[server] Fix ReplicaFetcherThread keeps throwing OutOfOrderSequenceException because of writer id expire

Author: LiebingYu

fluss-server/src/main/java/org/apache/fluss/server/log/LogTablet.java

@@ -643,7 +643,7 @@ private LogAppendInfo append(MemoryLogRecords records, boolean appendAsLeader)
             // now that we have valid records, offsets assigned, we need to validate the idempotent
             // state of the writers and collect some metadata.
             Either<WriterStateEntry.BatchMetadata, Collection<WriterAppendInfo>> validateResult =
-                    analyzeAndValidateWriterState(validRecords);
+                    analyzeAndValidateWriterState(validRecords, appendAsLeader);
 
             if (validateResult.isLeft()) {
                 // have duplicated batch metadata, skip the append and update append info.
@@ -1006,7 +1006,7 @@ private LogAppendInfo analyzeAndValidateRecords(MemoryLogRecords records) {
 
     /** Returns either the duplicated batch metadata (left) or the updated writers (right). */
     private Either<WriterStateEntry.BatchMetadata, Collection<WriterAppendInfo>>
-            analyzeAndValidateWriterState(MemoryLogRecords records) {
+            analyzeAndValidateWriterState(MemoryLogRecords records, boolean isAppendAsLeader) {
         Map<Long, WriterAppendInfo> updatedWriters = new HashMap<>();
 
         for (LogRecordBatch batch : records.batches()) {
@@ -1023,14 +1023,15 @@ private LogAppendInfo analyzeAndValidateRecords(MemoryLogRecords records) {
                 }
 
                 // update write append info.
-                updateWriterAppendInfo(writerStateManager, batch, updatedWriters);
+                updateWriterAppendInfo(writerStateManager, batch, updatedWriters, isAppendAsLeader);
             }
         }
 
         return Either.right(updatedWriters.values());
     }
 
-    void removeExpiredWriter(long currentTimeMs) {
+    @VisibleForTesting
+    public void removeExpiredWriter(long currentTimeMs) {
         synchronized (lock) {
             writerStateManager.removeExpiredWriters(currentTimeMs);
         }
@@ -1110,14 +1111,16 @@ private void deleteSegments(List<LogSegment> deletableSegments, SegmentDeletionR
     private static void updateWriterAppendInfo(
             WriterStateManager writerStateManager,
             LogRecordBatch batch,
-            Map<Long, WriterAppendInfo> writers) {
+            Map<Long, WriterAppendInfo> writers,
+            boolean isAppendAsLeader) {
         long writerId = batch.writerId();
         // update writers.
         WriterAppendInfo appendInfo =
                 writers.computeIfAbsent(writerId, id -> writerStateManager.prepareUpdate(writerId));
         appendInfo.append(
                 batch,
-                writerStateManager.isWriterInBatchExpired(System.currentTimeMillis(), batch));
+                writerStateManager.isWriterInBatchExpired(System.currentTimeMillis(), batch),
+                isAppendAsLeader);
     }
 
     static void rebuildWriterState(
@@ -1230,7 +1233,7 @@ private static void loadWritersFromRecords(
         Map<Long, WriterAppendInfo> loadedWriters = new HashMap<>();
         for (LogRecordBatch batch : records.batches()) {
             if (batch.hasWriterId()) {
-                updateWriterAppendInfo(writerStateManager, batch, loadedWriters);
+                updateWriterAppendInfo(writerStateManager, batch, loadedWriters, true);
             }
         }
         loadedWriters.values().forEach(writerStateManager::update);

fluss-server/src/main/java/org/apache/fluss/server/log/WriterAppendInfo.java

@@ -44,13 +44,15 @@ public long writerId() {
         return writerId;
     }
 
-    public void append(LogRecordBatch batch, boolean isWriterInBatchExpired) {
+    public void append(
+            LogRecordBatch batch, boolean isWriterInBatchExpired, boolean isAppendAsLeader) {
         LogOffsetMetadata firstOffsetMetadata = new LogOffsetMetadata(batch.baseLogOffset());
         appendDataBatch(
                 batch.batchSequence(),
                 firstOffsetMetadata,
                 batch.lastLogOffset(),
                 isWriterInBatchExpired,
+                isAppendAsLeader,
                 batch.commitTimestamp());
     }
 
@@ -59,8 +61,9 @@ public void appendDataBatch(
             LogOffsetMetadata firstOffsetMetadata,
             long lastOffset,
             boolean isWriterInBatchExpired,
+            boolean isAppendAsLeader,
             long batchTimestamp) {
-        maybeValidateDataBatch(batchSequence, isWriterInBatchExpired, lastOffset);
+        maybeValidateDataBatch(batchSequence, isWriterInBatchExpired, lastOffset, isAppendAsLeader);
         updatedEntry.addBath(
                 batchSequence,
                 lastOffset,
@@ -69,13 +72,16 @@ public void appendDataBatch(
     }
 
     private void maybeValidateDataBatch(
-            int appendFirstSeq, boolean isWriterInBatchExpired, long lastOffset) {
+            int appendFirstSeq,
+            boolean isWriterInBatchExpired,
+            long lastOffset,
+            boolean isAppendAsLeader) {
         int currentLastSeq =
                 !updatedEntry.isEmpty()
                         ? updatedEntry.lastBatchSequence()
                         : currentEntry.lastBatchSequence();
         // must be in sequence, even for the first batch should start from 0
-        if (!inSequence(currentLastSeq, appendFirstSeq, isWriterInBatchExpired)) {
+        if (!inSequence(currentLastSeq, appendFirstSeq, isWriterInBatchExpired, isAppendAsLeader)) {
             throw new OutOfOrderSequenceException(
                     String.format(
                             "Out of order batch sequence for writer %s at offset %s in "
@@ -93,16 +99,52 @@ public WriterStateEntry toEntry() {
      * three scenarios will be judged as in sequence:
      *
      * <ul>
-     *   <li>If lastBatchSeq equals NO_BATCH_SEQUENCE, we need to check whether the committed
-     *       timestamp of the next batch under the current writerId has expired. If it has expired,
-     *       we consider this a special case caused by writerId expiration, for this case, to ensure
-     *       the correctness of follower sync, we still treat it as in sequence.
+     *   <li>If lastBatchSeq equals NO_BATCH_SEQUENCE, the following two scenarios will be judged as
+     *       in sequence:
+     *       <ul>
+     *         <li>If the committed timestamp of the next batch under the current writerId has
+     *             expired, we consider this a special case caused by writerId expiration, for this
+     *             case, to ensure the correctness of follower sync, we still treat it as in
+     *             sequence.
+     *         <li>If the append request is from the follower, we consider this is a special case
+     *             caused by inconsistent expiration of writerId between the leader and follower. To
+     *             prevent continuous fetch failures on the follower side, we still treat it as in
+     *             sequence. Here is a detailed example: The expiration of a writer is triggered
+     *             asynchronously by the {@code PeriodicWriterIdExpirationCheck} thread at intervals
+     *             defined by {@code server.writer-id.expiration-check-interval}, which can result
+     *             in slight differences in the actual expiration times of the same writer on the
+     *             leader replica and follower replicas. This slight difference leads to a dreadful
+     *             corner case. Imagine the following scenario(set {@code
+     *             server.writer-id.expiration-check-interval}: 10min, {@code
+     *             server.writer-id.expiration-time}: 12h):
+     *             <pre>{@code
+     * Step     Time         Action of Leader                  Action of Follower
+     * 1        00:03:38     receive batch 0 of writer 101
+     * 2        00:03:38                                       fetch batch 0 of writer 101
+     * 3        12:05:00                                       remove state of writer 101
+     * 4        12:10:02     receive batch 1 of writer 101
+     * 5        12:10:02                                       fetch batch 0 of writer 101
+     * 6        12:11:00     remove state of writer 101
+     * }</pre>
+     *             In step 3, the follower removes the state of writer 101 first, since it has been
+     *             more than 12 hours since writer 101's last batch write, making it safe to remove.
+     *             However, since the expiration of writer 101 has not yet occurred on the leader,
+     *             and a new batch 1 is received at this time, it is successfully written on the
+     *             leader. At this point, the fetcher pulls batch 1 from the leader, but since the
+     *             state of writer 101 has already been cleaned up, an {@link
+     *             OutOfOrderSequenceException} will occur during to write if we don't treat it as
+     *             in sequence.
+     *       </ul>
      *   <li>nextBatchSeq == lastBatchSeq + 1L
      *   <li>lastBatchSeq reaches its maximum value
      * </ul>
      */
-    private boolean inSequence(int lastBatchSeq, int nextBatchSeq, boolean isWriterInBatchExpired) {
-        return (lastBatchSeq == NO_BATCH_SEQUENCE && isWriterInBatchExpired)
+    private boolean inSequence(
+            int lastBatchSeq,
+            int nextBatchSeq,
+            boolean isWriterInBatchExpired,
+            boolean isAppendAsLeader) {
+        return (lastBatchSeq == NO_BATCH_SEQUENCE && (isWriterInBatchExpired || !isAppendAsLeader))
                 || nextBatchSeq == lastBatchSeq + 1L
                 || (nextBatchSeq == 0 && lastBatchSeq == Integer.MAX_VALUE);
     }

fluss-server/src/test/java/org/apache/fluss/server/log/WriterStateManagerTest.java

@@ -126,14 +126,14 @@ void testValidationOnFirstEntryWhenLoadingLog() {
     void testPrepareUpdateDoesNotMutate() {
         WriterAppendInfo appendInfo = stateManager.prepareUpdate(writerId);
         appendInfo.appendDataBatch(
-                0, new LogOffsetMetadata(15L), 20L, false, System.currentTimeMillis());
+                0, new LogOffsetMetadata(15L), 20L, false, false, System.currentTimeMillis());
         assertThat(stateManager.lastEntry(writerId)).isNotPresent();
         stateManager.update(appendInfo);
         assertThat(stateManager.lastEntry(writerId)).isPresent();
 
         WriterAppendInfo nextAppendInfo = stateManager.prepareUpdate(writerId);
         nextAppendInfo.appendDataBatch(
-                1, new LogOffsetMetadata(26L), 30L, false, System.currentTimeMillis());
+                1, new LogOffsetMetadata(26L), 30L, false, false, System.currentTimeMillis());
         assertThat(stateManager.lastEntry(writerId)).isPresent();
 
         WriterStateEntry lastEntry = stateManager.lastEntry(writerId).get();
@@ -521,6 +521,7 @@ private void append(
                 new LogOffsetMetadata(offset),
                 offset,
                 isWriterInBatchExpired,
+                true,
                 lastTimestamp);
         stateManager.update(appendInfo);
         stateManager.updateMapEndOffset(offset + 1);

fluss-server/src/test/java/org/apache/fluss/server/replica/ReplicaTest.java

@@ -18,6 +18,7 @@
 package org.apache.fluss.server.replica;
 
 import org.apache.fluss.config.ConfigOptions;
+import org.apache.fluss.exception.OutOfOrderSequenceException;
 import org.apache.fluss.metadata.LogFormat;
 import org.apache.fluss.metadata.PhysicalTablePath;
 import org.apache.fluss.metadata.TableBucket;
@@ -48,6 +49,7 @@
 import java.io.File;
 import java.io.IOException;
 import java.nio.file.Path;
+import java.time.Duration;
 import java.util.ArrayList;
 import java.util.Arrays;
 import java.util.Collections;
@@ -75,10 +77,12 @@
 import static org.apache.fluss.testutils.DataTestUtils.genKvRecordBatch;
 import static org.apache.fluss.testutils.DataTestUtils.genKvRecords;
 import static org.apache.fluss.testutils.DataTestUtils.genMemoryLogRecordsByObject;
+import static org.apache.fluss.testutils.DataTestUtils.genMemoryLogRecordsWithWriterId;
 import static org.apache.fluss.testutils.DataTestUtils.getKeyValuePairs;
 import static org.apache.fluss.testutils.LogRecordsAssert.assertThatLogRecords;
 import static org.apache.fluss.utils.Preconditions.checkNotNull;
 import static org.assertj.core.api.Assertions.assertThat;
+import static org.assertj.core.api.Assertions.assertThatThrownBy;
 
 /** Test for {@link Replica}. */
 final class ReplicaTest extends ReplicaTestBase {
@@ -129,6 +133,40 @@ void testAppendRecordsToLeader() throws Exception {
         assertLogRecordsEquals(DATA1_ROW_TYPE, logReadInfo.getFetchedData().getRecords(), DATA1);
     }
 
+    @Test
+    void testAppendRecordsWithOutOfOrderBatchSequence() throws Exception {
+        Replica logReplica =
+                makeLogReplica(DATA1_PHYSICAL_TABLE_PATH, new TableBucket(DATA1_TABLE_ID, 1));
+        makeLogReplicaAsLeader(logReplica);
+
+        long writerId = 101L;
+
+        // 1. append a batch with batchSequence = 0
+        logReplica.appendRecordsToLeader(genMemoryLogRecordsWithWriterId(DATA1, writerId, 0, 0), 0);
+
+        // manual advance time and remove expired writer, the state of writer 101 will be removed
+        manualClock.advanceTime(Duration.ofHours(12));
+        manualClock.advanceTime(Duration.ofSeconds(1));
+        assertThat(logReplica.getLogTablet().writerStateManager().activeWriters().size())
+                .isEqualTo(1);
+        logReplica.getLogTablet().removeExpiredWriter(manualClock.milliseconds());
+        assertThat(logReplica.getLogTablet().writerStateManager().activeWriters().size())
+                .isEqualTo(0);
+
+        // 2. try to append an out of ordered batch as leader, will throw
+        // OutOfOrderSequenceException
+        assertThatThrownBy(
+                        () ->
+                                logReplica.appendRecordsToLeader(
+                                        genMemoryLogRecordsWithWriterId(DATA1, writerId, 2, 10), 0))
+                .isInstanceOf(OutOfOrderSequenceException.class);
+        assertThat(logReplica.getLocalLogEndOffset()).isEqualTo(10);
+
+        // 3. try to append an out of ordered batch as follower
+        logReplica.appendRecordsToFollower(genMemoryLogRecordsWithWriterId(DATA1, writerId, 2, 10));
+        assertThat(logReplica.getLocalLogEndOffset()).isEqualTo(20);
+    }
+
     @Test
     void testPartialPutRecordsToLeader() throws Exception {
         Replica kvReplica =

fluss-server/src/test/java/org/apache/fluss/server/replica/ReplicaTestBase.java

@@ -174,6 +174,8 @@ public void setup() throws Exception {
         conf.set(ConfigOptions.CLIENT_WRITER_BUFFER_PAGE_SIZE, MemorySize.parse("512b"));
         conf.set(ConfigOptions.CLIENT_WRITER_BATCH_SIZE, MemorySize.parse("1kb"));
 
+        conf.set(ConfigOptions.WRITER_ID_EXPIRATION_TIME, Duration.ofHours(12));
+
         scheduler = new FlussScheduler(2);
         scheduler.startup();