write_concurrency=1
        //
        // Channel and task breakdown:
        // - Checkpoint->Processor channel: 3 slots (TEST_CHECKPOINT_BUFFER_SIZE)
        // - Processor tasks: 2 tasks (fanout=2)
        // - Processor->Collector channel: 7 slots (processor_channel_size=7)
        // - Collector->Committer channel: 6 slots (collector_channel_size=6)
        // - Committer task: 1 task (blocked by slow commit)
        //
        // Total theoretical capacity: 3 + 2 + 7 + 6 + 1 = 19 checkpoints

        // Fill pipeline to theoretical capacity - these should all succeed
        for i in 0..19 {
            setup
                .send_checkpoint_with_timeout(i, Duration::from_millis(100))
                .await
                .unwrap();
        }

        // Find the actual back pressure point
        // Due to non-determinism in collector's tokio::select!, the collector may consume
        // up to 2 checkpoints (filling MAX_PENDING_ROWS=4) before applying back pressure.
        // This means back pressure occurs somewhere in range 19-21.
        let mut back_pressure_checkpoint = None;
        for checkpoint in 19..22 {
            if setup
                .send_checkpoint_with_timeout(checkpoint, Duration::from_millis(100))
                .await
                .is_err()
            {
                back_pressure_checkpoint = Some(checkpoint);
                break;
            }
        }
        assert!(
            back_pressure_checkpoint.is_some(),
            "Back pressure should occur between checkpoints 19-21"
        );

        // Verify that some data has been processed (pipeline is working)
        setup
            .store
            .wait_for_any_data(DataPipeline::NAME, TEST_TIMEOUT)
            .await;

        // Allow pipeline to drain by sending the blocked checkpoint with longer timeout
        setup
            .send_checkpoint_with_timeout(back_pressure_checkpoint.unwrap(), TEST_TIMEOUT)
            .await
            .unwrap();
    }

    // ==================== FAILURE TESTING ====================

    #[tokio::test]
    async fn test_commit_failure_retry() {
        let config = ConcurrentConfig::default();
        let store = MockStore::default().with_commit_failures(2); // Fail 2 times, then succeed
        let setup = TestSetup::new(config, store, 0).await;

        // Send a checkpoint
        setup
            .send_checkpoint_with_timeout(0, Duration::from_millis(200))
            .await
            .unwrap();

        // Should eventually succeed despite initial commit failures
        setup
            .store
            .wait_for_watermark(DataPipeline::NAME, 0, TEST_TIMEOUT)
            .await;

        // Verify data was eventually committed
        let data = setup
            .store
            .wait_for_data(DataPipeline::NAME, 0, Duration::from_secs(1))
            .await;
        assert_eq!(data, vec![1, 2]);
    }

    #[tokio::test]
    async fn test_prune_failure_retry() {
        let config = ConcurrentConfig {
            pruner: Some(PrunerConfig {
                interval_ms: 2000, // 2 seconds interval for testing
                delay_ms: 100,     // Short delay
                retention: 2,      // Keep only 2 checkpoints
                ..Default::default()
            }),
            ..Default::default()
        };

        // Configure prune failures for range [0, 2) - fail twice then succeed
        let store = MockStore::default().with_prune_failures(0, 2, 1);
        let setup = TestSetup::new(config, store, 0).await;

        // Send enough checkpoints to trigger pruning
        for i in 0..4 {
            setup
                .send_checkpoint_with_timeout(i, Duration::from_millis(200))
                .await
                .unwrap();
        }

        // Verify data is still available BEFORE pruning kicks in
        // With long pruning interval (5s), we can safely verify data without race conditions
        for i in 0..4 {
            let data = setup
                .store
                .wait_for_data(DataPipeline::NAME, i, Duration::from_secs(1))
                .await;
            assert_eq!(data, vec![i * 10 + 1, i * 10 + 2]);
        }

        // Wait for first pruning attempt (should fail) and verify no data has been pruned
        setup
            .store
            .wait_for_prune_attempts(0, 2, 1, TEST_TIMEOUT)
            .await;
        {
            let data = setup.store.data.get(DataPipeline::NAME).unwrap();
            for i in 0..4 {
                assert!(data.contains_key(&i));
            }
        };

        // Wait for second pruning attempt (should succeed)
        setup
            .store
            .wait_for_prune_attempts(0, 2, 2, TEST_TIMEOUT)
            .await;
        {
            let data = setup.store.data.get(DataPipeline::NAME).unwrap();
            // Verify recent checkpoints are still available
            assert!(data.contains_key(&2));
            assert!(data.contains_key(&3));

            // Verify old checkpoints are pruned
            assert!(!data.contains_key(&0));
            assert!(!data.contains_key(&1));
        };
    }

    #[tokio::test]
    async fn test_reader_watermark_failure_retry() {
        let config = ConcurrentConfig {
            pruner: Some(PrunerConfig {
                interval_ms: 100, // Fast interval for testing
                delay_ms: 100,    // Short delay
                retention: 3,     // Keep 3 checkpoints
                ..Default::default()
            }),
            ..Default::default()
        };

        // Configure reader watermark failures - fail 2 times then succeed
        let store = MockStore::default().with_reader_watermark_failures(2);
        let setup = TestSetup::new(config, store, 0).await;

        // Send checkpoints to trigger reader watermark updates
        for i in 0..6 {
            setup
                .send_checkpoint_with_timeout(i, Duration::from_millis(200))
                .await
                .unwrap();
        }

        // Wait for processing to complete
        setup
            .store
            .wait_for_watermark(DataPipeline::NAME, 5, TEST_TIMEOUT)
            .await;

        // Wait for reader watermark task to attempt updates (with failures and retries)
        tokio::time::sleep(Duration::from_secs(2)).await;

        // Verify that reader watermark was eventually updated despite failures
        let watermark = setup.store.watermark(DataPipeline::NAME).unwrap();
        assert_eq!(watermark.reader_lo, 3);
    }

    #[tokio::test]
    async fn test_database_connection_failure_retry() {
        let config = ConcurrentConfig::default();
        let store = MockStore::default().with_connection_failures(2); // Fail 2 times, then succeed
        let setup = TestSetup::new(config, store, 0).await;

        // Send a checkpoint
        setup
            .send_checkpoint_with_timeout(0, Duration::from_millis(200))
            .await
            .unwrap();

        // Should eventually succeed despite initial failures
        setup
            .store
            .wait_for_watermark(DataPipeline::NAME, 0, TEST_TIMEOUT)
            .await;

        // Verify data was eventually committed
        let data = setup
            .store
            .wait_for_data(DataPipeline::NAME, 0, TEST_TIMEOUT)
            .await;
        assert_eq!(data, vec![1, 2]);
    }
}