JingsongLi commented on code in PR #8401: URL: https://github.com/apache/paimon/pull/8401#discussion_r3534270073
########## paimon-flink/paimon-flink-common/src/main/java/org/apache/paimon/flink/sink/coordinator/CommittingWriteOperatorCoordinator.java: ########## @@ -0,0 +1,580 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one + * or more contributor license agreements. See the NOTICE file + * distributed with this work for additional information + * regarding copyright ownership. The ASF licenses this file + * to you 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. + */ + +package org.apache.paimon.flink.sink.coordinator; + +import org.apache.paimon.annotation.VisibleForTesting; +import org.apache.paimon.flink.sink.Committable; +import org.apache.paimon.flink.sink.Committer; +import org.apache.paimon.flink.sink.state.CoordinatorState; +import org.apache.paimon.flink.sink.state.CoordinatorStateSerializer; +import org.apache.paimon.flink.sink.state.MemoryBackendStateStore; +import org.apache.paimon.manifest.ManifestCommittable; + +import org.apache.flink.api.common.typeutils.base.ListSerializer; +import org.apache.flink.core.io.SimpleVersionedSerialization; +import org.apache.flink.runtime.jobgraph.OperatorID; +import org.apache.flink.runtime.operators.coordination.OperatorCoordinator; +import org.apache.flink.runtime.operators.coordination.OperatorEvent; +import org.apache.flink.runtime.operators.coordination.RecreateOnResetOperatorCoordinator; +import org.apache.flink.util.function.ThrowingRunnable; +import org.slf4j.Logger; +import org.slf4j.LoggerFactory; + +import java.io.IOException; +import java.util.ArrayList; +import java.util.Collections; +import java.util.List; +import java.util.Map; +import java.util.NavigableMap; +import java.util.TreeMap; +import java.util.concurrent.CompletableFuture; +import java.util.concurrent.ExecutorService; +import java.util.concurrent.Executors; + +import static org.apache.paimon.utils.Preconditions.checkState; + +/** + * {@link OperatorCoordinator} that runs the Paimon committer on the JobManager for the + * unaware-bucket append write path. Writers stream per-checkpoint committables to this coordinator + * over the {@link OperatorEvent} channel; on {@link #notifyCheckpointComplete} the coordinator + * aligns committables across subtasks and commits them from a dedicated single-thread executor, so + * the JM main thread is never blocked by table I/O. + * + * <p>Wrap this class with a {@link RecreateOnResetOperatorCoordinator} (see {@link Provider}). The + * wrapper discards this instance on global failover and creates a new one in its place, which keeps + * the lifecycle inside a single instance simple. See {@link #resetToCheckpoint} and {@link State} + * for the resulting state machine. + */ +public class CommittingWriteOperatorCoordinator implements OperatorCoordinator { + + private static final Logger LOG = + LoggerFactory.getLogger(CommittingWriteOperatorCoordinator.class); + + private final OperatorCoordinator.Context context; + private final Committer.Factory<Committable, ManifestCommittable> committerFactory; + private final boolean streamingCheckpointEnabled; + private final boolean failoverAfterRecovery; + private final int parallelism; + + private final WriterCommittables[] subtaskCommittables; + private final ListSerializer<Committable> committableSerializer; + private final CoordinatorStateSerializer stateSerializer; + private final ExecutorService commitExecutor; + private final SimpleWatermarkValve watermarkValve; + + // Populated by resetToCheckpoint and consumed by start. Plain fields are sufficient: both + // callbacks run on the same scheduler thread in order. + private long restoredCheckpointId = OperatorCoordinator.NO_CHECKPOINT; + private byte[] restoredCheckpointData; + + private State state; + private Committer<Committable, ManifestCommittable> committer; + private String commitUser; + private MemoryBackendStateStore stateStore; + + public CommittingWriteOperatorCoordinator( + OperatorCoordinator.Context context, + Committer.Factory<Committable, ManifestCommittable> committerFactory, + boolean streamingCheckpointEnabled, + String initialCommitUser, + boolean failoverAfterRecovery) { + this.context = context; + this.committerFactory = committerFactory; + this.streamingCheckpointEnabled = streamingCheckpointEnabled; + this.commitUser = initialCommitUser; + this.failoverAfterRecovery = failoverAfterRecovery; + this.parallelism = context.currentParallelism(); + this.subtaskCommittables = new WriterCommittables[parallelism]; + this.committableSerializer = CommittableEvent.createCommittableListSerializer(); + this.stateSerializer = new CoordinatorStateSerializer(); + this.commitExecutor = + Executors.newSingleThreadExecutor( + new CoordinatorExecutorThreadFactory("WriteCommitCoordinator", context)); + this.watermarkValve = new SimpleWatermarkValve(parallelism); + this.state = State.CREATED; + } + + @Override + public void start() throws Exception { + // Invoked at most once. If resetToCheckpoint ran first it already moved state to RESTORING + // and stashed the bytes; otherwise we are in CREATED and there is nothing to restore. + checkState( + state == State.CREATED || state == State.RESTORING, + "Coordinator already started, illegal state %s", + state); + runInEventLoop( + () -> { + if (state == State.RESTORING) { + restoreState(restoredCheckpointId, restoredCheckpointData); + // not needed after deserialization; release the reference + restoredCheckpointData = null; + initializeCommitter(true); + // stay in RESTORING until writers re-emit committables and align catches up + } else { + restoreState(OperatorCoordinator.NO_CHECKPOINT, null); + initializeCommitter(false); + transitionState(State.RUNNING); + } + }, + "starting"); + } + + @Override + public void close() throws Exception { + transitionState(State.CLOSED); + if (commitExecutor != null) { + commitExecutor.shutdownNow(); + } + if (committer != null) { + committer.close(); + committer = null; + } + } + + @Override + public void checkpointCoordinator(long checkpointId, CompletableFuture<byte[]> result) { + runCheckpointInEventLoop( + () -> { + if (state != State.RUNNING) { + // if checkpoint is executed before finishing restoring, just fail it + result.completeExceptionally( + new IllegalStateException( + "Checkpoint of commit coordinator should be taken in RUNNING state, while current state is " + + state)); + return; + } + committer.snapshotState(); + byte[] checkpointData = + SimpleVersionedSerialization.writeVersionAndSerialize( + stateSerializer, + new CoordinatorState( + commitUser, + watermarkValve.getCurrentWatermark(), + stateStore.getSerializedStates())); + result.complete(checkpointData); + }, + result, + "taking checkpoint %d", + checkpointId); + } + + @Override + public void handleEventFromOperator(int subtask, int attemptNumber, OperatorEvent event) { + runInEventLoop( + () -> { + if (event instanceof CommittableEvent) { + handleCommittableEvent(subtask, (CommittableEvent) event); + } else if (event instanceof WatermarkEvent) { + handleWatermarkEvent(subtask, (WatermarkEvent) event); + } else { + // TODO: end input handling + throw new UnsupportedOperationException("Unsupported event type: " + event); + } + }, + "handling operator event %s from subtask %d (#%d)", + event, + subtask, + attemptNumber); + } + + @Override + public void notifyCheckpointComplete(long checkpointId) { + runInEventLoop( + () -> { + if (state != State.RUNNING) { + throw new IllegalStateException( + "Completing checkpoint should be notified in RUNNING state, while current state is " + + state); + } + // writers always report a committable per (subtask, checkpoint) during + // snapshot, even if empty; missing means the writer is broken + if (!alignCommittables(checkpointId)) { + throw new IllegalStateException("Not all committables reported by writer"); + } + commitUpToCheckpoint( + checkpointId, + pollManifestCommittablesForCheckpoint( + checkpointId, + subtaskCommittables, + committer, + watermarkValve.getCurrentWatermark()), Review Comment: This still does not preserve the old checkpoint watermark semantics. The previous `CommitterOperator` calls `pollInputs()` from `snapshotState`, so each checkpoint gets a `ManifestCommittable` with the watermark as of that checkpoint barrier; later watermarks processed before `notifyCheckpointComplete` do not change that checkpoint. Here the coordinator reads `watermarkValve.getCurrentWatermark()` only when the checkpoint completes, so a watermark event processed after checkpoint N was snapshotted but before checkpoint N completes can be written into snapshot N. That can mark partitions done too early. Could we carry the writer/coordinator watermark snapshot with the checkpoint committable event (or otherwise store a per-checkpoint watermark) and add a regression where a later watermark arrives before `notifyCheckpointComplete(N)`? -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
