A replica can be promoted and started in one cluster state update (#32042)

When a replica is fully recovered (i.e., in `POST_RECOVERY` state) we send a request to the master
to start the shard. The master changes the state of the replica and publishes a cluster state to that
effect. In certain cases, that cluster state can be processed on the node hosting the replica
*together* with a cluster state that promotes that, now started, replica to a primary. This can
happen due to cluster state batched processing or if the master died after having committed the
cluster state that starts the shard but before publishing it to the node with the replica. If the master
also held the primary shard, the new master node will remove the primary (as it failed) and will also
immediately promote the replica (thinking it is started). 

Sadly our code in IndexShard didn't allow for this which caused [assertions](https://github.com/elastic/elasticsearch/blob/13917162ad5c59a96ccb4d6a81a5044546c45c22/server/src/main/java/org/elasticsearch/index/seqno/ReplicationTracker.java#L482) to be tripped in some of our tests runs.

Author: bleskes

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

@@ -413,10 +413,9 @@ public void updateShardState(final ShardRouting newRouting,
 
             if (state == IndexShardState.POST_RECOVERY && newRouting.active()) {
                 assert currentRouting.active() == false : "we are in POST_RECOVERY, but our shard routing is active " + currentRouting;
-
-                if (newRouting.primary() && currentRouting.isRelocationTarget() == false) {
-                    replicationTracker.activatePrimaryMode(getLocalCheckpoint());
-                }
+                assert currentRouting.isRelocationTarget()  == false || currentRouting.primary() == false ||
+                        replicationTracker.isPrimaryMode() :
+                    "a primary relocation is completed by the master, but primary mode is not active " + currentRouting;
 
                 changeState(IndexShardState.STARTED, "global state is [" + newRouting.state() + "]");
             } else if (currentRouting.primary() && currentRouting.relocating() && replicationTracker.isPrimaryMode() == false &&
@@ -432,7 +431,12 @@ public void updateShardState(final ShardRouting newRouting,
             final CountDownLatch shardStateUpdated = new CountDownLatch(1);
 
             if (newRouting.primary()) {
-                if (newPrimaryTerm != primaryTerm) {
+                if (newPrimaryTerm == primaryTerm) {
+                    if (currentRouting.initializing() && currentRouting.isRelocationTarget() == false && newRouting.active()) {
+                        // the master started a recovering primary, activate primary mode.
+                        replicationTracker.activatePrimaryMode(getLocalCheckpoint());
+                    }
+                } else {
                     assert currentRouting.primary() == false : "term is only increased as part of primary promotion";
                     /* Note that due to cluster state batching an initializing primary shard term can failed and re-assigned
                      * in one state causing it's term to be incremented. Note that if both current shard state and new
@@ -521,6 +525,11 @@ public void onFailure(Exception e) {
             }
             // set this last, once we finished updating all internal state.
             this.shardRouting = newRouting;
+
+            assert this.shardRouting.primary() == false ||
+                this.shardRouting.started() == false || // note that we use started and not active to avoid relocating shards
+                this.replicationTracker.isPrimaryMode()
+                : "an started primary must be in primary mode " + this.shardRouting;
             shardStateUpdated.countDown();
         }
         if (currentRouting != null && currentRouting.active() == false && newRouting.active()) {

server/src/test/java/org/elasticsearch/index/replication/IndexLevelReplicationTests.java

@@ -363,7 +363,7 @@ public void testSeqNoCollision() throws Exception {
             logger.info("--> Promote replica2 as the primary");
             shards.promoteReplicaToPrimary(replica2);
             logger.info("--> Recover replica3 from replica2");
-            recoverReplica(replica3, replica2);
+            recoverReplica(replica3, replica2, true);
             try (Translog.Snapshot snapshot = getTranslog(replica3).newSnapshot()) {
                 assertThat(snapshot.totalOperations(), equalTo(initDocs + 1));
                 assertThat(snapshot.next(), equalTo(op2));

server/src/test/java/org/elasticsearch/index/shard/IndexShardTests.java

@@ -226,6 +226,7 @@ public void testPersistenceStateMetadataPersistence() throws Exception {
     }
 
     public void testFailShard() throws Exception {
+        allowShardFailures();
         IndexShard shard = newStartedShard();
         final ShardPath shardPath = shard.shardPath();
         assertNotNull(shardPath);
@@ -309,7 +310,8 @@ public void testRejectOperationPermitWithHigherTermWhenNotStarted() throws IOExc
     }
 
     public void testPrimaryPromotionDelaysOperations() throws IOException, BrokenBarrierException, InterruptedException {
-        final IndexShard indexShard = newStartedShard(false);
+        final IndexShard indexShard = newShard(false);
+        recoveryEmptyReplica(indexShard, randomBoolean());
 
         final int operations = scaledRandomIntBetween(1, 64);
         final CyclicBarrier barrier = new CyclicBarrier(1 + operations);
@@ -353,20 +355,10 @@ public void onFailure(Exception e) {
         barrier.await();
         latch.await();
 
-        // promote the replica
         final ShardRouting replicaRouting = indexShard.routingEntry();
-        final ShardRouting primaryRouting =
-                newShardRouting(
-                        replicaRouting.shardId(),
-                        replicaRouting.currentNodeId(),
-                        null,
-                        true,
-                        ShardRoutingState.STARTED,
-                        replicaRouting.allocationId());
-        indexShard.updateShardState(primaryRouting, indexShard.getPrimaryTerm() + 1, (shard, listener) -> {},
-            0L, Collections.singleton(primaryRouting.allocationId().getId()),
-            new IndexShardRoutingTable.Builder(primaryRouting.shardId()).addShard(primaryRouting).build(),
-            Collections.emptySet());
+        promoteReplica(indexShard, Collections.singleton(replicaRouting.allocationId().getId()),
+            new IndexShardRoutingTable.Builder(replicaRouting.shardId()).addShard(replicaRouting).build());
+
 
         final int delayedOperations = scaledRandomIntBetween(1, 64);
         final CyclicBarrier delayedOperationsBarrier = new CyclicBarrier(1 + delayedOperations);
@@ -428,8 +420,9 @@ public void onFailure(Exception e) {
      * 1) Internal state (ala ReplicationTracker) have been updated
      * 2) Primary term is set to the new term
      */
-    public void testPublishingOrderOnPromotion() throws IOException, BrokenBarrierException, InterruptedException {
-        final IndexShard indexShard = newStartedShard(false);
+    public void testPublishingOrderOnPromotion() throws IOException, InterruptedException, BrokenBarrierException {
+        final IndexShard indexShard = newShard(false);
+        recoveryEmptyReplica(indexShard, randomBoolean());
         final long promotedTerm = indexShard.getPrimaryTerm() + 1;
         final CyclicBarrier barrier = new CyclicBarrier(2);
         final AtomicBoolean stop = new AtomicBoolean();
@@ -448,18 +441,10 @@ public void testPublishingOrderOnPromotion() throws IOException, BrokenBarrierEx
         });
         thread.start();
 
-        final ShardRouting replicaRouting = indexShard.routingEntry();
-        final ShardRouting primaryRouting = newShardRouting(replicaRouting.shardId(), replicaRouting.currentNodeId(), null, true,
-            ShardRoutingState.STARTED, replicaRouting.allocationId());
-
-
-        final Set<String> inSyncAllocationIds = Collections.singleton(primaryRouting.allocationId().getId());
-        final IndexShardRoutingTable routingTable =
-            new IndexShardRoutingTable.Builder(primaryRouting.shardId()).addShard(primaryRouting).build();
         barrier.await();
-        // promote the replica
-        indexShard.updateShardState(primaryRouting, promotedTerm, (shard, listener) -> {}, 0L, inSyncAllocationIds, routingTable,
-            Collections.emptySet());
+        final ShardRouting replicaRouting = indexShard.routingEntry();
+        promoteReplica(indexShard, Collections.singleton(replicaRouting.allocationId().getId()),
+            new IndexShardRoutingTable.Builder(replicaRouting.shardId()).addShard(replicaRouting).build());
 
         stop.set(true);
         thread.join();
@@ -468,7 +453,8 @@ public void testPublishingOrderOnPromotion() throws IOException, BrokenBarrierEx
 
 
     public void testPrimaryFillsSeqNoGapsOnPromotion() throws Exception {
-        final IndexShard indexShard = newStartedShard(false);
+        final IndexShard indexShard = newShard(false);
+        recoveryEmptyReplica(indexShard, randomBoolean());
 
         // most of the time this is large enough that most of the time there will be at least one gap
         final int operations = 1024 - scaledRandomIntBetween(0, 1024);
@@ -479,17 +465,8 @@ public void testPrimaryFillsSeqNoGapsOnPromotion() throws Exception {
 
         // promote the replica
         final ShardRouting replicaRouting = indexShard.routingEntry();
-        final ShardRouting primaryRouting =
-                newShardRouting(
-                        replicaRouting.shardId(),
-                        replicaRouting.currentNodeId(),
-                        null,
-                        true,
-                        ShardRoutingState.STARTED,
-                        replicaRouting.allocationId());
-        indexShard.updateShardState(primaryRouting, indexShard.getPrimaryTerm() + 1, (shard, listener) -> {},
-            0L, Collections.singleton(primaryRouting.allocationId().getId()),
-            new IndexShardRoutingTable.Builder(primaryRouting.shardId()).addShard(primaryRouting).build(), Collections.emptySet());
+        promoteReplica(indexShard, Collections.singleton(replicaRouting.allocationId().getId()),
+            new IndexShardRoutingTable.Builder(replicaRouting.shardId()).addShard(replicaRouting).build());
 
         /*
          * This operation completing means that the delay operation executed as part of increasing the primary term has completed and the
@@ -506,7 +483,7 @@ public void onResponse(Releasable releasable) {
 
                     @Override
                     public void onFailure(Exception e) {
-                        throw new RuntimeException(e);
+                        throw new AssertionError(e);
                     }
                 },
                 ThreadPool.Names.GENERIC, "");
@@ -846,7 +823,7 @@ public void testGlobalCheckpointSync() throws IOException {
         // add a replica
         recoverShardFromStore(primaryShard);
         final IndexShard replicaShard = newShard(shardId, false);
-        recoverReplica(replicaShard, primaryShard);
+        recoverReplica(replicaShard, primaryShard, true);
         final int maxSeqNo = randomIntBetween(0, 128);
         for (int i = 0; i <= maxSeqNo; i++) {
             EngineTestCase.generateNewSeqNo(primaryShard.getEngine());
@@ -1625,7 +1602,7 @@ public void testPrimaryHandOffUpdatesLocalCheckpoint() throws IOException {
         IndexShardTestCase.updateRoutingEntry(primarySource, primarySource.routingEntry().relocate(randomAlphaOfLength(10), -1));
         final IndexShard primaryTarget = newShard(primarySource.routingEntry().getTargetRelocatingShard());
         updateMappings(primaryTarget, primarySource.indexSettings().getIndexMetaData());
-        recoverReplica(primaryTarget, primarySource);
+        recoverReplica(primaryTarget, primarySource, true);
 
         // check that local checkpoint of new primary is properly tracked after primary relocation
         assertThat(primaryTarget.getLocalCheckpoint(), equalTo(totalOps - 1L));
@@ -2082,7 +2059,7 @@ public long indexTranslogOperations(List<Translog.Operation> operations, int tot
                     assertFalse(replica.isSyncNeeded());
                     return localCheckpoint;
                 }
-            }, true);
+            }, true, true);
 
         closeShards(primary, replica);
     }
@@ -2189,7 +2166,7 @@ public long indexTranslogOperations(List<Translog.Operation> operations, int tot
                     assertTrue(replica.isActive());
                     return localCheckpoint;
                 }
-            }, false);
+            }, false, true);
 
         closeShards(primary, replica);
     }
@@ -2241,7 +2218,7 @@ public void finalizeRecovery(long globalCheckpoint) throws IOException {
                     super.finalizeRecovery(globalCheckpoint);
                     assertListenerCalled.accept(replica);
                 }
-            }, false);
+            }, false, true);
 
         closeShards(primary, replica);
     }

server/src/test/java/org/elasticsearch/indices/cluster/AbstractIndicesClusterStateServiceTestCase.java

@@ -357,6 +357,14 @@ public void updateShardState(ShardRouting shardRouting,
                 assertTrue("and active shard must stay active, current: " + this.shardRouting + ", got: " + shardRouting,
                     shardRouting.active());
             }
+            if (this.shardRouting.primary()) {
+                assertTrue("a primary shard can't be demoted", shardRouting.primary());
+            } else if (shardRouting.primary()) {
+                // note: it's ok for a replica in post recovery to be started and promoted at once
+                // this can happen when the primary failed after we sent the start shard message
+                assertTrue("a replica can only be promoted when active. current: " + this.shardRouting + " new: " + shardRouting,
+                    shardRouting.active());
+            }
             this.shardRouting = shardRouting;
             if (shardRouting.primary()) {
                 term = newPrimaryTerm;

server/src/test/java/org/elasticsearch/indices/recovery/PeerRecoveryTargetServiceTests.java

@@ -43,7 +43,7 @@ public void testGetStartingSeqNo() throws Exception {
         try {
             // Empty store
             {
-                recoveryEmptyReplica(replica);
+                recoveryEmptyReplica(replica, true);
                 final RecoveryTarget recoveryTarget = new RecoveryTarget(replica, null, null, null);
                 assertThat(PeerRecoveryTargetService.getStartingSeqNo(logger, recoveryTarget), equalTo(0L));
                 recoveryTarget.decRef();

server/src/test/java/org/elasticsearch/indices/recovery/RecoveryTests.java

@@ -261,7 +261,7 @@ public void testPeerRecoverySendSafeCommitInFileBased() throws Exception {
         }
         IndexShard replicaShard = newShard(primaryShard.shardId(), false);
         updateMappings(replicaShard, primaryShard.indexSettings().getIndexMetaData());
-        recoverReplica(replicaShard, primaryShard);
+        recoverReplica(replicaShard, primaryShard, true);
         List<IndexCommit> commits = DirectoryReader.listCommits(replicaShard.store().directory());
         long maxSeqNo = Long.parseLong(commits.get(0).getUserData().get(SequenceNumbers.MAX_SEQ_NO));
         assertThat(maxSeqNo, lessThanOrEqualTo(globalCheckpoint));

test/framework/src/main/java/org/elasticsearch/index/replication/ESIndexLevelReplicationTestCase.java

@@ -265,7 +265,7 @@ public synchronized IndexShard addReplicaWithExistingPath(final ShardPath shardP
                 RecoverySource.PeerRecoverySource.INSTANCE);
 
             final IndexShard newReplica =
-                    newShard(shardRouting, shardPath, indexMetaData, null, getEngineFactory(shardRouting), () -> {}, EMPTY_EVENT_LISTENER);
+                newShard(shardRouting, shardPath, indexMetaData, null, getEngineFactory(shardRouting), () -> {}, EMPTY_EVENT_LISTENER);
             replicas.add(newReplica);
             updateAllocationIDsOnPrimary();
             return newReplica;
@@ -341,8 +341,11 @@ public void recoverReplica(
             IndexShard replica,
             BiFunction<IndexShard, DiscoveryNode, RecoveryTarget> targetSupplier,
             boolean markAsRecovering) throws IOException {
-            ESIndexLevelReplicationTestCase.this.recoverReplica(replica, primary, targetSupplier, markAsRecovering, activeIds(),
-                routingTable(Function.identity()));
+            final IndexShardRoutingTable routingTable = routingTable(Function.identity());
+            final Set<String> inSyncIds = activeIds();
+            ESIndexLevelReplicationTestCase.this.recoverUnstartedReplica(replica, primary, targetSupplier, markAsRecovering, inSyncIds,
+                routingTable);
+            ESIndexLevelReplicationTestCase.this.startReplicaAfterRecovery(replica, primary, inSyncIds, routingTable);
         }
 
         public synchronized DiscoveryNode getPrimaryNode() {