risingwave_meta/barrier/

schedule.rs

// Copyright 2022 RisingWave Labs
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//     http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.

use std::collections::hash_map::Entry;
use std::collections::{HashMap, HashSet, VecDeque};
use std::future::Future;
use std::sync::Arc;

use anyhow::{Context, anyhow};
use assert_matches::assert_matches;
use await_tree::InstrumentAwait;
use itertools::Itertools;
use parking_lot::Mutex;
use risingwave_common::catalog::{DatabaseId, TableId};
use risingwave_common::id::JobId;
use risingwave_hummock_sdk::HummockVersionId;
use risingwave_pb::catalog::Database;
use rw_futures_util::pending_on_none;
use tokio::select;
use tokio::sync::watch;
use tokio::time::{Duration, Instant};
use tokio_stream::wrappers::IntervalStream;
use tokio_stream::{StreamExt, StreamMap};
use tracing::{info, warn};

use super::{Command, Scheduled};
use crate::barrier::context::GlobalBarrierWorkerContext;
use crate::hummock::HummockManagerRef;
use crate::notification::{Notifier, wait_collection};
use crate::rpc::metrics::GLOBAL_META_METRICS;
use crate::{MetaError, MetaResult};

pub(super) struct NewBarrier<'a> {
    pub database_id: DatabaseId,
    pub command: Option<(Command, Notifier)>,
    pub span: tracing::Span,
    pub checkpoint: bool,
    pub periodic_barriers: &'a mut PeriodicBarriers,
    periodic: bool,
}

impl NewBarrier<'_> {
    /// Reset the periodic timer of the database, unless the barrier is the periodic one, and
    /// update its checkpoint counter. Call this only after the barrier is injected; a barrier
    /// discarded before injection, e.g. a no-op or rejected reschedule intent, is dropped
    /// instead and leaves both untouched.
    pub(super) fn injected(self) {
        if !self.periodic {
            self.periodic_barriers
                .reset_database_timer(self.database_id);
        }
        self.periodic_barriers
            .update_num_uncheckpointed_barrier(self.database_id, self.checkpoint);
    }

    #[cfg(test)]
    pub(super) fn for_test(
        database_id: DatabaseId,
        command: Option<(Command, Notifier)>,
        checkpoint: bool,
        periodic_barriers: &mut PeriodicBarriers,
    ) -> NewBarrier<'_> {
        NewBarrier {
            database_id,
            command,
            span: tracing::Span::none(),
            checkpoint,
            periodic_barriers,
            periodic: false,
        }
    }
}

/// A queue for scheduling barriers.
struct Inner {
    queue: Mutex<ScheduledQueue>,

    /// When `queue` is not empty anymore, all subscribers of this watcher will be notified.
    changed_tx: watch::Sender<()>,
}

#[derive(Debug)]
enum QueueStatus {
    /// The queue is ready to accept new command.
    Ready,
    /// The queue is blocked to accept new command with the given reason.
    Blocked(String),
}

impl QueueStatus {
    fn is_blocked(&self) -> bool {
        matches!(self, Self::Blocked(_))
    }
}

struct ScheduledQueueItem {
    command: Command,
    notifier: Notifier,
    span: tracing::Span,
}

struct StatusQueue<T> {
    queue: T,
    status: QueueStatus,
}

type DatabaseScheduledQueue = StatusQueue<VecDeque<ScheduledQueueItem>>;
type ScheduledQueue = StatusQueue<HashMap<DatabaseId, DatabaseScheduledQueue>>;

impl DatabaseScheduledQueue {
    fn new(status: QueueStatus) -> Self {
        Self {
            queue: Default::default(),
            status,
        }
    }
}

impl<T> StatusQueue<T> {
    fn mark_blocked(&mut self, reason: String) {
        self.status = QueueStatus::Blocked(reason);
    }

    fn mark_ready(&mut self) -> bool {
        let prev_blocked = self.status.is_blocked();
        self.status = QueueStatus::Ready;
        prev_blocked
    }

    fn validate_item(&mut self, command: &Command) -> MetaResult<()> {
        // We don't allow any command to be scheduled when the queue is blocked, except for dropping streaming jobs.
        // Because we allow dropping streaming jobs when the cluster is under recovery, so we have to buffer the drop
        // command and execute it when the cluster is ready to clean up it.
        // TODO: this is just a workaround to allow dropping streaming jobs when the cluster is under recovery,
        // we need to refine it when catalog and streaming metadata can be handled in a transactional way.
        if let QueueStatus::Blocked(reason) = &self.status
            && !matches!(
                command,
                Command::DropStreamingJobs { .. } | Command::DropSubscription { .. }
            )
        {
            return Err(MetaError::unavailable(reason));
        }
        Ok(())
    }
}

fn tracing_span() -> tracing::Span {
    if tracing::Span::current().is_none() {
        tracing::Span::none()
    } else {
        tracing::info_span!(
            "barrier",
            checkpoint = tracing::field::Empty,
            epoch = tracing::field::Empty
        )
    }
}

/// The sender side of the barrier scheduling queue.
/// Can be cloned and held by other managers to schedule and run barriers.
#[derive(Clone)]
pub struct BarrierScheduler {
    inner: Arc<Inner>,

    /// Used for getting the latest snapshot after `FLUSH`.
    hummock_manager: HummockManagerRef,
}

impl BarrierScheduler {
    /// Create a pair of [`BarrierScheduler`] and [`ScheduledBarriers`], for scheduling barriers
    /// from different managers, and executing them in the barrier manager, respectively.
    pub fn new_pair(hummock_manager: HummockManagerRef) -> (Self, ScheduledBarriers) {
        let inner = Arc::new(Inner {
            queue: Mutex::new(ScheduledQueue {
                queue: Default::default(),
                status: QueueStatus::Ready,
            }),
            changed_tx: watch::channel(()).0,
        });

        (
            Self {
                inner: inner.clone(),
                hummock_manager,
            },
            ScheduledBarriers { inner },
        )
    }

    /// Push a scheduled barrier into the queue.
    fn push(
        &self,
        database_id: DatabaseId,
        scheduleds: impl IntoIterator<Item = (Command, Notifier)>,
    ) -> MetaResult<()> {
        let mut queue = self.inner.queue.lock();
        let scheduleds = scheduleds.into_iter().collect_vec();
        scheduleds
            .iter()
            .try_for_each(|(command, _)| queue.validate_item(command))?;
        let queue = queue
            .queue
            .entry(database_id)
            .or_insert_with(|| DatabaseScheduledQueue::new(QueueStatus::Ready));
        scheduleds
            .iter()
            .try_for_each(|(command, _)| queue.validate_item(command))?;
        for (command, notifier) in scheduleds {
            queue.queue.push_back(ScheduledQueueItem {
                command,
                notifier,
                span: tracing_span(),
            });
            if queue.queue.len() == 1 {
                self.inner.changed_tx.send(()).ok();
            }
        }
        Ok(())
    }

    /// Try to cancel scheduled cmd for create streaming job, return true if the command exists previously and get cancelled.
    pub fn try_cancel_scheduled_create(&self, database_id: DatabaseId, job_id: JobId) -> bool {
        let queue = &mut self.inner.queue.lock();
        let Some(queue) = queue.queue.get_mut(&database_id) else {
            return false;
        };

        if let Some(idx) = queue.queue.iter().position(|scheduled| {
            if let Command::CreateStreamingJob { info, .. } = &scheduled.command
                && info.stream_job_fragments.stream_job_id() == job_id
            {
                true
            } else {
                false
            }
        }) {
            queue.queue.remove(idx).unwrap();
            true
        } else {
            false
        }
    }

    /// Run multiple commands and return when they're all completely finished (i.e., collected). It's ensured that
    /// multiple commands are executed continuously.
    ///
    /// Run a command and return when it's completely finished (i.e., collected).
    #[await_tree::instrument("run_command({})", command)]
    pub async fn run_command(&self, database_id: DatabaseId, command: Command) -> MetaResult<()> {
        self.schedule_command(database_id, command)?.await
    }

    /// Schedule a command and return a future that resolves once it is collected.
    pub fn schedule_command(
        &self,
        database_id: DatabaseId,
        command: Command,
    ) -> MetaResult<impl Future<Output = MetaResult<()>> + use<>> {
        tracing::trace!("schedule_command: {:?}", command);
        let (notifier, started_rx) = Notifier::new();
        self.push(database_id, vec![(command, notifier)])?;
        Ok(async move {
            let collect_rxs = started_rx
                .instrument_await("wait_injected")
                .await
                .ok()
                .context("failed to inject barrier")??;
            tracing::trace!(
                collection_count = collect_rxs.len(),
                "waiting for collect_rx"
            );
            // Wait for every part before returning the first collection error.
            wait_collection(collect_rxs)
                .instrument_await("wait_collected")
                .await
        })
    }

    /// Schedule a command without waiting for it to be executed.
    pub fn run_command_no_wait(&self, database_id: DatabaseId, command: Command) -> MetaResult<()> {
        tracing::trace!("run_command_no_wait: {:?}", command);
        let (notifier, _started_rx) = Notifier::new();
        self.push(database_id, vec![(command, notifier)])
    }

    /// Flush means waiting for the next barrier to collect.
    pub async fn flush(&self, database_id: DatabaseId) -> MetaResult<HummockVersionId> {
        let start = Instant::now();

        tracing::debug!("start barrier flush");
        self.run_command(database_id, Command::Flush).await?;

        let elapsed = Instant::now().duration_since(start);
        tracing::debug!("barrier flushed in {:?}", elapsed);

        let version_id = self.hummock_manager.get_version_id().await;
        Ok(version_id)
    }
}

/// The receiver side of the barrier scheduling queue.
pub struct ScheduledBarriers {
    inner: Arc<Inner>,
}

/// State specific to each database for barrier generation.
#[derive(Debug)]
pub struct DatabaseBarrierState {
    barrier_interval: Option<Duration>,
    checkpoint_frequency: Option<u64>,
    // The numbers of barrier (checkpoint = false) since the last barrier (checkpoint = true)
    num_uncheckpointed_barrier: u64,
}

impl DatabaseBarrierState {
    fn new(barrier_interval_ms: Option<u32>, checkpoint_frequency: Option<u64>) -> Self {
        Self {
            barrier_interval: barrier_interval_ms.map(|ms| Duration::from_millis(ms as u64)),
            checkpoint_frequency,
            num_uncheckpointed_barrier: 0,
        }
    }
}

/// Held by the [`crate::barrier::worker::GlobalBarrierWorker`] to execute these commands.
#[derive(Default, Debug)]
pub struct PeriodicBarriers {
    /// Default system params for barrier interval and checkpoint frequency.
    sys_barrier_interval: Duration,
    sys_checkpoint_frequency: u64,
    /// Per-database state.
    databases: HashMap<DatabaseId, DatabaseBarrierState>,
    /// Holds `IntervalStream` for each database, keyed by `DatabaseId`.
    /// `StreamMap` will yield `(DatabaseId, Instant)` when a timer ticks.
    timer_streams: StreamMap<DatabaseId, IntervalStream>,
    force_checkpoint_databases: HashSet<DatabaseId>,
}

impl PeriodicBarriers {
    pub(super) fn new(
        sys_barrier_interval: Duration,
        sys_checkpoint_frequency: u64,
        database_infos: Vec<Database>,
    ) -> Self {
        let mut databases = HashMap::with_capacity(database_infos.len());
        let mut timer_streams = StreamMap::with_capacity(database_infos.len());
        database_infos.into_iter().for_each(|database| {
            let database_id: DatabaseId = database.id;
            let barrier_interval_ms = database.barrier_interval_ms;
            let checkpoint_frequency = database.checkpoint_frequency;
            databases.insert(
                database_id,
                DatabaseBarrierState::new(barrier_interval_ms, checkpoint_frequency),
            );
            let duration = if let Some(ms) = barrier_interval_ms {
                Duration::from_millis(ms as u64)
            } else {
                sys_barrier_interval
            };

            // Create an `IntervalStream` for the database with the specified interval.
            let interval_stream = Self::new_interval_stream(duration, &database_id);
            timer_streams.insert(database_id, interval_stream);
        });
        Self {
            sys_barrier_interval,
            sys_checkpoint_frequency,
            databases,
            timer_streams,
            force_checkpoint_databases: Default::default(),
        }
    }

    // Create a new interval stream with the specified duration.
    fn new_interval_stream(duration: Duration, database_id: &DatabaseId) -> IntervalStream {
        GLOBAL_META_METRICS
            .barrier_interval_by_database
            .with_label_values(&[&database_id.to_string()])
            .set(duration.as_millis_f64());
        let mut interval = tokio::time::interval(duration);
        interval.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Delay);
        IntervalStream::new(interval)
    }

    /// Update the system barrier interval.
    pub(super) fn set_sys_barrier_interval(&mut self, duration: Duration) {
        if self.sys_barrier_interval == duration {
            return;
        }
        self.sys_barrier_interval = duration;
        // Reset the `IntervalStream` for all databases that use default param.
        for (db_id, db_state) in &mut self.databases {
            if db_state.barrier_interval.is_none() {
                let interval_stream = Self::new_interval_stream(duration, db_id);
                self.timer_streams.insert(*db_id, interval_stream);
            }
        }
    }

    /// Update the system checkpoint frequency.
    pub fn set_sys_checkpoint_frequency(&mut self, frequency: u64) {
        if self.sys_checkpoint_frequency == frequency {
            return;
        }
        self.sys_checkpoint_frequency = frequency;
        // Reset the `num_uncheckpointed_barrier` for all databases that use default param.
        for db_state in self.databases.values_mut() {
            if db_state.checkpoint_frequency.is_none() {
                db_state.num_uncheckpointed_barrier = 0;
            }
        }
    }

    pub(super) fn update_database_barrier(
        &mut self,
        database_id: DatabaseId,
        barrier_interval_ms: Option<u32>,
        checkpoint_frequency: Option<u64>,
    ) {
        match self.databases.entry(database_id) {
            Entry::Occupied(mut entry) => {
                let db_state = entry.get_mut();
                db_state.barrier_interval =
                    barrier_interval_ms.map(|ms| Duration::from_millis(ms as u64));
                db_state.checkpoint_frequency = checkpoint_frequency;
                // Reset the `num_uncheckpointed_barrier` since the barrier interval or checkpoint frequency is changed.
                db_state.num_uncheckpointed_barrier = 0;
            }
            Entry::Vacant(entry) => {
                entry.insert(DatabaseBarrierState::new(
                    barrier_interval_ms,
                    checkpoint_frequency,
                ));
            }
        }

        // If the database already has a timer stream, reset it with the new interval.
        let duration = if let Some(ms) = barrier_interval_ms {
            Duration::from_millis(ms as u64)
        } else {
            self.sys_barrier_interval
        };

        let interval_stream = Self::new_interval_stream(duration, &database_id);
        self.timer_streams.insert(database_id, interval_stream);
    }

    /// Make the `checkpoint` of the next barrier must be true.
    pub fn force_checkpoint_in_next_barrier(&mut self, database_id: DatabaseId) {
        if self.databases.contains_key(&database_id) {
            self.force_checkpoint_databases.insert(database_id);
        } else {
            warn!(
                ?database_id,
                "force checkpoint in next barrier for non-existing database"
            );
        }
    }

    fn reset_database_timer(&mut self, database_id: DatabaseId) {
        // Check if the database exists.
        assert!(
            self.databases.contains_key(&database_id),
            "database {} not found in scheduled barriers",
            database_id
        );
        assert!(
            self.timer_streams.contains_key(&database_id),
            "timer stream for database {} not found in scheduled barriers",
            database_id
        );
        // New command will trigger the barriers, so reset the timer for the specific database.
        for (db_id, timer_stream) in self.timer_streams.iter_mut() {
            if *db_id == database_id {
                timer_stream.as_mut().reset();
            }
        }
    }

    #[await_tree::instrument]
    pub(super) async fn next_barrier(
        &mut self,
        context: &impl GlobalBarrierWorkerContext,
    ) -> NewBarrier<'_> {
        let force_checkpoint_database = self.force_checkpoint_databases.extract_if(|_| true).next();
        if let Some(database_id) = force_checkpoint_database {
            NewBarrier {
                database_id,
                command: None,
                span: tracing_span(),
                checkpoint: true,
                periodic_barriers: self,
                periodic: false,
            }
        } else {
            select! {
                biased;
                scheduled = context.next_scheduled() => {
                    let database_id = scheduled.database_id;
                    let checkpoint = scheduled.command.need_checkpoint() || self.try_get_checkpoint(database_id);
                    NewBarrier {
                        database_id: scheduled.database_id,
                        command: Some((scheduled.command, scheduled.notifier)),
                        span: scheduled.span,
                        checkpoint,
                        periodic_barriers: self,
                        periodic: false,
                    }
                },
                // If there is no database, we won't wait for `Interval`, but only wait for command.
                // Normally it will not return None, because there is always at least one database.
                (database_id, _instant) = pending_on_none(self.timer_streams.next()) => {
                    let checkpoint = self.try_get_checkpoint(database_id);
                    NewBarrier {
                        database_id,
                        command: None,
                        span: tracing_span(),
                        checkpoint,
                        periodic_barriers: self,
                        periodic: true,
                    }
                }
            }
        }
    }

    /// Whether the barrier(checkpoint = true) should be injected.
    fn try_get_checkpoint(&self, database_id: DatabaseId) -> bool {
        let db_state = self.databases.get(&database_id).unwrap();
        let checkpoint_frequency = db_state
            .checkpoint_frequency
            .unwrap_or(self.sys_checkpoint_frequency);
        db_state.num_uncheckpointed_barrier + 1 >= checkpoint_frequency
    }

    /// Update the `num_uncheckpointed_barrier`
    fn update_num_uncheckpointed_barrier(&mut self, database_id: DatabaseId, checkpoint: bool) {
        let db_state = self.databases.get_mut(&database_id).unwrap();
        if checkpoint {
            db_state.num_uncheckpointed_barrier = 0;
        } else {
            db_state.num_uncheckpointed_barrier += 1;
        }
    }
}

impl ScheduledBarriers {
    pub(super) async fn next_scheduled(&self) -> Scheduled {
        'outer: loop {
            let mut rx = self.inner.changed_tx.subscribe();
            {
                let mut queue = self.inner.queue.lock();
                if queue.status.is_blocked() {
                    continue;
                }
                for (database_id, queue) in &mut queue.queue {
                    if queue.status.is_blocked() {
                        continue;
                    }
                    if let Some(item) = queue.queue.pop_front() {
                        break 'outer Scheduled {
                            database_id: *database_id,
                            command: item.command,
                            notifier: item.notifier,
                            span: item.span,
                        };
                    }
                }
            }
            rx.changed().await.unwrap();
        }
    }
}

pub(super) enum MarkReadyOptions {
    Database(DatabaseId),
    Global {
        failed_databases: HashMap<DatabaseId, HashSet<JobId>>,
    },
}

pub(super) struct PreApplyDropCancel {
    pub streaming_job_ids: Vec<JobId>,
    pub dropped_state_table_ids: Vec<TableId>,
}

impl ScheduledBarriers {
    /// Pre buffered drop and cancel command, return all dropped state tables if any.
    pub(super) fn pre_apply_drop_cancel(
        &self,
        database_id: Option<DatabaseId>,
    ) -> PreApplyDropCancel {
        self.pre_apply_drop_cancel_scheduled(database_id)
    }

    /// Mark command scheduler as blocked and abort all queued scheduled command and notify with
    /// specific reason.
    pub(super) fn abort_and_mark_blocked(
        &self,
        database_id: Option<DatabaseId>,
        reason: impl Into<String>,
    ) {
        let mut queue = self.inner.queue.lock();
        fn database_blocked_reason(database_id: DatabaseId, reason: &String) -> String {
            format!("database {} unavailable {}", database_id, reason)
        }
        fn mark_blocked_and_notify_failed(
            database_id: DatabaseId,
            queue: &mut DatabaseScheduledQueue,
            reason: &String,
        ) {
            let reason = database_blocked_reason(database_id, reason);
            let err: MetaError = anyhow!("{}", reason).into();
            queue.mark_blocked(reason);
            while let Some(ScheduledQueueItem { notifier, .. }) = queue.queue.pop_front() {
                notifier.notify_start_failed(err.clone());
            }
        }
        if let Some(database_id) = database_id {
            let reason = reason.into();
            match queue.queue.entry(database_id) {
                Entry::Occupied(entry) => {
                    let queue = entry.into_mut();
                    if queue.status.is_blocked() {
                        if cfg!(debug_assertions) {
                            panic!("database {} marked as blocked twice", database_id);
                        } else {
                            warn!(?database_id, "database marked as blocked twice");
                        }
                    }
                    info!(?database_id, "database marked as blocked");
                    mark_blocked_and_notify_failed(database_id, queue, &reason);
                }
                Entry::Vacant(entry) => {
                    entry.insert(DatabaseScheduledQueue::new(QueueStatus::Blocked(
                        database_blocked_reason(database_id, &reason),
                    )));
                }
            }
        } else {
            let reason = reason.into();
            if queue.status.is_blocked() {
                if cfg!(debug_assertions) {
                    panic!("cluster marked as blocked twice");
                } else {
                    warn!("cluster marked as blocked twice");
                }
            }
            info!("cluster marked as blocked");
            queue.mark_blocked(reason.clone());
            for (database_id, queue) in &mut queue.queue {
                mark_blocked_and_notify_failed(*database_id, queue, &reason);
            }
        }
    }

    /// Mark command scheduler as ready to accept new command.
    pub(super) fn mark_ready(&self, options: MarkReadyOptions) {
        let mut queue = self.inner.queue.lock();
        let queue = &mut *queue;
        match options {
            MarkReadyOptions::Database(database_id) => {
                info!(?database_id, "database marked as ready");
                let database_queue = queue
                    .queue
                    .entry(database_id)
                    .or_insert_with(|| DatabaseScheduledQueue::new(QueueStatus::Ready));
                if !database_queue.status.is_blocked() {
                    if cfg!(debug_assertions) {
                        panic!("database {} marked as ready twice", database_id);
                    } else {
                        warn!(?database_id, "database marked as ready twice");
                    }
                }
                if database_queue.mark_ready()
                    && !queue.status.is_blocked()
                    && !database_queue.queue.is_empty()
                {
                    self.inner.changed_tx.send(()).ok();
                }
            }
            MarkReadyOptions::Global { failed_databases } => {
                if !queue.status.is_blocked() {
                    if cfg!(debug_assertions) {
                        panic!("cluster marked as ready twice");
                    } else {
                        warn!("cluster marked as ready twice");
                    }
                }
                info!(
                    failed_database_ids = ?failed_databases.keys().collect_vec(),
                    "cluster marked as ready"
                );
                let prev_blocked = queue.mark_ready();
                for database_id in failed_databases.keys() {
                    queue.queue.entry(*database_id).or_insert_with(|| {
                        DatabaseScheduledQueue::new(QueueStatus::Blocked(format!(
                            "database {} failed to recover in global recovery",
                            database_id
                        )))
                    });
                }
                for (database_id, queue) in &mut queue.queue {
                    if !failed_databases.contains_key(database_id) {
                        queue.mark_ready();
                    }
                }
                if prev_blocked
                    && queue
                        .queue
                        .values()
                        .any(|database_queue| !database_queue.queue.is_empty())
                {
                    self.inner.changed_tx.send(()).ok();
                }
            }
        }
    }

    /// Try to pre apply drop and cancel scheduled command and return all dropped state tables if any.
    /// It should only be called in recovery.
    pub(super) fn pre_apply_drop_cancel_scheduled(
        &self,
        database_id: Option<DatabaseId>,
    ) -> PreApplyDropCancel {
        let mut queue = self.inner.queue.lock();
        let mut drop_cancel = PreApplyDropCancel {
            streaming_job_ids: vec![],
            dropped_state_table_ids: vec![],
        };

        let mut pre_apply_drop_cancel = |queue: &mut DatabaseScheduledQueue| {
            while let Some(ScheduledQueueItem {
                notifier, command, ..
            }) = queue.queue.pop_front()
            {
                match command {
                    Command::DropStreamingJobs {
                        streaming_job_ids,
                        unregistered_state_table_ids,
                        ..
                    } => {
                        drop_cancel.streaming_job_ids.extend(streaming_job_ids);
                        drop_cancel
                            .dropped_state_table_ids
                            .extend(unregistered_state_table_ids);
                    }
                    Command::DropSubscription { .. } => {}
                    _ => {
                        unreachable!("only drop and cancel streaming jobs should be buffered");
                    }
                }
                // These buffered commands are pre-applied without injecting a real barrier.
                notifier.start().started();
            }
        };

        if let Some(database_id) = database_id {
            assert_matches!(queue.status, QueueStatus::Ready);
            if let Some(queue) = queue.queue.get_mut(&database_id) {
                assert_matches!(queue.status, QueueStatus::Blocked(_));
                pre_apply_drop_cancel(queue);
            }
        } else {
            assert_matches!(queue.status, QueueStatus::Blocked(_));
            for queue in queue.queue.values_mut() {
                pre_apply_drop_cancel(queue);
            }
        }

        drop_cancel
    }
}

#[cfg(test)]
mod tests {
    use futures::FutureExt;
    use risingwave_meta_model::PartialGraphId;

    use super::*;
    use crate::barrier::RescheduleContext;

    fn create_test_database(
        id: u32,
        barrier_interval_ms: Option<u32>,
        checkpoint_frequency: Option<u64>,
    ) -> Database {
        Database {
            id: id.into(),
            name: format!("test_db_{}", id),
            barrier_interval_ms,
            checkpoint_frequency,
            ..Default::default()
        }
    }

    fn noop_reschedule_intent(database_id: DatabaseId) -> Scheduled {
        let (notifier, _) = Notifier::new();
        Scheduled {
            database_id,
            command: Command::RescheduleIntent {
                context: RescheduleContext::empty(),
                reschedule_plan: None,
            },
            notifier,
            span: tracing::Span::none(),
        }
    }

    // Mock context for testing next_barrier
    struct MockGlobalBarrierWorkerContext {
        scheduled_rx: tokio::sync::Mutex<tokio::sync::mpsc::UnboundedReceiver<Scheduled>>,
    }

    impl MockGlobalBarrierWorkerContext {
        fn new() -> (Self, tokio::sync::mpsc::UnboundedSender<Scheduled>) {
            let (tx, rx) = tokio::sync::mpsc::unbounded_channel();
            (
                Self {
                    scheduled_rx: tokio::sync::Mutex::new(rx),
                },
                tx,
            )
        }
    }

    impl GlobalBarrierWorkerContext for MockGlobalBarrierWorkerContext {
        async fn next_scheduled(&self) -> Scheduled {
            self.scheduled_rx.lock().await.recv().await.unwrap()
        }

        async fn commit_epoch(
            &self,
            _commit_info: crate::hummock::CommitEpochInfo,
        ) -> MetaResult<risingwave_pb::hummock::HummockVersionStats> {
            unimplemented!()
        }

        async fn abort_and_mark_blocked(
            &self,
            _recovery: crate::manager::sink_coordination::RecoveryStart,
            _recovery_reason: &crate::barrier::RecoveryReason,
        ) -> MetaResult<()> {
            unimplemented!()
        }

        async fn mark_ready(
            &self,
            _options: MarkReadyOptions,
            _running_job_ids: HashSet<JobId>,
        ) -> MetaResult<()> {
            unimplemented!()
        }

        async fn resolve_log_store_epoch<'a>(
            &'a self,
            _upstream_table_ids: impl Iterator<Item = risingwave_common::catalog::TableId> + Send + 'a,
            _since_epoch: u64,
        ) -> MetaResult<crate::barrier::command::SinceTimestampResolvedEpoch> {
            Ok(Default::default())
        }

        async fn post_collect_command(
            &self,
            _command: crate::barrier::command::PostCollectCommand,
        ) -> MetaResult<()> {
            unimplemented!()
        }

        async fn notify_creating_job_failed(&self, _database_id: Option<DatabaseId>, _err: String) {
            unimplemented!()
        }

        async fn finish_creating_job(
            &self,
            _job: crate::barrier::progress::TrackingJob,
        ) -> MetaResult<()> {
            unimplemented!()
        }

        async fn new_control_stream(
            &self,
            _node: &risingwave_pb::common::WorkerNode,
        ) -> MetaResult<risingwave_rpc_client::StreamingControlHandle> {
            unimplemented!()
        }

        async fn reload_runtime_info(
            &self,
        ) -> MetaResult<crate::barrier::BarrierWorkerRuntimeInfoSnapshot> {
            unimplemented!()
        }

        async fn reload_database_runtime_info(
            &self,
            _database_id: DatabaseId,
        ) -> MetaResult<crate::barrier::DatabaseRuntimeInfoSnapshot> {
            unimplemented!()
        }

        async fn handle_list_finished_source_ids(
            &self,
            _list_finished_source_ids: Vec<
                risingwave_pb::stream_service::barrier_complete_response::PbListFinishedSource,
            >,
        ) -> MetaResult<()> {
            unimplemented!()
        }

        async fn handle_load_finished_source_ids(
            &self,
            _load_finished_source_ids: Vec<
                risingwave_pb::stream_service::barrier_complete_response::PbLoadFinishedSource,
            >,
        ) -> MetaResult<()> {
            unimplemented!()
        }

        async fn finish_cdc_table_backfill(&self, _job_id: JobId) -> MetaResult<()> {
            unimplemented!()
        }

        async fn handle_refresh_finished_actors(
            &self,
            _refresh_finished_actors: Vec<
                risingwave_pb::stream_service::barrier_complete_response::PbRefreshFinishedActor,
            >,
        ) -> MetaResult<()> {
            unimplemented!()
        }

        async fn load_batch_refresh_trigger_context(
            &self,
            _job_id: JobId,
            _database_id: DatabaseId,
            _last_committed_epoch: u64,
        ) -> MetaResult<crate::barrier::checkpoint::independent_job::BatchRefreshJobTriggerContext>
        {
            unimplemented!()
        }

        async fn pre_commit_iceberg_pk_index_sink_metadata(
            &self,
            _metadata: Vec<crate::manager::iceberg_pk_index_sink::IcebergPkIndexPreCommitMetadata>,
        ) -> MetaResult<Vec<risingwave_meta_model::SinkId>> {
            unimplemented!()
        }

        async fn commit_iceberg_pk_index_sink_metadata(
            &self,
            _sink_ids: Vec<risingwave_meta_model::SinkId>,
        ) -> MetaResult<()> {
            unimplemented!()
        }

        fn advance_iceberg_pk_index_sink_committed_epochs(
            &self,
            _epochs: impl IntoIterator<Item = (PartialGraphId, u64)>,
        ) {
            unimplemented!()
        }
    }

    #[tokio::test(start_paused = true)]
    async fn test_next_barrier_with_different_intervals() {
        // Create databases with different intervals
        let databases = vec![
            create_test_database(1, Some(50), Some(2)), // 50ms interval, checkpoint every 2
            create_test_database(2, Some(100), Some(3)), // 100ms interval, checkpoint every 3
            create_test_database(3, None, Some(5)), /* Use system default (200ms), checkpoint every 5 */
        ];

        let mut periodic = PeriodicBarriers::new(
            Duration::from_millis(200), // System default
            10,                         // System checkpoint frequency
            databases,
        );

        let (context, _tx) = MockGlobalBarrierWorkerContext::new();

        // Call next_barrier for each database once, because the first tick is returned immediately
        for _ in 0..3 {
            let barrier = periodic.next_barrier(&context).await;
            assert!(barrier.command.is_none()); // Should be a periodic barrier, not a scheduled command
            assert!(!barrier.checkpoint); // First barrier shouldn't be a checkpoint
            barrier.injected();
        }

        // Since we have 3 databases with intervals 50ms, 100ms, and 200ms,
        // the first barrier should come from database 1 (50ms interval)
        let start_time = Instant::now();
        let barrier = periodic.next_barrier(&context).await;
        let mut elapsed = start_time.elapsed();

        // Verify the barrier properties
        assert_eq!(barrier.database_id, DatabaseId::from(1));
        assert!(barrier.command.is_none()); // Should be a periodic barrier, not a scheduled command
        assert!(barrier.checkpoint); // Second barrier should be checkpoint for database 1
        barrier.injected();
        // Use tokio's time pause mechanism, so it will be exactly 50ms here.
        assert_eq!(
            elapsed,
            Duration::from_millis(50),
            "Elapsed time exceeded: {:?}",
            elapsed
        );

        // Verify that the checkpoint frequency works
        let db1_id = DatabaseId::from(1);
        let db1_state = periodic.databases.get_mut(&db1_id).unwrap();
        assert_eq!(db1_state.num_uncheckpointed_barrier, 0); // Should reset after checkpoint

        // Next barrier should come from database 1 and database 2 at 100ms
        for _ in 0..2 {
            let barrier = periodic.next_barrier(&context).await;
            assert!(barrier.command.is_none()); // Should be a periodic barrier, not a scheduled command
            assert!(!barrier.checkpoint); // Next two barriers shouldn't be checkpoints
            barrier.injected();
        }

        elapsed = start_time.elapsed();

        assert_eq!(
            elapsed,
            Duration::from_millis(100),
            "Elapsed time exceeded: {:?}",
            elapsed
        );
    }

    #[tokio::test(start_paused = true)]
    async fn test_next_barrier_with_scheduled_command() {
        let databases = vec![
            create_test_database(1, Some(1000), Some(2)), // Long interval to avoid interference
        ];

        let mut periodic = PeriodicBarriers::new(Duration::from_millis(1000), 10, databases);

        let (context, tx) = MockGlobalBarrierWorkerContext::new();

        // Skip the first barrier to let the timers start
        periodic.next_barrier(&context).await.injected();

        // Schedule a command
        let (notifier, _started_rx) = Notifier::new();
        let scheduled_command = Scheduled {
            database_id: DatabaseId::from(1),
            command: Command::Flush,
            notifier,
            span: tracing::Span::none(),
        };

        // Send scheduled command in background
        let tx_clone = tx.clone();
        tokio::spawn(async move {
            tokio::time::sleep(Duration::from_millis(10)).await;
            tx_clone.send(scheduled_command).unwrap();
        });

        let barrier = periodic.next_barrier(&context).await;

        // Should return the scheduled command
        assert!(barrier.command.is_some());
        assert_eq!(barrier.database_id, DatabaseId::from(1));
        assert!(!barrier.periodic);

        if let Some((command, _)) = &barrier.command {
            assert!(matches!(command, Command::Flush));
        }

        barrier.injected();
        let injected_at = Instant::now();

        // Without the reset the tick would come at the original deadline, 10ms earlier.
        let barrier = periodic.next_barrier(&context).await;
        assert!(barrier.periodic);
        assert!(injected_at.elapsed() >= Duration::from_millis(1000));
    }

    #[tokio::test(start_paused = true)]
    async fn test_next_barrier_multiple_databases_timing() {
        let databases = vec![
            create_test_database(1, Some(30), Some(10)), // Fast interval
            create_test_database(2, Some(100), Some(10)), // Slower interval
        ];

        let mut periodic = PeriodicBarriers::new(Duration::from_millis(500), 10, databases);

        let (context, _tx) = MockGlobalBarrierWorkerContext::new();

        // Skip first 2 barriers to let the timers start
        for _ in 0..2 {
            periodic.next_barrier(&context).await;
        }

        let mut barrier_counts = HashMap::new();

        // Collect barriers for a short period
        let mut barriers = Vec::new();
        for _ in 0..5 {
            let barrier = periodic.next_barrier(&context).await;
            barriers.push(barrier.database_id);
        }

        // Count barriers per database
        for database_id in barriers {
            *barrier_counts.entry(database_id).or_insert(0) += 1;
        }

        // Database 1 (30ms interval) should have more barriers than database 2 (100ms interval)
        let db1_count = barrier_counts.get(&DatabaseId::from(1)).unwrap_or(&0);
        let db2_count = barrier_counts.get(&DatabaseId::from(2)).unwrap_or(&0);

        // Due to timing, db1 should generally have more barriers, but allow for some variance
        assert_eq!(*db1_count, 4);
        assert_eq!(*db2_count, 1);
    }

    #[tokio::test(start_paused = true)]
    async fn test_next_barrier_force_checkpoint() {
        let databases = vec![create_test_database(1, Some(100), Some(10))];

        let mut periodic = PeriodicBarriers::new(Duration::from_millis(100), 10, databases);

        let (context, _tx) = MockGlobalBarrierWorkerContext::new();

        // Force checkpoint for next barrier
        periodic.force_checkpoint_in_next_barrier(DatabaseId::from(1));

        let barrier = periodic.next_barrier(&context).now_or_never().unwrap();

        // Should be a checkpoint barrier due to force_checkpoint
        assert!(barrier.checkpoint);
        assert_eq!(barrier.database_id, DatabaseId::from(1));
        assert!(barrier.command.is_none());
        assert!(!barrier.periodic);

        // The forced barrier is not injected, so the timer is not reset and the periodic
        // barrier arrives without waiting, as the first tick of the timer is due at creation.
        let start_time = Instant::now();
        // A reset would have pushed the tick a full 100ms interval out.
        let barrier = periodic.next_barrier(&context).await;
        assert!(barrier.periodic);
        assert_eq!(barrier.database_id, DatabaseId::from(1));
        assert!(start_time.elapsed() < Duration::from_millis(100));
    }

    #[tokio::test(start_paused = true)]
    async fn test_next_barrier_multiple_force_checkpoints() {
        let databases = vec![
            create_test_database(1, Some(100), Some(10)),
            create_test_database(2, Some(100), Some(10)),
        ];

        let mut periodic = PeriodicBarriers::new(Duration::from_millis(100), 10, databases);

        let (context, _tx) = MockGlobalBarrierWorkerContext::new();

        periodic.force_checkpoint_in_next_barrier(DatabaseId::from(1));
        periodic.force_checkpoint_in_next_barrier(DatabaseId::from(2));

        let barrier1 = periodic.next_barrier(&context).now_or_never().unwrap();
        assert!(barrier1.checkpoint);
        assert!(barrier1.command.is_none());
        let database_id1 = barrier1.database_id;

        let barrier2 = periodic.next_barrier(&context).now_or_never().unwrap();
        assert!(barrier2.checkpoint);
        assert!(barrier2.command.is_none());
        assert_eq!(
            HashSet::from([database_id1, barrier2.database_id]),
            HashSet::from([DatabaseId::from(1), DatabaseId::from(2)])
        );
        assert!(periodic.force_checkpoint_databases.is_empty());

        // Neither forced barrier is injected, so neither timer is reset and both periodic
        // barriers arrive without waiting, as the first tick of a timer is due at creation.
        let start_time = Instant::now();
        let barrier3 = periodic.next_barrier(&context).await;
        assert!(barrier3.periodic);
        let database_id3 = barrier3.database_id;

        let barrier4 = periodic.next_barrier(&context).await;
        assert!(barrier4.periodic);
        assert_eq!(
            HashSet::from([database_id3, barrier4.database_id]),
            HashSet::from([DatabaseId::from(1), DatabaseId::from(2)])
        );
        // A reset would have pushed either tick a full 100ms interval out.
        assert!(start_time.elapsed() < Duration::from_millis(100));
    }

    #[tokio::test(start_paused = true)]
    async fn test_next_barrier_checkpoint_frequency() {
        let databases = vec![create_test_database(1, Some(50), Some(2))]; // Checkpoint every 2 barriers

        let mut periodic = PeriodicBarriers::new(Duration::from_millis(50), 10, databases);

        let (context, tx) = MockGlobalBarrierWorkerContext::new();

        // First barrier - should not be checkpoint
        let barrier1 = periodic.next_barrier(&context).await;
        assert!(!barrier1.checkpoint);
        barrier1.injected();

        // A no-op reschedule intent arrives exactly when the next tick is due. The biased
        // select takes the intent, which is then discarded without injection, so neither the
        // due tick nor the checkpoint counter must be touched.
        let start_time = Instant::now();
        tokio::time::sleep(Duration::from_millis(50)).await;
        tx.send(noop_reschedule_intent(DatabaseId::from(1)))
            .unwrap();
        let discarded = periodic.next_barrier(&context).now_or_never().unwrap();
        assert!(discarded.command.is_some());
        assert!(!discarded.periodic);

        // Second barrier - should be checkpoint (frequency = 2)
        let barrier2 = periodic.next_barrier(&context).now_or_never().unwrap();
        assert!(barrier2.periodic);
        assert!(barrier2.checkpoint);
        // A reset at the tie would have moved the tick from 50ms to 100ms.
        assert!(start_time.elapsed() < Duration::from_millis(100));
        barrier2.injected();

        // Third barrier - should not be checkpoint (counter reset)
        let barrier3 = periodic.next_barrier(&context).await;
        assert!(!barrier3.checkpoint);
        barrier3.injected();
    }

    #[tokio::test]
    async fn test_update_database_barrier() {
        let databases = vec![create_test_database(1, Some(1000), Some(10))];

        let mut periodic = PeriodicBarriers::new(Duration::from_millis(500), 20, databases);

        let database_id = DatabaseId::new(1);

        // Update existing database
        periodic.update_database_barrier(database_id, Some(2000), Some(15));

        let db_state = periodic.databases.get(&database_id).unwrap();
        assert_eq!(db_state.barrier_interval, Some(Duration::from_millis(2000)));
        assert_eq!(db_state.checkpoint_frequency, Some(15));
        assert_eq!(db_state.num_uncheckpointed_barrier, 0);
        assert!(!periodic.force_checkpoint_databases.contains(&database_id));

        // Add new database
        periodic.update_database_barrier(DatabaseId::from(2), None, None);

        assert!(periodic.databases.contains_key(&DatabaseId::from(2)));
        let db2_state = periodic.databases.get(&DatabaseId::from(2)).unwrap();
        assert_eq!(db2_state.barrier_interval, None);
        assert_eq!(db2_state.checkpoint_frequency, None);
    }
}