Harden periodically check to avoid endless flush loop (#29125)

In #28350, we fixed an endless flushing loop which may happen on
replicas by tightening the relation between the flush action and the
periodically flush condition.

1. The periodically flush condition is enabled only if it is disabled
after a flush.

2. If the periodically flush condition is enabled then a flush will
actually happen regardless of Lucene state.

(1) and (2) guarantee that a flushing loop will be terminated. Sadly,
the condition 1 can be violated in edge cases as we used two different
algorithms to evaluate the current and future uncommitted translog size.

- We use method `uncommittedSizeInBytes` to calculate current
  uncommitted size. It is the sum of translogs whose generation at least
the minGen (determined by a given seqno). We pick a continuous range of
translogs since the minGen to evaluate the current uncommitted size.

- We use method `sizeOfGensAboveSeqNoInBytes` to calculate the future
  uncommitted size. It is the sum of translogs whose maxSeqNo at least
the given seqNo. Here we don't pick a range but select translog one by
one.

Suppose we have 3 translogs `gen1={#1,#2}, gen2={}, gen3={#3} and
seqno=#1`, `uncommittedSizeInBytes` is the sum of gen1, gen2, and gen3
while `sizeOfGensAboveSeqNoInBytes` is the sum of gen1 and gen3. Gen2 is
excluded because its maxSeqno is still -1.

This commit removes both `sizeOfGensAboveSeqNoInBytes` and
`uncommittedSizeInBytes` methods, then enforces an engine to use only
`sizeInBytesByMinGen` method to evaluate the periodically flush condition.

Closes #29097
Relates ##28350

Author: dnhatn

server/src/main/java/org/elasticsearch/index/engine/InternalEngine.java

@@ -1479,7 +1479,8 @@ final boolean tryRenewSyncCommit() {
             ensureOpen();
             ensureCanFlush();
             String syncId = lastCommittedSegmentInfos.getUserData().get(SYNC_COMMIT_ID);
-            if (syncId != null && translog.uncommittedOperations() == 0 && indexWriter.hasUncommittedChanges()) {
+            long translogGenOfLastCommit = Long.parseLong(lastCommittedSegmentInfos.userData.get(Translog.TRANSLOG_GENERATION_KEY));
+            if (syncId != null && indexWriter.hasUncommittedChanges() && translog.totalOperationsByMinGen(translogGenOfLastCommit) == 0) {
                 logger.trace("start renewing sync commit [{}]", syncId);
                 commitIndexWriter(indexWriter, translog, syncId);
                 logger.debug("successfully sync committed. sync id [{}].", syncId);
@@ -1501,26 +1502,30 @@ final boolean tryRenewSyncCommit() {
     @Override
     public boolean shouldPeriodicallyFlush() {
         ensureOpen();
+        final long translogGenerationOfLastCommit = Long.parseLong(lastCommittedSegmentInfos.userData.get(Translog.TRANSLOG_GENERATION_KEY));
         final long flushThreshold = config().getIndexSettings().getFlushThresholdSize().getBytes();
-        final long uncommittedSizeOfCurrentCommit = translog.uncommittedSizeInBytes();
-        if (uncommittedSizeOfCurrentCommit < flushThreshold) {
+        if (translog.sizeInBytesByMinGen(translogGenerationOfLastCommit) < flushThreshold) {
             return false;
         }
         /*
-         * We should only flush ony if the shouldFlush condition can become false after flushing.
-         * This condition will change if the `uncommittedSize` of the new commit is smaller than
-         * the `uncommittedSize` of the current commit. This method is to maintain translog only,
-         * thus the IndexWriter#hasUncommittedChanges condition is not considered.
-         */
-        final long uncommittedSizeOfNewCommit = translog.sizeOfGensAboveSeqNoInBytes(localCheckpointTracker.getCheckpoint() + 1);
-        /*
-         * If flushThreshold is too small, we may repeatedly flush even there is no uncommitted operation
-         * as #sizeOfGensAboveSeqNoInByte and #uncommittedSizeInBytes can return different values.
-         * An empty translog file has non-zero `uncommittedSize` (the translog header), and method #sizeOfGensAboveSeqNoInBytes can
-         * return 0 now(no translog gen contains ops above local checkpoint) but method #uncommittedSizeInBytes will return an actual
-         * non-zero value after rolling a new translog generation. This can be avoided by checking the actual uncommitted operations.
+         * We flush to reduce the size of uncommitted translog but strictly speaking the uncommitted size won't always be
+         * below the flush-threshold after a flush. To avoid getting into an endless loop of flushing, we only enable the
+         * periodically flush condition if this condition is disabled after a flush. The condition will change if the new
+         * commit points to the later generation the last commit's(eg. gen-of-last-commit < gen-of-new-commit)[1].
+         *
+         * When the local checkpoint equals to max_seqno, and translog-gen of the last commit equals to translog-gen of
+         * the new commit, we know that the last generation must contain operations because its size is above the flush
+         * threshold and the flush-threshold is guaranteed to be higher than an empty translog by the setting validation.
+         * This guarantees that the new commit will point to the newly rolled generation. In fact, this scenario only
+         * happens when the generation-threshold is close to or above the flush-threshold; otherwise we have rolled
+         * generations as the generation-threshold was reached, then the first condition (eg. [1]) is already satisfied.
+         *
+         * This method is to maintain translog only, thus IndexWriter#hasUncommittedChanges condition is not considered.
          */
-        return uncommittedSizeOfNewCommit < uncommittedSizeOfCurrentCommit && translog.uncommittedOperations() > 0;
+        final long translogGenerationOfNewCommit =
+            translog.getMinGenerationForSeqNo(localCheckpointTracker.getCheckpoint() + 1).translogFileGeneration;
+        return translogGenerationOfLastCommit < translogGenerationOfNewCommit
+            || localCheckpointTracker.getCheckpoint() == localCheckpointTracker.getMaxSeqNo();
     }
 
     @Override

server/src/main/java/org/elasticsearch/index/translog/Translog.java

@@ -106,6 +106,7 @@ public class Translog extends AbstractIndexShardComponent implements IndexShardC
     public static final String TRANSLOG_FILE_SUFFIX = ".tlog";
     public static final String CHECKPOINT_SUFFIX = ".ckp";
     public static final String CHECKPOINT_FILE_NAME = "translog" + CHECKPOINT_SUFFIX;
+    public static final int DEFAULT_HEADER_SIZE_IN_BYTES = TranslogWriter.getHeaderLength(UUIDs.randomBase64UUID());
 
     static final Pattern PARSE_STRICT_ID_PATTERN = Pattern.compile("^" + TRANSLOG_FILE_PREFIX + "(\\d+)(\\.tlog)$");
 
@@ -372,26 +373,11 @@ public long getMinFileGeneration() {
         }
     }
 
-
-    /**
-     * Returns the number of operations in the translog files that aren't committed to lucene.
-     */
-    public int uncommittedOperations() {
-        return totalOperations(deletionPolicy.getTranslogGenerationOfLastCommit());
-    }
-
-    /**
-     * Returns the size in bytes of the translog files that aren't committed to lucene.
-     */
-    public long uncommittedSizeInBytes() {
-        return sizeInBytesByMinGen(deletionPolicy.getTranslogGenerationOfLastCommit());
-    }
-
     /**
      * Returns the number of operations in the translog files
      */
     public int totalOperations() {
-        return totalOperations(-1);
+        return totalOperationsByMinGen(-1);
     }
 
     /**
@@ -402,9 +388,9 @@ public long sizeInBytes() {
     }
 
     /**
-     * Returns the number of operations in the transaction files that aren't committed to lucene..
+     * Returns the number of operations in the translog files at least the given generation
      */
-    private int totalOperations(long minGeneration) {
+    public int totalOperationsByMinGen(long minGeneration) {
         try (ReleasableLock ignored = readLock.acquire()) {
             ensureOpen();
             return Stream.concat(readers.stream(), Stream.of(current))
@@ -425,9 +411,9 @@ public int estimateTotalOperationsFromMinSeq(long minSeqNo) {
     }
 
     /**
-     * Returns the size in bytes of the translog files above the given generation
+     * Returns the size in bytes of the translog files at least the given generation
      */
-    private long sizeInBytesByMinGen(long minGeneration) {
+    public long sizeInBytesByMinGen(long minGeneration) {
         try (ReleasableLock ignored = readLock.acquire()) {
             ensureOpen();
             return Stream.concat(readers.stream(), Stream.of(current))
@@ -437,16 +423,6 @@ private long sizeInBytesByMinGen(long minGeneration) {
         }
     }
 
-    /**
-     * Returns the size in bytes of the translog files with ops above the given seqNo
-     */
-    public long sizeOfGensAboveSeqNoInBytes(long minSeqNo) {
-        try (ReleasableLock ignored = readLock.acquire()) {
-            ensureOpen();
-            return readersAboveMinSeqNo(minSeqNo).mapToLong(BaseTranslogReader::sizeInBytes).sum();
-        }
-    }
-
     /**
      * Creates a new translog for the specified generation.
      *
@@ -751,7 +727,8 @@ private void closeOnTragicEvent(Exception ex) {
     public TranslogStats stats() {
         // acquire lock to make the two numbers roughly consistent (no file change half way)
         try (ReleasableLock lock = readLock.acquire()) {
-            return new TranslogStats(totalOperations(), sizeInBytes(), uncommittedOperations(), uncommittedSizeInBytes());
+            final long uncommittedGen = deletionPolicy.getTranslogGenerationOfLastCommit();
+            return new TranslogStats(totalOperations(), sizeInBytes(), totalOperationsByMinGen(uncommittedGen), sizeInBytesByMinGen(uncommittedGen));
         }
     }
 
@@ -1508,7 +1485,7 @@ public static void writeOperationNoSize(BufferedChecksumStreamOutput out, Transl
      * @return the minimum generation for the sequence number
      */
     public TranslogGeneration getMinGenerationForSeqNo(final long seqNo) {
-        try (ReleasableLock ignored = writeLock.acquire()) {
+        try (ReleasableLock ignored = readLock.acquire()) {
             /*
              * When flushing, the engine will ask the translog for the minimum generation that could contain any sequence number after the
              * local checkpoint. Immediately after flushing, there will be no such generation, so this minimum generation in this case will

server/src/main/java/org/elasticsearch/index/translog/TranslogDeletionPolicy.java

@@ -211,7 +211,6 @@ public synchronized long getMinTranslogGenerationForRecovery() {
 
     /**
      * Returns a translog generation that will be used to calculate the number of uncommitted operations since the last index commit.
-     * See {@link Translog#uncommittedOperations()} and {@link Translog#uncommittedSizeInBytes()}
      */
     public synchronized long getTranslogGenerationOfLastCommit() {
         return translogGenerationOfLastCommit;

server/src/test/java/org/elasticsearch/index/engine/InternalEngineTests.java

@@ -723,14 +723,13 @@ public void testTranslogRecoveryDoesNotReplayIntoTranslog() throws IOException {
             recoveringEngine = new InternalEngine(copy(initialEngine.config(), EngineConfig.OpenMode.OPEN_INDEX_AND_TRANSLOG)) {
                 @Override
                 public CommitId flush(boolean force, boolean waitIfOngoing) throws EngineException {
-                    assertThat(getTranslog().uncommittedOperations(), equalTo(docs));
+                    assertThat(getTranslog().stats().getUncommittedOperations(), equalTo(docs));
                     final CommitId commitId = super.flush(force, waitIfOngoing);
                     flushed.set(true);
                     return commitId;
                 }
             };
-
-            assertThat(recoveringEngine.getTranslog().uncommittedOperations(), equalTo(docs));
+            assertThat(recoveringEngine.getTranslog().stats().getUncommittedOperations(), equalTo(docs));
             recoveringEngine.recoverFromTranslog();
             assertTrue(flushed.get());
         } finally {
@@ -2884,7 +2883,7 @@ public void testCurrentTranslogIDisCommitted() throws IOException {
                     assertEquals(engine.getTranslog().getTranslogUUID(), userData.get(Translog.TRANSLOG_UUID_KEY));
                     expectThrows(IllegalStateException.class, () -> engine.recoverFromTranslog());
                     assertEquals(1, engine.getTranslog().currentFileGeneration());
-                    assertEquals(0L, engine.getTranslog().uncommittedOperations());
+                    assertEquals(0L, engine.getTranslog().stats().getUncommittedOperations());
                 }
             }
 
@@ -3840,7 +3839,7 @@ protected long doGenerateSeqNoForOperation(Operation operation) {
                             System.nanoTime(),
                             reason));
             assertThat(noOpEngine.getLocalCheckpointTracker().getCheckpoint(), equalTo((long) (maxSeqNo + 1)));
-            assertThat(noOpEngine.getTranslog().uncommittedOperations(), equalTo(1 + gapsFilled));
+            assertThat(noOpEngine.getTranslog().stats().getUncommittedOperations(), equalTo(1 + gapsFilled));
             // skip to the op that we added to the translog
             Translog.Operation op;
             Translog.Operation last = null;
@@ -4041,7 +4040,7 @@ public void testFillUpSequenceIdGapsOnRecovery() throws IOException {
             assertEquals(checkpointOnReplica, replicaEngine.getLocalCheckpointTracker().getCheckpoint());
             recoveringEngine = new InternalEngine(copy(
                 replicaEngine.config(), EngineConfig.OpenMode.OPEN_INDEX_AND_TRANSLOG, globalCheckpoint::get));
-            assertEquals(numDocsOnReplica, recoveringEngine.getTranslog().uncommittedOperations());
+            assertEquals(numDocsOnReplica, recoveringEngine.getTranslog().stats().getUncommittedOperations());
             recoveringEngine.recoverFromTranslog();
             assertEquals(maxSeqIDOnReplica, recoveringEngine.getLocalCheckpointTracker().getMaxSeqNo());
             assertEquals(checkpointOnReplica, recoveringEngine.getLocalCheckpointTracker().getCheckpoint());
@@ -4076,7 +4075,7 @@ public void testFillUpSequenceIdGapsOnRecovery() throws IOException {
             recoveringEngine = new InternalEngine(
                 copy(replicaEngine.config(), EngineConfig.OpenMode.OPEN_INDEX_AND_TRANSLOG, globalCheckpoint::get));
             if (flushed) {
-                assertEquals(0, recoveringEngine.getTranslog().uncommittedOperations());
+                assertThat(recoveringEngine.getTranslog().stats().getUncommittedOperations(), equalTo(0));
             }
             recoveringEngine.recoverFromTranslog();
             assertEquals(maxSeqIDOnReplica, recoveringEngine.getLocalCheckpointTracker().getMaxSeqNo());
@@ -4451,31 +4450,71 @@ public void testShouldPeriodicallyFlush() throws Exception {
             engine.index(indexForDoc(doc));
         }
         assertThat("Not exceeded translog flush threshold yet", engine.shouldPeriodicallyFlush(), equalTo(false));
-        long flushThreshold = RandomNumbers.randomLongBetween(random(), 100, engine.getTranslog().uncommittedSizeInBytes());
+        long flushThreshold = RandomNumbers.randomLongBetween(random(), 100, engine.getTranslog().stats().getUncommittedSizeInBytes());
         final IndexSettings indexSettings = engine.config().getIndexSettings();
         final IndexMetaData indexMetaData = IndexMetaData.builder(indexSettings.getIndexMetaData())
             .settings(Settings.builder().put(indexSettings.getSettings())
                 .put(IndexSettings.INDEX_TRANSLOG_FLUSH_THRESHOLD_SIZE_SETTING.getKey(), flushThreshold + "b")).build();
         indexSettings.updateIndexMetaData(indexMetaData);
         engine.onSettingsChanged();
-        assertThat(engine.getTranslog().uncommittedOperations(), equalTo(numDocs));
+        assertThat(engine.getTranslog().stats().getUncommittedOperations(), equalTo(numDocs));
         assertThat(engine.shouldPeriodicallyFlush(), equalTo(true));
         engine.flush();
-        assertThat(engine.getTranslog().uncommittedOperations(), equalTo(0));
+        assertThat(engine.getTranslog().stats().getUncommittedOperations(), equalTo(0));
         // Stale operations skipped by Lucene but added to translog - still able to flush
         for (int id = 0; id < numDocs; id++) {
             final ParsedDocument doc = testParsedDocument(Integer.toString(id), null, testDocumentWithTextField(), SOURCE, null);
             final Engine.IndexResult result = engine.index(replicaIndexForDoc(doc, 1L, id, false));
             assertThat(result.isCreated(), equalTo(false));
         }
         SegmentInfos lastCommitInfo = engine.getLastCommittedSegmentInfos();
-        assertThat(engine.getTranslog().uncommittedOperations(), equalTo(numDocs));
+        assertThat(engine.getTranslog().stats().getUncommittedOperations(), equalTo(numDocs));
         assertThat(engine.shouldPeriodicallyFlush(), equalTo(true));
         engine.flush(false, false);
         assertThat(engine.getLastCommittedSegmentInfos(), not(sameInstance(lastCommitInfo)));
-        assertThat(engine.getTranslog().uncommittedOperations(), equalTo(0));
+        assertThat(engine.getTranslog().stats().getUncommittedOperations(), equalTo(0));
+        // If the new index commit still points to the same translog generation as the current index commit,
+        // we should not enable the periodically flush condition; otherwise we can get into an infinite loop of flushes.
+        engine.getLocalCheckpointTracker().generateSeqNo(); // create a gap here
+        for (int id = 0; id < numDocs; id++) {
+            if (randomBoolean()) {
+                engine.getTranslog().rollGeneration();
+            }
+            final ParsedDocument doc = testParsedDocument("new" + id, null, testDocumentWithTextField(), SOURCE, null);
+            engine.index(replicaIndexForDoc(doc, 2L, engine.getLocalCheckpointTracker().generateSeqNo(), false));
+            if (engine.shouldPeriodicallyFlush()) {
+                engine.flush();
+                assertThat(engine.getLastCommittedSegmentInfos(), not(sameInstance(lastCommitInfo)));
+                assertThat(engine.shouldPeriodicallyFlush(), equalTo(false));
+            }
+        }
     }
 
+    public void testStressShouldPeriodicallyFlush() throws Exception {
+        final long flushThreshold = randomLongBetween(100, 5000);
+        final long generationThreshold = randomLongBetween(1000, 5000);
+        final IndexSettings indexSettings = engine.config().getIndexSettings();
+        final IndexMetaData indexMetaData = IndexMetaData.builder(indexSettings.getIndexMetaData())
+            .settings(Settings.builder().put(indexSettings.getSettings())
+                .put(IndexSettings.INDEX_TRANSLOG_GENERATION_THRESHOLD_SIZE_SETTING.getKey(), generationThreshold + "b")
+                .put(IndexSettings.INDEX_TRANSLOG_FLUSH_THRESHOLD_SIZE_SETTING.getKey(), flushThreshold + "b")).build();
+        indexSettings.updateIndexMetaData(indexMetaData);
+        engine.onSettingsChanged();
+        final int numOps = scaledRandomIntBetween(100, 10_000);
+        for (int i = 0; i < numOps; i++) {
+            final long localCheckPoint = engine.getLocalCheckpointTracker().getCheckpoint();
+            final long seqno = randomLongBetween(Math.max(0, localCheckPoint), localCheckPoint + 5);
+            final ParsedDocument doc = testParsedDocument(Long.toString(seqno), null, testDocumentWithTextField(), SOURCE, null);
+            engine.index(replicaIndexForDoc(doc, 1L, seqno, false));
+            if (rarely() && engine.getTranslog().shouldRollGeneration()) {
+                engine.rollTranslogGeneration();
+            }
+            if (rarely() || engine.shouldPeriodicallyFlush()) {
+                engine.flush();
+                assertThat(engine.shouldPeriodicallyFlush(), equalTo(false));
+            }
+        }
+    }
 
     public void testStressUpdateSameDocWhileGettingIt() throws IOException, InterruptedException {
         final int iters = randomIntBetween(1, 15);