Segment Replication - Recover all translog ops when flipping to write…

…able engine. (#6352) (#6371) Signed-off-by: Marc Handalian <[email protected]> Co-authored-by: Marc Handalian <[email protected]>
opensearch-project · Feb 18, 2023 · d34596e · d34596e
1 parent dd2727f
commit d34596e
Showing 1 changed file with 3 additions and 1 deletion.
diff --git a/server/src/main/java/org/opensearch/index/shard/IndexShard.java b/server/src/main/java/org/opensearch/index/shard/IndexShard.java
@@ -4323,7 +4323,9 @@ public void close() throws IOException {
 
         // When the new engine is created, translogs are synced from remote store onto local. Since remote store is the source
         // of truth for translog, we play all translogs that exists locally. Otherwise, the recoverUpto happens upto global checkpoint.
-        long recoverUpto = this.isRemoteTranslogEnabled() ? Long.MAX_VALUE : globalCheckpoint;
+        // We also replay all local translog ops with Segment replication, because on engine swap our local translog may
+        // hold more ops than the global checkpoint.
+        long recoverUpto = this.isRemoteTranslogEnabled() || indexSettings().isSegRepEnabled() ? Long.MAX_VALUE : globalCheckpoint;
         newEngineReference.get().recoverFromTranslog(translogRunner, recoverUpto);
         newEngineReference.get().refresh("reset_engine");
         synchronized (engineMutex) {