feat(storage): per table try wait epoch

proto/stream_service.proto

@@ -49,6 +49,7 @@ message BarrierCompleteResponse {
 
 message WaitEpochCommitRequest {
   uint64 epoch = 1;
+  uint32 table_id = 2;
 }
 
 message WaitEpochCommitResponse {

src/compute/src/rpc/service/stream_service.rs

@@ -17,6 +17,7 @@ use futures::{Stream, StreamExt, TryStreamExt};
 use risingwave_pb::stream_service::stream_service_server::StreamService;
 use risingwave_pb::stream_service::*;
 use risingwave_storage::dispatch_state_store;
+use risingwave_storage::store::TryWaitEpochOptions;
 use risingwave_stream::error::StreamError;
 use risingwave_stream::task::{LocalStreamManager, StreamEnvironment};
 use tokio::sync::mpsc::unbounded_channel;
@@ -45,14 +46,20 @@ impl StreamService for StreamServiceImpl {
         &self,
         request: Request<WaitEpochCommitRequest>,
     ) -> Result<Response<WaitEpochCommitResponse>, Status> {
-        let epoch = request.into_inner().epoch;
+        let request = request.into_inner();
+        let epoch = request.epoch;
 
         dispatch_state_store!(self.env.state_store(), store, {
             use risingwave_hummock_sdk::HummockReadEpoch;
             use risingwave_storage::StateStore;
 
             store
-                .try_wait_epoch(HummockReadEpoch::Committed(epoch))
+                .try_wait_epoch(
+                    HummockReadEpoch::Committed(epoch),
+                    TryWaitEpochOptions {
+                        table_id: request.table_id.into(),
+                    },
+                )
                 .instrument_await(format!("wait_epoch_commit (epoch {})", epoch))
                 .await
                 .map_err(StreamError::from)?;

src/meta/src/barrier/command.rs

@@ -22,7 +22,6 @@ use risingwave_common::hash::ActorMapping;
 use risingwave_common::types::Timestamptz;
 use risingwave_common::util::epoch::Epoch;
 use risingwave_connector::source::SplitImpl;
-use risingwave_hummock_sdk::HummockEpoch;
 use risingwave_pb::catalog::{CreateType, Table};
 use risingwave_pb::common::PbWorkerNode;
 use risingwave_pb::meta::table_fragments::PbActorStatus;
@@ -442,6 +441,8 @@ pub struct CommandContext {
     pub prev_epoch: TracedEpoch,
     pub curr_epoch: TracedEpoch,
 
+    pub table_ids_to_commit: HashSet<TableId>,
+
     pub current_paused_reason: Option<PausedReason>,
 
     pub command: Command,
@@ -470,12 +471,12 @@ impl std::fmt::Debug for CommandContext {
 }
 
 impl CommandContext {
-    #[allow(clippy::too_many_arguments)]
     pub(super) fn new(
         node_map: HashMap<WorkerId, PbWorkerNode>,
         subscription_info: InflightSubscriptionInfo,
         prev_epoch: TracedEpoch,
         curr_epoch: TracedEpoch,
+        table_ids_to_commit: HashSet<TableId>,
         current_paused_reason: Option<PausedReason>,
         command: Command,
         kind: BarrierKind,
@@ -487,6 +488,7 @@ impl CommandContext {
             subscription_info,
             prev_epoch,
             curr_epoch,
+            table_ids_to_commit,
             current_paused_reason,
             command,
             kind,
@@ -945,15 +947,24 @@ impl Command {
 }
 
 impl CommandContext {
-    pub async fn wait_epoch_commit(&self, epoch: HummockEpoch) -> MetaResult<()> {
+    pub async fn wait_epoch_commit(&self) -> MetaResult<()> {
+        let table_id = self.table_ids_to_commit.iter().next().cloned();
+        // try wait epoch on an existing random table id
+        let Some(table_id) = table_id else {
+            // no need to wait epoch when there is no table id
+            return Ok(());
+        };
         let futures = self.node_map.values().map(|worker_node| async {
             let client = self
                 .barrier_manager_context
                 .env
                 .stream_client_pool()
                 .get(worker_node)
                 .await?;
-            let request = WaitEpochCommitRequest { epoch };
+            let request = WaitEpochCommitRequest {
+                epoch: self.prev_epoch.value().0,
+                table_id: table_id.table_id,
+            };
             client.wait_epoch_commit(request).await
         });
 
@@ -976,7 +987,7 @@ impl CommandContext {
                     // storage version with this epoch is synced to all compute nodes before the
                     // execution of the next command of `Update`, as some newly created operators
                     // may immediately initialize their states on that barrier.
-                    self.wait_epoch_commit(self.prev_epoch.value().0).await?;
+                    self.wait_epoch_commit().await?;
                 }
             }
 

src/meta/src/barrier/mod.rs

@@ -1028,11 +1028,14 @@ impl GlobalBarrierManager {
         });
         span.record("epoch", curr_epoch.value().0);
 
+        let table_ids_to_commit: HashSet<_> = pre_applied_graph_info.existing_table_ids().collect();
+
         let command_ctx = Arc::new(CommandContext::new(
             self.active_streaming_nodes.current().clone(),
             pre_applied_subscription_info,
             prev_epoch.clone(),
             curr_epoch.clone(),
+            table_ids_to_commit.clone(),
             self.state.paused_reason(),
             command,
             kind,
@@ -1042,8 +1045,6 @@ impl GlobalBarrierManager {
 
         send_latency_timer.observe_duration();
 
-        let table_ids_to_commit: HashSet<_> = pre_applied_graph_info.existing_table_ids().collect();
-
         let mut jobs_to_wait = HashSet::new();
 
         for (table_id, creating_job) in &mut self.checkpoint_control.creating_streaming_job_controls

src/storage/hummock_sdk/src/version.rs

@@ -422,6 +422,13 @@ impl HummockVersion {
         self.max_committed_epoch
     }
 
+    pub fn table_committed_epoch(&self, table_id: TableId) -> Option<u64> {
+        self.state_table_info
+            .info()
+            .get(&table_id)
+            .map(|info| info.committed_epoch)
+    }
+
     pub fn visible_table_committed_epoch(&self) -> u64 {
         self.max_committed_epoch
     }

src/storage/hummock_test/src/bin/replay/replay_impl.rs

@@ -25,7 +25,7 @@ use risingwave_hummock_sdk::{HummockReadEpoch, HummockVersionId, SyncResult};
 use risingwave_hummock_trace::{
     GlobalReplay, LocalReplay, LocalReplayRead, ReplayItem, ReplayRead, ReplayStateStore,
     ReplayWrite, Result, TraceError, TracedBytes, TracedInitOptions, TracedNewLocalOptions,
-    TracedReadOptions, TracedSealCurrentEpochOptions, TracedSubResp,
+    TracedReadOptions, TracedSealCurrentEpochOptions, TracedSubResp, TracedTryWaitEpochOptions,
 };
 use risingwave_meta::manager::{MessageStatus, MetaSrvEnv, NotificationManagerRef, WorkerKey};
 use risingwave_pb::common::WorkerNode;
@@ -170,9 +170,13 @@ impl ReplayStateStore for GlobalReplayImpl {
         Box::new(LocalReplayImpl(local_storage))
     }
 
-    async fn try_wait_epoch(&self, epoch: HummockReadEpoch) -> Result<()> {
+    async fn try_wait_epoch(
+        &self,
+        epoch: HummockReadEpoch,
+        options: TracedTryWaitEpochOptions,
+    ) -> Result<()> {
         self.store
-            .try_wait_epoch(epoch)
+            .try_wait_epoch(epoch, options.into())
             .await
             .map_err(|_| TraceError::TryWaitEpochFailed)?;
         Ok(())

src/storage/hummock_test/src/failpoint_tests.rs

@@ -30,7 +30,8 @@ use risingwave_storage::hummock::test_utils::{count_stream, default_opts_for_tes
 use risingwave_storage::hummock::{CachePolicy, HummockStorage};
 use risingwave_storage::storage_value::StorageValue;
 use risingwave_storage::store::{
-    LocalStateStore, NewLocalOptions, PrefetchOptions, ReadOptions, StateStoreRead, WriteOptions,
+    LocalStateStore, NewLocalOptions, PrefetchOptions, ReadOptions, StateStoreRead,
+    TryWaitEpochOptions, WriteOptions,
 };
 use risingwave_storage::StateStore;
 
@@ -148,7 +149,10 @@ async fn test_failpoints_state_store_read_upload() {
         .unwrap();
     meta_client.commit_epoch(1, res, false).await.unwrap();
     hummock_storage
-        .try_wait_epoch(HummockReadEpoch::Committed(1))
+        .try_wait_epoch(
+            HummockReadEpoch::Committed(1),
+            TryWaitEpochOptions::for_test(local.table_id()),
+        )
         .await
         .unwrap();
     // clear block cache
@@ -225,7 +229,10 @@ async fn test_failpoints_state_store_read_upload() {
         .unwrap();
     meta_client.commit_epoch(3, res, false).await.unwrap();
     hummock_storage
-        .try_wait_epoch(HummockReadEpoch::Committed(3))
+        .try_wait_epoch(
+            HummockReadEpoch::Committed(3),
+            TryWaitEpochOptions::for_test(local.table_id()),
+        )
         .await
         .unwrap();
 

src/storage/hummock_test/src/hummock_storage_tests.rs

@@ -568,7 +568,7 @@ async fn test_state_store_sync() {
         .commit_epoch(epoch1, res, false)
         .await
         .unwrap();
-    test_env.storage.try_wait_epoch_for_test(epoch1).await;
+    test_env.wait_sync_committed_version().await;
     {
         // after sync 1 epoch
         let read_version = hummock_storage.read_version();
@@ -614,7 +614,7 @@ async fn test_state_store_sync() {
         .commit_epoch(epoch2, res, false)
         .await
         .unwrap();
-    test_env.storage.try_wait_epoch_for_test(epoch2).await;
+    test_env.wait_sync_committed_version().await;
     {
         // after sync all epoch
         let read_version = hummock_storage.read_version();
@@ -912,7 +912,7 @@ async fn test_delete_get() {
         .commit_epoch(epoch2, res, false)
         .await
         .unwrap();
-    test_env.storage.try_wait_epoch_for_test(epoch2).await;
+    test_env.wait_sync_committed_version().await;
     assert!(test_env
         .storage
         .get(
@@ -1112,7 +1112,7 @@ async fn test_multiple_epoch_sync() {
         .commit_epoch(epoch3, sync_result3, false)
         .await
         .unwrap();
-    test_env.storage.try_wait_epoch_for_test(epoch3).await;
+    test_env.wait_sync_committed_version().await;
     test_get(true).await;
 }
 
@@ -1286,7 +1286,7 @@ async fn test_iter_with_min_epoch() {
             .commit_epoch(epoch2, sync_result2, false)
             .await
             .unwrap();
-        test_env.storage.try_wait_epoch_for_test(epoch2).await;
+        test_env.wait_sync_committed_version().await;
 
         {
             let iter = test_env
@@ -1578,7 +1578,7 @@ async fn test_hummock_version_reader() {
                 .commit_epoch(epoch1, sync_result1, false)
                 .await
                 .unwrap();
-            test_env.storage.try_wait_epoch_for_test(epoch1).await;
+            test_env.wait_sync_committed_version().await;
 
             let sync_result2 = test_env
                 .storage
@@ -1590,7 +1590,7 @@ async fn test_hummock_version_reader() {
                 .commit_epoch(epoch2, sync_result2, false)
                 .await
                 .unwrap();
-            test_env.storage.try_wait_epoch_for_test(epoch2).await;
+            test_env.wait_sync_committed_version().await;
 
             let sync_result3 = test_env
                 .storage
@@ -1602,7 +1602,7 @@ async fn test_hummock_version_reader() {
                 .commit_epoch(epoch3, sync_result3, false)
                 .await
                 .unwrap();
-            test_env.storage.try_wait_epoch_for_test(epoch3).await;
+            test_env.wait_sync_committed_version().await;
             {
                 let (_, read_snapshot) = read_filter_for_version(
                     epoch1,
@@ -1996,7 +1996,7 @@ async fn test_get_with_min_epoch() {
         .await
         .unwrap();
 
-    test_env.storage.try_wait_epoch_for_test(epoch2).await;
+    test_env.wait_sync_committed_version().await;
     let k = gen_key(0);
     let prefix_hint = {
         let mut ret = Vec::with_capacity(TABLE_PREFIX_LEN + k.len());
@@ -2433,7 +2433,7 @@ async fn test_table_watermark() {
     };
 
     test_env.commit_epoch(epoch1).await;
-    test_env.storage.try_wait_epoch_for_test(epoch1).await;
+    test_env.wait_sync_committed_version().await;
 
     let (local1, local2) = test_after_epoch2(local1, local2).await;
 
@@ -2522,7 +2522,7 @@ async fn test_table_watermark() {
     let (local1, local2) = test_after_epoch2(local1, local2).await;
 
     test_env.commit_epoch(epoch2).await;
-    test_env.storage.try_wait_epoch_for_test(epoch2).await;
+    test_env.wait_sync_committed_version().await;
 
     test_global_read(test_env.storage.clone(), epoch2).await;
 
@@ -2557,7 +2557,7 @@ async fn test_table_watermark() {
     let (local1, local2) = test_after_epoch2(local1, local2).await;
 
     test_env.commit_epoch(epoch3).await;
-    test_env.storage.try_wait_epoch_for_test(epoch3).await;
+    test_env.wait_sync_committed_version().await;
 
     check_version_table_watermark(test_env.storage.get_pinned_version());
 

src/storage/hummock_test/src/snapshot_tests.rs

@@ -27,7 +27,7 @@ use risingwave_storage::hummock::{CachePolicy, HummockStorage};
 use risingwave_storage::storage_value::StorageValue;
 use risingwave_storage::store::{
     LocalStateStore, NewLocalOptions, PrefetchOptions, ReadOptions, SealCurrentEpochOptions,
-    StateStoreRead, WriteOptions,
+    StateStoreRead, TryWaitEpochOptions, WriteOptions,
 };
 use risingwave_storage::StateStore;
 
@@ -157,7 +157,10 @@ async fn test_snapshot_inner(
                 .await
                 .unwrap();
             hummock_storage
-                .try_wait_epoch(HummockReadEpoch::Committed(epoch1))
+                .try_wait_epoch(
+                    HummockReadEpoch::Committed(epoch1),
+                    TryWaitEpochOptions::for_test(Default::default()),
+                )
                 .await
                 .unwrap();
         }
@@ -201,7 +204,10 @@ async fn test_snapshot_inner(
                 .await
                 .unwrap();
             hummock_storage
-                .try_wait_epoch(HummockReadEpoch::Committed(epoch2))
+                .try_wait_epoch(
+                    HummockReadEpoch::Committed(epoch2),
+                    TryWaitEpochOptions::for_test(Default::default()),
+                )
                 .await
                 .unwrap();
         }
@@ -251,7 +257,10 @@ async fn test_snapshot_inner(
                 .await
                 .unwrap();
             hummock_storage
-                .try_wait_epoch(HummockReadEpoch::Committed(epoch3))
+                .try_wait_epoch(
+                    HummockReadEpoch::Committed(epoch3),
+                    TryWaitEpochOptions::for_test(Default::default()),
+                )
                 .await
                 .unwrap();
         }
@@ -334,7 +343,10 @@ async fn test_snapshot_range_scan_inner(
                 .await
                 .unwrap();
             hummock_storage
-                .try_wait_epoch(HummockReadEpoch::Committed(epoch))
+                .try_wait_epoch(
+                    HummockReadEpoch::Committed(epoch),
+                    TryWaitEpochOptions::for_test(local.table_id()),
+                )
                 .await
                 .unwrap();
         }