[Segment Replication] - Update replicas to commit SegmentInfos instead of relying on segments_N from primary shards.

CHANGELOG.md

@@ -24,6 +24,7 @@ Inspired from [Keep a Changelog](https://keepachangelog.com/en/1.0.0/)
 - Use RemoteSegmentStoreDirectory instead of RemoteDirectory ([#4240](https://github.com/opensearch-project/OpenSearch/pull/4240))
 - Plugin ZIP publication groupId value is configurable ([#4156](https://github.com/opensearch-project/OpenSearch/pull/4156))
 - Add index specific setting for remote repository ([#4253](https://github.com/opensearch-project/OpenSearch/pull/4253))
+- [Segment Replication] Update replicas to commit SegmentInfos instead of relying on SIS files from primary shards. ([#4402](https://github.com/opensearch-project/OpenSearch/pull/4402))
 
 ### Deprecated
 
@@ -70,4 +71,4 @@ Inspired from [Keep a Changelog](https://keepachangelog.com/en/1.0.0/)
 
 
 [Unreleased]: https://github.com/opensearch-project/OpenSearch/compare/2.2.0...HEAD
-[2.x]: https://github.com/opensearch-project/OpenSearch/compare/2.2.0...2.x
+[2.x]: https://github.com/opensearch-project/OpenSearch/compare/2.2.0...2.x

server/src/main/java/org/opensearch/index/engine/NRTReplicationEngine.java

@@ -54,6 +54,8 @@ public class NRTReplicationEngine extends Engine {
     private final LocalCheckpointTracker localCheckpointTracker;
     private final WriteOnlyTranslogManager translogManager;
 
+    private volatile long lastReceivedGen = SequenceNumbers.NO_OPS_PERFORMED;
+
     private static final int SI_COUNTER_INCREMENT = 10;
 
     public NRTReplicationEngine(EngineConfig engineConfig) {
@@ -120,14 +122,16 @@ public TranslogManager translogManager() {
 
     public synchronized void updateSegments(final SegmentInfos infos, long seqNo) throws IOException {
         // Update the current infos reference on the Engine's reader.
+        final long incomingGeneration = infos.getGeneration();
         readerManager.updateSegments(infos);
 
-        // only update the persistedSeqNo and "lastCommitted" infos reference if the incoming segments have a higher
-        // generation. We can still refresh with incoming SegmentInfos that are not part of a commit point.
-        if (infos.getGeneration() > lastCommittedSegmentInfos.getGeneration()) {
-            this.lastCommittedSegmentInfos = infos;
+        // Commit and roll the xlog when we receive a different generation than what was last received.
+        // lower/higher gens are possible from a new primary that was just elected.
+        if (incomingGeneration != lastReceivedGen) {
+            commitSegmentInfos();
             translogManager.rollTranslogGeneration();
         }
+        lastReceivedGen = incomingGeneration;
         localCheckpointTracker.fastForwardProcessedSeqNo(seqNo);
     }
 
@@ -141,20 +145,16 @@ public synchronized void updateSegments(final SegmentInfos infos, long seqNo) th
      *
      * @throws IOException - When there is an IO error committing the SegmentInfos.
      */
-    public void commitSegmentInfos() throws IOException {
-        // TODO: This method should wait for replication events to finalize.
-        final SegmentInfos latestSegmentInfos = getLatestSegmentInfos();
-        /*
-         This is a workaround solution which decreases the chances of conflict on replica nodes when same file is copied
-         from two different primaries during failover. Increasing counter helps in avoiding this conflict as counter is
-         used to generate new segment file names. The ideal solution is to identify the counter from previous primary.
-         */
-        latestSegmentInfos.counter = latestSegmentInfos.counter + SI_COUNTER_INCREMENT;
-        latestSegmentInfos.changed();
-        store.commitSegmentInfos(latestSegmentInfos, localCheckpointTracker.getMaxSeqNo(), localCheckpointTracker.getProcessedCheckpoint());
+    private void commitSegmentInfos(SegmentInfos infos) throws IOException {
+        store.commitSegmentInfos(infos, localCheckpointTracker.getMaxSeqNo(), localCheckpointTracker.getProcessedCheckpoint());
+        this.lastCommittedSegmentInfos = store.readLastCommittedSegmentsInfo();
         translogManager.syncTranslog();
     }
 
+    protected void commitSegmentInfos() throws IOException {
+        commitSegmentInfos(getLatestSegmentInfos());
+    }
+
     @Override
     public String getHistoryUUID() {
         return loadHistoryUUID(lastCommittedSegmentInfos.userData);
@@ -354,6 +354,15 @@ protected final void closeNoLock(String reason, CountDownLatch closedLatch) {
             assert rwl.isWriteLockedByCurrentThread() || failEngineLock.isHeldByCurrentThread()
                 : "Either the write lock must be held or the engine must be currently be failing itself";
             try {
+                final SegmentInfos latestSegmentInfos = getLatestSegmentInfos();
+                /*
+                 This is a workaround solution which decreases the chances of conflict on replica nodes when same file is copied
+                 from two different primaries during failover. Increasing counter helps in avoiding this conflict as counter is
+                 used to generate new segment file names. The ideal solution is to identify the counter from previous primary.
+                 */
+                latestSegmentInfos.counter = latestSegmentInfos.counter + SI_COUNTER_INCREMENT;
+                latestSegmentInfos.changed();
+                commitSegmentInfos(latestSegmentInfos);
                 IOUtils.close(readerManager, translogManager, store::decRef);
             } catch (Exception e) {
                 logger.warn("failed to close engine", e);

server/src/main/java/org/opensearch/index/engine/NRTReplicationReaderManager.java

@@ -74,6 +74,9 @@ protected OpenSearchDirectoryReader refreshIfNeeded(OpenSearchDirectoryReader re
      * @throws IOException - When Refresh fails with an IOException.
      */
     public synchronized void updateSegments(SegmentInfos infos) throws IOException {
+        // roll over the currentInfo's generation, this ensures the on-disk gen
+        // is always increased.
+        infos.updateGeneration(currentInfos);
         currentInfos = infos;
         maybeRefresh();
     }

server/src/main/java/org/opensearch/index/shard/IndexShard.java

@@ -623,7 +623,7 @@ public void updateShardState(
                             if (indexSettings.isSegRepEnabled()) {
                                 // this Shard's engine was read only, we need to update its engine before restoring local history from xlog.
                                 assert newRouting.primary() && currentRouting.primary() == false;
-                                promoteNRTReplicaToPrimary();
+                                resetEngineToGlobalCheckpoint();
                             }
                             replicationTracker.activatePrimaryMode(getLocalCheckpoint());
                             ensurePeerRecoveryRetentionLeasesExist();
@@ -3557,7 +3557,9 @@ private void innerAcquireReplicaOperationPermit(
                             currentGlobalCheckpoint,
                             maxSeqNo
                         );
-                        if (currentGlobalCheckpoint < maxSeqNo) {
+                        // With Segment Replication enabled, we never want to reset a replica's engine unless
+                        // it is promoted to primary.
+                        if (currentGlobalCheckpoint < maxSeqNo && indexSettings.isSegRepEnabled() == false) {
                             resetEngineToGlobalCheckpoint();
                         } else {
                             getEngine().translogManager().rollTranslogGeneration();
@@ -4120,26 +4122,4 @@ RetentionLeaseSyncer getRetentionLeaseSyncer() {
     public GatedCloseable<SegmentInfos> getSegmentInfosSnapshot() {
         return getEngine().getSegmentInfosSnapshot();
     }
-
-    /**
-     * With segment replication enabled - prepare the shard's engine to be promoted as the new primary.
-     *
-     * If this shard is currently using a replication engine, this method:
-     * 1. Invokes {@link NRTReplicationEngine#commitSegmentInfos()} to ensure the engine can be reopened as writeable from the latest refresh point.
-     * InternalEngine opens its IndexWriter from an on-disk commit point, but this replica may have recently synced from a primary's refresh point, meaning it has documents searchable in its in-memory SegmentInfos
-     * that are not part of a commit point.  This ensures that those documents are made part of a commit and do not need to be reindexed after promotion.
-     * 2. Invokes resetEngineToGlobalCheckpoint - This call performs the engine swap, opening up as a writeable engine and replays any operations in the xlog. The operations indexed from xlog here will be
-     * any ack'd writes that were not copied to this replica before promotion.
-     */
-    private void promoteNRTReplicaToPrimary() {
-        assert shardRouting.primary() && indexSettings.isSegRepEnabled();
-        getReplicationEngine().ifPresentOrElse(engine -> {
-            try {
-                engine.commitSegmentInfos();
-                resetEngineToGlobalCheckpoint();
-            } catch (IOException e) {
-                throw new EngineException(shardId, "Unable to update  replica to writeable engine, failing shard", e);
-            }
-        }, () -> { throw new EngineException(shardId, "Expected replica engine to be of type NRTReplicationEngine"); });
-    }
 }

server/src/main/java/org/opensearch/indices/recovery/MultiFileWriter.java

@@ -156,7 +156,10 @@ private void innerWriteFileChunk(StoreFileMetadata fileMetadata, long position,
                 + temporaryFileName
                 + "] in "
                 + Arrays.toString(store.directory().listAll());
-            store.directory().sync(Collections.singleton(temporaryFileName));
+            // With Segment Replication, we will fsync after a full commit has been received.
+            if (store.indexSettings().isSegRepEnabled() == false) {
+                store.directory().sync(Collections.singleton(temporaryFileName));
+            }
             IndexOutput remove = removeOpenIndexOutputs(name);
             assert remove == null || remove == indexOutput; // remove maybe null if we got finished
         }