BUG-8792: allow transactions to not time out after reconnect
[controller.git] / opendaylight / md-sal / sal-distributed-datastore / src / main / java / org / opendaylight / controller / cluster / datastore / Shard.java
1 /*
2  * Copyright (c) 2014 Cisco Systems, Inc. and others.  All rights reserved.
3  *
4  * This program and the accompanying materials are made available under the
5  * terms of the Eclipse Public License v1.0 which accompanies this distribution,
6  * and is available at http://www.eclipse.org/legal/epl-v10.html
7  */
8
9 package org.opendaylight.controller.cluster.datastore;
10
11 import akka.actor.ActorRef;
12 import akka.actor.ActorSelection;
13 import akka.actor.Cancellable;
14 import akka.actor.Props;
15 import akka.actor.Status;
16 import akka.actor.Status.Failure;
17 import akka.serialization.Serialization;
18 import com.google.common.annotations.VisibleForTesting;
19 import com.google.common.base.Optional;
20 import com.google.common.base.Preconditions;
21 import com.google.common.base.Ticker;
22 import com.google.common.base.Verify;
23 import com.google.common.collect.ImmutableList;
24 import com.google.common.collect.ImmutableMap;
25 import com.google.common.collect.Range;
26 import java.io.IOException;
27 import java.util.Arrays;
28 import java.util.Collection;
29 import java.util.Collections;
30 import java.util.Map;
31 import java.util.concurrent.TimeUnit;
32 import javax.annotation.Nonnull;
33 import javax.annotation.Nullable;
34 import org.opendaylight.controller.cluster.access.ABIVersion;
35 import org.opendaylight.controller.cluster.access.commands.ConnectClientRequest;
36 import org.opendaylight.controller.cluster.access.commands.ConnectClientSuccess;
37 import org.opendaylight.controller.cluster.access.commands.LocalHistoryRequest;
38 import org.opendaylight.controller.cluster.access.commands.NotLeaderException;
39 import org.opendaylight.controller.cluster.access.commands.OutOfSequenceEnvelopeException;
40 import org.opendaylight.controller.cluster.access.commands.TransactionRequest;
41 import org.opendaylight.controller.cluster.access.concepts.ClientIdentifier;
42 import org.opendaylight.controller.cluster.access.concepts.FrontendIdentifier;
43 import org.opendaylight.controller.cluster.access.concepts.LocalHistoryIdentifier;
44 import org.opendaylight.controller.cluster.access.concepts.Request;
45 import org.opendaylight.controller.cluster.access.concepts.RequestEnvelope;
46 import org.opendaylight.controller.cluster.access.concepts.RequestException;
47 import org.opendaylight.controller.cluster.access.concepts.RequestSuccess;
48 import org.opendaylight.controller.cluster.access.concepts.RetiredGenerationException;
49 import org.opendaylight.controller.cluster.access.concepts.RuntimeRequestException;
50 import org.opendaylight.controller.cluster.access.concepts.TransactionIdentifier;
51 import org.opendaylight.controller.cluster.access.concepts.UnsupportedRequestException;
52 import org.opendaylight.controller.cluster.common.actor.CommonConfig;
53 import org.opendaylight.controller.cluster.common.actor.MessageTracker;
54 import org.opendaylight.controller.cluster.common.actor.MessageTracker.Error;
55 import org.opendaylight.controller.cluster.common.actor.MeteringBehavior;
56 import org.opendaylight.controller.cluster.datastore.exceptions.NoShardLeaderException;
57 import org.opendaylight.controller.cluster.datastore.identifiers.ShardIdentifier;
58 import org.opendaylight.controller.cluster.datastore.jmx.mbeans.shard.ShardMBeanFactory;
59 import org.opendaylight.controller.cluster.datastore.jmx.mbeans.shard.ShardStats;
60 import org.opendaylight.controller.cluster.datastore.messages.AbortTransaction;
61 import org.opendaylight.controller.cluster.datastore.messages.ActorInitialized;
62 import org.opendaylight.controller.cluster.datastore.messages.BatchedModifications;
63 import org.opendaylight.controller.cluster.datastore.messages.CanCommitTransaction;
64 import org.opendaylight.controller.cluster.datastore.messages.CloseTransactionChain;
65 import org.opendaylight.controller.cluster.datastore.messages.CommitTransaction;
66 import org.opendaylight.controller.cluster.datastore.messages.CreateTransaction;
67 import org.opendaylight.controller.cluster.datastore.messages.CreateTransactionReply;
68 import org.opendaylight.controller.cluster.datastore.messages.ForwardedReadyTransaction;
69 import org.opendaylight.controller.cluster.datastore.messages.GetShardDataTree;
70 import org.opendaylight.controller.cluster.datastore.messages.MakeLeaderLocal;
71 import org.opendaylight.controller.cluster.datastore.messages.OnDemandShardState;
72 import org.opendaylight.controller.cluster.datastore.messages.PeerAddressResolved;
73 import org.opendaylight.controller.cluster.datastore.messages.PersistAbortTransactionPayload;
74 import org.opendaylight.controller.cluster.datastore.messages.ReadyLocalTransaction;
75 import org.opendaylight.controller.cluster.datastore.messages.RegisterChangeListener;
76 import org.opendaylight.controller.cluster.datastore.messages.RegisterDataTreeChangeListener;
77 import org.opendaylight.controller.cluster.datastore.messages.ShardLeaderStateChanged;
78 import org.opendaylight.controller.cluster.datastore.messages.UpdateSchemaContext;
79 import org.opendaylight.controller.cluster.datastore.persisted.AbortTransactionPayload;
80 import org.opendaylight.controller.cluster.datastore.persisted.DatastoreSnapshot;
81 import org.opendaylight.controller.cluster.datastore.persisted.DatastoreSnapshot.ShardSnapshot;
82 import org.opendaylight.controller.cluster.datastore.utils.Dispatchers;
83 import org.opendaylight.controller.cluster.notifications.LeaderStateChanged;
84 import org.opendaylight.controller.cluster.notifications.RegisterRoleChangeListener;
85 import org.opendaylight.controller.cluster.notifications.RoleChangeNotifier;
86 import org.opendaylight.controller.cluster.raft.LeadershipTransferFailedException;
87 import org.opendaylight.controller.cluster.raft.RaftActor;
88 import org.opendaylight.controller.cluster.raft.RaftActorRecoveryCohort;
89 import org.opendaylight.controller.cluster.raft.RaftActorSnapshotCohort;
90 import org.opendaylight.controller.cluster.raft.RaftState;
91 import org.opendaylight.controller.cluster.raft.base.messages.FollowerInitialSyncUpStatus;
92 import org.opendaylight.controller.cluster.raft.client.messages.OnDemandRaftState;
93 import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
94 import org.opendaylight.controller.cluster.raft.messages.RequestLeadership;
95 import org.opendaylight.controller.cluster.raft.messages.ServerRemoved;
96 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload;
97 import org.opendaylight.yangtools.concepts.Identifier;
98 import org.opendaylight.yangtools.yang.data.api.schema.tree.DataValidationFailedException;
99 import org.opendaylight.yangtools.yang.data.api.schema.tree.TipProducingDataTree;
100 import org.opendaylight.yangtools.yang.data.api.schema.tree.TreeType;
101 import org.opendaylight.yangtools.yang.model.api.SchemaContext;
102 import org.opendaylight.yangtools.yang.model.api.SchemaContextProvider;
103 import scala.concurrent.duration.Duration;
104 import scala.concurrent.duration.FiniteDuration;
105
106 /**
107  * A Shard represents a portion of the logical data tree.
108  *
109  * <p>
110  * Our Shard uses InMemoryDataTree as it's internal representation and delegates all requests it
111  */
112 public class Shard extends RaftActor {
113
114     @VisibleForTesting
115     static final Object TX_COMMIT_TIMEOUT_CHECK_MESSAGE = new Object() {
116         @Override
117         public String toString() {
118             return "txCommitTimeoutCheck";
119         }
120     };
121
122     @VisibleForTesting
123     static final Object GET_SHARD_MBEAN_MESSAGE = new Object() {
124         @Override
125         public String toString() {
126             return "getShardMBeanMessage";
127         }
128     };
129
130     static final Object RESUME_NEXT_PENDING_TRANSACTION = new Object() {
131         @Override
132         public String toString() {
133             return "resumeNextPendingTransaction";
134         }
135     };
136
137     // FIXME: shard names should be encapsulated in their own class and this should be exposed as a constant.
138     public static final String DEFAULT_NAME = "default";
139
140     private static final Collection<ABIVersion> SUPPORTED_ABIVERSIONS;
141
142     static {
143         final ABIVersion[] values = ABIVersion.values();
144         final ABIVersion[] real = Arrays.copyOfRange(values, 1, values.length - 1);
145         SUPPORTED_ABIVERSIONS = ImmutableList.copyOf(real).reverse();
146     }
147
148     // FIXME: make this a dynamic property based on mailbox size and maximum number of clients
149     private static final int CLIENT_MAX_MESSAGES = 1000;
150
151     // The state of this Shard
152     private final ShardDataTree store;
153
154     /// The name of this shard
155     private final String name;
156
157     private final ShardStats shardMBean;
158
159     private DatastoreContext datastoreContext;
160
161     private final ShardCommitCoordinator commitCoordinator;
162
163     private long transactionCommitTimeout;
164
165     private Cancellable txCommitTimeoutCheckSchedule;
166
167     private final Optional<ActorRef> roleChangeNotifier;
168
169     private final MessageTracker appendEntriesReplyTracker;
170
171     private final ShardTransactionActorFactory transactionActorFactory;
172
173     private final ShardSnapshotCohort snapshotCohort;
174
175     private final DataTreeChangeListenerSupport treeChangeSupport = new DataTreeChangeListenerSupport(this);
176     private final DataChangeListenerSupport changeSupport = new DataChangeListenerSupport(this);
177
178
179     private ShardSnapshot restoreFromSnapshot;
180
181     private final ShardTransactionMessageRetrySupport messageRetrySupport;
182
183     private final FrontendMetadata frontendMetadata;
184     private Map<FrontendIdentifier, LeaderFrontendState> knownFrontends = ImmutableMap.of();
185     private boolean paused;
186
187     protected Shard(final AbstractBuilder<?, ?> builder) {
188         super(builder.getId().toString(), builder.getPeerAddresses(),
189                 Optional.of(builder.getDatastoreContext().getShardRaftConfig()), DataStoreVersions.CURRENT_VERSION);
190
191         this.name = builder.getId().toString();
192         this.datastoreContext = builder.getDatastoreContext();
193         this.restoreFromSnapshot = builder.getRestoreFromSnapshot();
194         this.frontendMetadata = new FrontendMetadata(name);
195
196         setPersistence(datastoreContext.isPersistent());
197
198         LOG.info("Shard created : {}, persistent : {}", name, datastoreContext.isPersistent());
199
200         ShardDataTreeChangeListenerPublisherActorProxy treeChangeListenerPublisher =
201                 new ShardDataTreeChangeListenerPublisherActorProxy(getContext(), name + "-DTCL-publisher", name);
202         ShardDataChangeListenerPublisherActorProxy dataChangeListenerPublisher =
203                 new ShardDataChangeListenerPublisherActorProxy(getContext(), name + "-DCL-publisher", name);
204         if (builder.getDataTree() != null) {
205             store = new ShardDataTree(this, builder.getSchemaContext(), builder.getDataTree(),
206                     treeChangeListenerPublisher, dataChangeListenerPublisher, name, frontendMetadata);
207         } else {
208             store = new ShardDataTree(this, builder.getSchemaContext(), builder.getTreeType(),
209                     builder.getDatastoreContext().getStoreRoot(), treeChangeListenerPublisher,
210                     dataChangeListenerPublisher, name, frontendMetadata);
211         }
212
213         shardMBean = ShardMBeanFactory.getShardStatsMBean(name, datastoreContext.getDataStoreMXBeanType(), this);
214
215         if (isMetricsCaptureEnabled()) {
216             getContext().become(new MeteringBehavior(this));
217         }
218
219         commitCoordinator = new ShardCommitCoordinator(store, LOG, this.name);
220
221         setTransactionCommitTimeout();
222
223         // create a notifier actor for each cluster member
224         roleChangeNotifier = createRoleChangeNotifier(name);
225
226         appendEntriesReplyTracker = new MessageTracker(AppendEntriesReply.class,
227                 getRaftActorContext().getConfigParams().getIsolatedCheckIntervalInMillis());
228
229         transactionActorFactory = new ShardTransactionActorFactory(store, datastoreContext,
230             new Dispatchers(context().system().dispatchers()).getDispatcherPath(Dispatchers.DispatcherType.Transaction),
231                 self(), getContext(), shardMBean, builder.getId().getShardName());
232
233         snapshotCohort = ShardSnapshotCohort.create(getContext(), builder.getId().getMemberName(), store, LOG,
234             this.name);
235
236         messageRetrySupport = new ShardTransactionMessageRetrySupport(this);
237     }
238
239     private void setTransactionCommitTimeout() {
240         transactionCommitTimeout = TimeUnit.MILLISECONDS.convert(
241                 datastoreContext.getShardTransactionCommitTimeoutInSeconds(), TimeUnit.SECONDS) / 2;
242     }
243
244     private Optional<ActorRef> createRoleChangeNotifier(final String shardId) {
245         ActorRef shardRoleChangeNotifier = this.getContext().actorOf(
246             RoleChangeNotifier.getProps(shardId), shardId + "-notifier");
247         return Optional.of(shardRoleChangeNotifier);
248     }
249
250     @Override
251     public void postStop() {
252         LOG.info("Stopping Shard {}", persistenceId());
253
254         super.postStop();
255
256         messageRetrySupport.close();
257
258         if (txCommitTimeoutCheckSchedule != null) {
259             txCommitTimeoutCheckSchedule.cancel();
260         }
261
262         commitCoordinator.abortPendingTransactions("Transaction aborted due to shutdown.", this);
263
264         shardMBean.unregisterMBean();
265     }
266
267     @Override
268     protected void handleRecover(final Object message) {
269         LOG.debug("{}: onReceiveRecover: Received message {} from {}", persistenceId(), message.getClass(),
270             getSender());
271
272         super.handleRecover(message);
273         if (LOG.isTraceEnabled()) {
274             appendEntriesReplyTracker.begin();
275         }
276     }
277
278     @SuppressWarnings("checkstyle:IllegalCatch")
279     @Override
280     protected void handleNonRaftCommand(final Object message) {
281         try (MessageTracker.Context context = appendEntriesReplyTracker.received(message)) {
282             final Optional<Error> maybeError = context.error();
283             if (maybeError.isPresent()) {
284                 LOG.trace("{} : AppendEntriesReply failed to arrive at the expected interval {}", persistenceId(),
285                     maybeError.get());
286             }
287
288             store.resetTransactionBatch();
289
290             if (message instanceof RequestEnvelope) {
291                 final long now = ticker().read();
292                 final RequestEnvelope envelope = (RequestEnvelope)message;
293
294                 try {
295                     final RequestSuccess<?, ?> success = handleRequest(envelope, now);
296                     if (success != null) {
297                         envelope.sendSuccess(success, ticker().read() - now);
298                     }
299                 } catch (RequestException e) {
300                     LOG.debug("{}: request {} failed", persistenceId(), envelope, e);
301                     envelope.sendFailure(e, ticker().read() - now);
302                 } catch (Exception e) {
303                     LOG.debug("{}: request {} caused failure", persistenceId(), envelope, e);
304                     envelope.sendFailure(new RuntimeRequestException("Request failed to process", e),
305                         ticker().read() - now);
306                 }
307             } else if (message instanceof ConnectClientRequest) {
308                 handleConnectClient((ConnectClientRequest)message);
309             } else if (CreateTransaction.isSerializedType(message)) {
310                 handleCreateTransaction(message);
311             } else if (message instanceof BatchedModifications) {
312                 handleBatchedModifications((BatchedModifications)message);
313             } else if (message instanceof ForwardedReadyTransaction) {
314                 handleForwardedReadyTransaction((ForwardedReadyTransaction) message);
315             } else if (message instanceof ReadyLocalTransaction) {
316                 handleReadyLocalTransaction((ReadyLocalTransaction)message);
317             } else if (CanCommitTransaction.isSerializedType(message)) {
318                 handleCanCommitTransaction(CanCommitTransaction.fromSerializable(message));
319             } else if (CommitTransaction.isSerializedType(message)) {
320                 handleCommitTransaction(CommitTransaction.fromSerializable(message));
321             } else if (AbortTransaction.isSerializedType(message)) {
322                 handleAbortTransaction(AbortTransaction.fromSerializable(message));
323             } else if (CloseTransactionChain.isSerializedType(message)) {
324                 closeTransactionChain(CloseTransactionChain.fromSerializable(message));
325             } else if (message instanceof RegisterChangeListener) {
326                 changeSupport.onMessage((RegisterChangeListener) message, isLeader(), hasLeader());
327             } else if (message instanceof RegisterDataTreeChangeListener) {
328                 treeChangeSupport.onMessage((RegisterDataTreeChangeListener) message, isLeader(), hasLeader());
329             } else if (message instanceof UpdateSchemaContext) {
330                 updateSchemaContext((UpdateSchemaContext) message);
331             } else if (message instanceof PeerAddressResolved) {
332                 PeerAddressResolved resolved = (PeerAddressResolved) message;
333                 setPeerAddress(resolved.getPeerId(), resolved.getPeerAddress());
334             } else if (TX_COMMIT_TIMEOUT_CHECK_MESSAGE.equals(message)) {
335                 commitTimeoutCheck();
336             } else if (message instanceof DatastoreContext) {
337                 onDatastoreContext((DatastoreContext)message);
338             } else if (message instanceof RegisterRoleChangeListener) {
339                 roleChangeNotifier.get().forward(message, context());
340             } else if (message instanceof FollowerInitialSyncUpStatus) {
341                 shardMBean.setFollowerInitialSyncStatus(((FollowerInitialSyncUpStatus) message).isInitialSyncDone());
342                 context().parent().tell(message, self());
343             } else if (GET_SHARD_MBEAN_MESSAGE.equals(message)) {
344                 sender().tell(getShardMBean(), self());
345             } else if (message instanceof GetShardDataTree) {
346                 sender().tell(store.getDataTree(), self());
347             } else if (message instanceof ServerRemoved) {
348                 context().parent().forward(message, context());
349             } else if (ShardTransactionMessageRetrySupport.TIMER_MESSAGE_CLASS.isInstance(message)) {
350                 messageRetrySupport.onTimerMessage(message);
351             } else if (message instanceof DataTreeCohortActorRegistry.CohortRegistryCommand) {
352                 store.processCohortRegistryCommand(getSender(),
353                         (DataTreeCohortActorRegistry.CohortRegistryCommand) message);
354             } else if (message instanceof PersistAbortTransactionPayload) {
355                 final TransactionIdentifier txId = ((PersistAbortTransactionPayload) message).getTransactionId();
356                 persistPayload(txId, AbortTransactionPayload.create(txId), true);
357             } else if (message instanceof MakeLeaderLocal) {
358                 onMakeLeaderLocal();
359             } else if (RESUME_NEXT_PENDING_TRANSACTION.equals(message)) {
360                 store.resumeNextPendingTransaction();
361             } else {
362                 super.handleNonRaftCommand(message);
363             }
364         }
365     }
366
367     private void commitTimeoutCheck() {
368         store.checkForExpiredTransactions(transactionCommitTimeout, this::updateAccess);
369         commitCoordinator.checkForExpiredTransactions(transactionCommitTimeout, this);
370     }
371
372     private Optional<Long> updateAccess(final SimpleShardDataTreeCohort cohort) {
373         final FrontendIdentifier frontend = cohort.getIdentifier().getHistoryId().getClientId().getFrontendId();
374         // If this frontend has freshly connected, give it some time to catch up before killing its transactions.
375         final LeaderFrontendState state = knownFrontends.get(frontend);
376         return state == null ? Optional.absent() : Optional.of(state.getLastConnectTicks());
377     }
378
379     private void onMakeLeaderLocal() {
380         LOG.debug("{}: onMakeLeaderLocal received", persistenceId());
381         if (isLeader()) {
382             getSender().tell(new Status.Success(null), getSelf());
383             return;
384         }
385
386         final ActorSelection leader = getLeader();
387
388         if (leader == null) {
389             // Leader is not present. The cluster is most likely trying to
390             // elect a leader and we should let that run its normal course
391
392             // TODO we can wait for the election to complete and retry the
393             // request. We can also let the caller retry by sending a flag
394             // in the response indicating the request is "reTryable".
395             getSender().tell(new Failure(
396                     new LeadershipTransferFailedException("We cannot initiate leadership transfer to local node. "
397                             + "Currently there is no leader for " + persistenceId())),
398                     getSelf());
399             return;
400         }
401
402         leader.tell(new RequestLeadership(getId(), getSender()), getSelf());
403     }
404
405     // Acquire our frontend tracking handle and verify generation matches
406     @Nullable
407     private LeaderFrontendState findFrontend(final ClientIdentifier clientId) throws RequestException {
408         final LeaderFrontendState existing = knownFrontends.get(clientId.getFrontendId());
409         if (existing != null) {
410             final int cmp = Long.compareUnsigned(existing.getIdentifier().getGeneration(), clientId.getGeneration());
411             if (cmp == 0) {
412                 existing.touch();
413                 return existing;
414             }
415             if (cmp > 0) {
416                 LOG.debug("{}: rejecting request from outdated client {}", persistenceId(), clientId);
417                 throw new RetiredGenerationException(existing.getIdentifier().getGeneration());
418             }
419
420             LOG.info("{}: retiring state {}, outdated by request from client {}", persistenceId(), existing, clientId);
421             existing.retire();
422             knownFrontends.remove(clientId.getFrontendId());
423         } else {
424             LOG.debug("{}: client {} is not yet known", persistenceId(), clientId);
425         }
426
427         return null;
428     }
429
430     private LeaderFrontendState getFrontend(final ClientIdentifier clientId) throws RequestException {
431         final LeaderFrontendState ret = findFrontend(clientId);
432         if (ret != null) {
433             return ret;
434         }
435
436         // TODO: a dedicated exception would be better, but this is technically true, too
437         throw new OutOfSequenceEnvelopeException(0);
438     }
439
440     private static @Nonnull ABIVersion selectVersion(final ConnectClientRequest message) {
441         final Range<ABIVersion> clientRange = Range.closed(message.getMinVersion(), message.getMaxVersion());
442         for (ABIVersion v : SUPPORTED_ABIVERSIONS) {
443             if (clientRange.contains(v)) {
444                 return v;
445             }
446         }
447
448         throw new IllegalArgumentException(String.format(
449             "No common version between backend versions %s and client versions %s", SUPPORTED_ABIVERSIONS,
450             clientRange));
451     }
452
453     @SuppressWarnings("checkstyle:IllegalCatch")
454     private void handleConnectClient(final ConnectClientRequest message) {
455         try {
456             final ClientIdentifier clientId = message.getTarget();
457             final LeaderFrontendState existing = findFrontend(clientId);
458             if (existing != null) {
459                 existing.touch();
460             }
461
462             if (!isLeader() || !isLeaderActive()) {
463                 LOG.info("{}: not currently leader, rejecting request {}. isLeader: {}, isLeaderActive: {},"
464                                 + "isLeadershipTransferInProgress: {}.",
465                         persistenceId(), message, isLeader(), isLeaderActive(), isLeadershipTransferInProgress());
466                 throw new NotLeaderException(getSelf());
467             }
468
469             final ABIVersion selectedVersion = selectVersion(message);
470             final LeaderFrontendState frontend;
471             if (existing == null) {
472                 frontend = new LeaderFrontendState(persistenceId(), clientId, store);
473                 knownFrontends.put(clientId.getFrontendId(), frontend);
474                 LOG.debug("{}: created state {} for client {}", persistenceId(), frontend, clientId);
475             } else {
476                 frontend = existing;
477             }
478
479             frontend.reconnect();
480             message.getReplyTo().tell(new ConnectClientSuccess(message.getTarget(), message.getSequence(), getSelf(),
481                 ImmutableList.of(), store.getDataTree(), CLIENT_MAX_MESSAGES).toVersion(selectedVersion),
482                 ActorRef.noSender());
483         } catch (RequestException | RuntimeException e) {
484             message.getReplyTo().tell(new Failure(e), ActorRef.noSender());
485         }
486     }
487
488     private @Nullable RequestSuccess<?, ?> handleRequest(final RequestEnvelope envelope, final long now)
489             throws RequestException {
490         // We are not the leader, hence we want to fail-fast.
491         if (!isLeader() || paused || !isLeaderActive()) {
492             LOG.debug("{}: not currently active leader, rejecting request {}. isLeader: {}, isLeaderActive: {},"
493                             + "isLeadershipTransferInProgress: {}, paused: {}",
494                     persistenceId(), envelope, isLeader(), isLeaderActive(), isLeadershipTransferInProgress(), paused);
495             throw new NotLeaderException(getSelf());
496         }
497
498         final Request<?, ?> request = envelope.getMessage();
499         if (request instanceof TransactionRequest) {
500             final TransactionRequest<?> txReq = (TransactionRequest<?>)request;
501             final ClientIdentifier clientId = txReq.getTarget().getHistoryId().getClientId();
502             return getFrontend(clientId).handleTransactionRequest(txReq, envelope, now);
503         } else if (request instanceof LocalHistoryRequest) {
504             final LocalHistoryRequest<?> lhReq = (LocalHistoryRequest<?>)request;
505             final ClientIdentifier clientId = lhReq.getTarget().getClientId();
506             return getFrontend(clientId).handleLocalHistoryRequest(lhReq, envelope, now);
507         } else {
508             LOG.warn("{}: rejecting unsupported request {}", persistenceId(), request);
509             throw new UnsupportedRequestException(request);
510         }
511     }
512
513     private boolean hasLeader() {
514         return getLeaderId() != null;
515     }
516
517     public int getPendingTxCommitQueueSize() {
518         return store.getQueueSize();
519     }
520
521     public int getCohortCacheSize() {
522         return commitCoordinator.getCohortCacheSize();
523     }
524
525     @Override
526     protected Optional<ActorRef> getRoleChangeNotifier() {
527         return roleChangeNotifier;
528     }
529
530     @Override
531     protected LeaderStateChanged newLeaderStateChanged(final String memberId, final String leaderId,
532             final short leaderPayloadVersion) {
533         return isLeader() ? new ShardLeaderStateChanged(memberId, leaderId, store.getDataTree(), leaderPayloadVersion)
534                 : new ShardLeaderStateChanged(memberId, leaderId, leaderPayloadVersion);
535     }
536
537     protected void onDatastoreContext(final DatastoreContext context) {
538         datastoreContext = context;
539
540         setTransactionCommitTimeout();
541
542         setPersistence(datastoreContext.isPersistent());
543
544         updateConfigParams(datastoreContext.getShardRaftConfig());
545     }
546
547     // applyState() will be invoked once consensus is reached on the payload
548     void persistPayload(final Identifier id, final Payload payload, final boolean batchHint) {
549         boolean canSkipPayload = !hasFollowers() && !persistence().isRecoveryApplicable();
550         if (canSkipPayload) {
551             applyState(self(), id, payload);
552         } else {
553             // We are faking the sender
554             persistData(self(), id, payload, batchHint);
555         }
556     }
557
558     private void handleCommitTransaction(final CommitTransaction commit) {
559         if (isLeader()) {
560             commitCoordinator.handleCommit(commit.getTransactionId(), getSender(), this);
561         } else {
562             ActorSelection leader = getLeader();
563             if (leader == null) {
564                 messageRetrySupport.addMessageToRetry(commit, getSender(),
565                         "Could not commit transaction " + commit.getTransactionId());
566             } else {
567                 LOG.debug("{}: Forwarding CommitTransaction to leader {}", persistenceId(), leader);
568                 leader.forward(commit, getContext());
569             }
570         }
571     }
572
573     private void handleCanCommitTransaction(final CanCommitTransaction canCommit) {
574         LOG.debug("{}: Can committing transaction {}", persistenceId(), canCommit.getTransactionId());
575
576         if (isLeader()) {
577             commitCoordinator.handleCanCommit(canCommit.getTransactionId(), getSender(), this);
578         } else {
579             ActorSelection leader = getLeader();
580             if (leader == null) {
581                 messageRetrySupport.addMessageToRetry(canCommit, getSender(),
582                         "Could not canCommit transaction " + canCommit.getTransactionId());
583             } else {
584                 LOG.debug("{}: Forwarding CanCommitTransaction to leader {}", persistenceId(), leader);
585                 leader.forward(canCommit, getContext());
586             }
587         }
588     }
589
590     @SuppressWarnings("checkstyle:IllegalCatch")
591     protected void handleBatchedModificationsLocal(final BatchedModifications batched, final ActorRef sender) {
592         try {
593             commitCoordinator.handleBatchedModifications(batched, sender, this);
594         } catch (Exception e) {
595             LOG.error("{}: Error handling BatchedModifications for Tx {}", persistenceId(),
596                     batched.getTransactionId(), e);
597             sender.tell(new Failure(e), getSelf());
598         }
599     }
600
601     private void handleBatchedModifications(final BatchedModifications batched) {
602         // This message is sent to prepare the modifications transaction directly on the Shard as an
603         // optimization to avoid the extra overhead of a separate ShardTransaction actor. On the last
604         // BatchedModifications message, the caller sets the ready flag in the message indicating
605         // modifications are complete. The reply contains the cohort actor path (this actor) for the caller
606         // to initiate the 3-phase commit. This also avoids the overhead of sending an additional
607         // ReadyTransaction message.
608
609         // If we're not the leader then forward to the leader. This is a safety measure - we shouldn't
610         // normally get here if we're not the leader as the front-end (TransactionProxy) should determine
611         // the primary/leader shard. However with timing and caching on the front-end, there's a small
612         // window where it could have a stale leader during leadership transitions.
613         //
614         boolean isLeaderActive = isLeaderActive();
615         if (isLeader() && isLeaderActive) {
616             handleBatchedModificationsLocal(batched, getSender());
617         } else {
618             ActorSelection leader = getLeader();
619             if (!isLeaderActive || leader == null) {
620                 messageRetrySupport.addMessageToRetry(batched, getSender(),
621                         "Could not commit transaction " + batched.getTransactionId());
622             } else {
623                 // If this is not the first batch and leadership changed in between batched messages,
624                 // we need to reconstruct previous BatchedModifications from the transaction
625                 // DataTreeModification, honoring the max batched modification count, and forward all the
626                 // previous BatchedModifications to the new leader.
627                 Collection<BatchedModifications> newModifications = commitCoordinator
628                         .createForwardedBatchedModifications(batched,
629                                 datastoreContext.getShardBatchedModificationCount());
630
631                 LOG.debug("{}: Forwarding {} BatchedModifications to leader {}", persistenceId(),
632                         newModifications.size(), leader);
633
634                 for (BatchedModifications bm : newModifications) {
635                     leader.forward(bm, getContext());
636                 }
637             }
638         }
639     }
640
641     private boolean failIfIsolatedLeader(final ActorRef sender) {
642         if (isIsolatedLeader()) {
643             sender.tell(new Failure(new NoShardLeaderException(String.format(
644                     "Shard %s was the leader but has lost contact with all of its followers. Either all"
645                     + " other follower nodes are down or this node is isolated by a network partition.",
646                     persistenceId()))), getSelf());
647             return true;
648         }
649
650         return false;
651     }
652
653     protected boolean isIsolatedLeader() {
654         return getRaftState() == RaftState.IsolatedLeader;
655     }
656
657     @SuppressWarnings("checkstyle:IllegalCatch")
658     private void handleReadyLocalTransaction(final ReadyLocalTransaction message) {
659         LOG.debug("{}: handleReadyLocalTransaction for {}", persistenceId(), message.getTransactionId());
660
661         boolean isLeaderActive = isLeaderActive();
662         if (isLeader() && isLeaderActive) {
663             try {
664                 commitCoordinator.handleReadyLocalTransaction(message, getSender(), this);
665             } catch (Exception e) {
666                 LOG.error("{}: Error handling ReadyLocalTransaction for Tx {}", persistenceId(),
667                         message.getTransactionId(), e);
668                 getSender().tell(new Failure(e), getSelf());
669             }
670         } else {
671             ActorSelection leader = getLeader();
672             if (!isLeaderActive || leader == null) {
673                 messageRetrySupport.addMessageToRetry(message, getSender(),
674                         "Could not commit transaction " + message.getTransactionId());
675             } else {
676                 LOG.debug("{}: Forwarding ReadyLocalTransaction to leader {}", persistenceId(), leader);
677                 message.setRemoteVersion(getCurrentBehavior().getLeaderPayloadVersion());
678                 leader.forward(message, getContext());
679             }
680         }
681     }
682
683     private void handleForwardedReadyTransaction(final ForwardedReadyTransaction forwardedReady) {
684         LOG.debug("{}: handleForwardedReadyTransaction for {}", persistenceId(), forwardedReady.getTransactionId());
685
686         boolean isLeaderActive = isLeaderActive();
687         if (isLeader() && isLeaderActive) {
688             commitCoordinator.handleForwardedReadyTransaction(forwardedReady, getSender(), this);
689         } else {
690             ActorSelection leader = getLeader();
691             if (!isLeaderActive || leader == null) {
692                 messageRetrySupport.addMessageToRetry(forwardedReady, getSender(),
693                         "Could not commit transaction " + forwardedReady.getTransactionId());
694             } else {
695                 LOG.debug("{}: Forwarding ForwardedReadyTransaction to leader {}", persistenceId(), leader);
696
697                 ReadyLocalTransaction readyLocal = new ReadyLocalTransaction(forwardedReady.getTransactionId(),
698                         forwardedReady.getTransaction().getSnapshot(), forwardedReady.isDoImmediateCommit());
699                 readyLocal.setRemoteVersion(getCurrentBehavior().getLeaderPayloadVersion());
700                 leader.forward(readyLocal, getContext());
701             }
702         }
703     }
704
705     private void handleAbortTransaction(final AbortTransaction abort) {
706         doAbortTransaction(abort.getTransactionId(), getSender());
707     }
708
709     void doAbortTransaction(final Identifier transactionID, final ActorRef sender) {
710         commitCoordinator.handleAbort(transactionID, sender, this);
711     }
712
713     private void handleCreateTransaction(final Object message) {
714         if (isLeader()) {
715             createTransaction(CreateTransaction.fromSerializable(message));
716         } else if (getLeader() != null) {
717             getLeader().forward(message, getContext());
718         } else {
719             getSender().tell(new Failure(new NoShardLeaderException(
720                     "Could not create a shard transaction", persistenceId())), getSelf());
721         }
722     }
723
724     private void closeTransactionChain(final CloseTransactionChain closeTransactionChain) {
725         final LocalHistoryIdentifier id = closeTransactionChain.getIdentifier();
726         store.closeTransactionChain(id, null);
727         store.purgeTransactionChain(id, null);
728     }
729
730     @SuppressWarnings("checkstyle:IllegalCatch")
731     private void createTransaction(final CreateTransaction createTransaction) {
732         try {
733             if (TransactionType.fromInt(createTransaction.getTransactionType()) != TransactionType.READ_ONLY
734                     && failIfIsolatedLeader(getSender())) {
735                 return;
736             }
737
738             ActorRef transactionActor = createTransaction(createTransaction.getTransactionType(),
739                 createTransaction.getTransactionId());
740
741             getSender().tell(new CreateTransactionReply(Serialization.serializedActorPath(transactionActor),
742                     createTransaction.getTransactionId(), createTransaction.getVersion()).toSerializable(), getSelf());
743         } catch (Exception e) {
744             getSender().tell(new Failure(e), getSelf());
745         }
746     }
747
748     private ActorRef createTransaction(final int transactionType, final TransactionIdentifier transactionId) {
749         LOG.debug("{}: Creating transaction : {} ", persistenceId(), transactionId);
750         return transactionActorFactory.newShardTransaction(TransactionType.fromInt(transactionType),
751             transactionId);
752     }
753
754     private void updateSchemaContext(final UpdateSchemaContext message) {
755         updateSchemaContext(message.getSchemaContext());
756     }
757
758     @VisibleForTesting
759     void updateSchemaContext(final SchemaContext schemaContext) {
760         store.updateSchemaContext(schemaContext);
761     }
762
763     private boolean isMetricsCaptureEnabled() {
764         CommonConfig config = new CommonConfig(getContext().system().settings().config());
765         return config.isMetricCaptureEnabled();
766     }
767
768     @Override
769     @VisibleForTesting
770     public RaftActorSnapshotCohort getRaftActorSnapshotCohort() {
771         return snapshotCohort;
772     }
773
774     @Override
775     @Nonnull
776     protected RaftActorRecoveryCohort getRaftActorRecoveryCohort() {
777         return new ShardRecoveryCoordinator(store,
778             restoreFromSnapshot != null ? restoreFromSnapshot.getSnapshot() : null, persistenceId(), LOG);
779     }
780
781     @Override
782     protected void onRecoveryComplete() {
783         restoreFromSnapshot = null;
784
785         //notify shard manager
786         getContext().parent().tell(new ActorInitialized(), getSelf());
787
788         // Being paranoid here - this method should only be called once but just in case...
789         if (txCommitTimeoutCheckSchedule == null) {
790             // Schedule a message to be periodically sent to check if the current in-progress
791             // transaction should be expired and aborted.
792             FiniteDuration period = Duration.create(transactionCommitTimeout / 3, TimeUnit.MILLISECONDS);
793             txCommitTimeoutCheckSchedule = getContext().system().scheduler().schedule(
794                     period, period, getSelf(),
795                     TX_COMMIT_TIMEOUT_CHECK_MESSAGE, getContext().dispatcher(), ActorRef.noSender());
796         }
797     }
798
799     @Override
800     protected void applyState(final ActorRef clientActor, final Identifier identifier, final Object data) {
801         if (data instanceof Payload) {
802             try {
803                 store.applyReplicatedPayload(identifier, (Payload)data);
804             } catch (DataValidationFailedException | IOException e) {
805                 LOG.error("{}: Error applying replica {}", persistenceId(), identifier, e);
806             }
807         } else {
808             LOG.error("{}: Unknown state for {} received {}", persistenceId(), identifier, data);
809         }
810     }
811
812     @Override
813     protected void onStateChanged() {
814         boolean isLeader = isLeader();
815         boolean hasLeader = hasLeader();
816         changeSupport.onLeadershipChange(isLeader, hasLeader);
817         treeChangeSupport.onLeadershipChange(isLeader, hasLeader);
818
819         // If this actor is no longer the leader close all the transaction chains
820         if (!isLeader) {
821             if (LOG.isDebugEnabled()) {
822                 LOG.debug(
823                     "{}: onStateChanged: Closing all transaction chains because shard {} is no longer the leader",
824                     persistenceId(), getId());
825             }
826
827             paused = false;
828             store.purgeLeaderState();
829         }
830
831         if (hasLeader && !isIsolatedLeader()) {
832             messageRetrySupport.retryMessages();
833         }
834     }
835
836     @Override
837     protected void onLeaderChanged(final String oldLeader, final String newLeader) {
838         shardMBean.incrementLeadershipChangeCount();
839         paused = false;
840
841         if (!isLeader()) {
842             if (!knownFrontends.isEmpty()) {
843                 LOG.debug("{}: removing frontend state for {}", persistenceId(), knownFrontends.keySet());
844                 knownFrontends = ImmutableMap.of();
845             }
846
847             if (!hasLeader()) {
848                 // No leader anywhere, nothing else to do
849                 return;
850             }
851
852             // Another leader was elected. If we were the previous leader and had pending transactions, convert
853             // them to transaction messages and send to the new leader.
854             ActorSelection leader = getLeader();
855             if (leader != null) {
856                 Collection<?> messagesToForward = convertPendingTransactionsToMessages();
857
858                 if (!messagesToForward.isEmpty()) {
859                     LOG.debug("{}: Forwarding {} pending transaction messages to leader {}", persistenceId(),
860                             messagesToForward.size(), leader);
861
862                     for (Object message : messagesToForward) {
863                         leader.tell(message, self());
864                     }
865                 }
866             } else {
867                 commitCoordinator.abortPendingTransactions("The transacton was aborted due to inflight leadership "
868                         + "change and the leader address isn't available.", this);
869             }
870         } else {
871             // We have become the leader, we need to reconstruct frontend state
872             knownFrontends = Verify.verifyNotNull(frontendMetadata.toLeaderState(this));
873             LOG.debug("{}: became leader with frontend state for {}", persistenceId(), knownFrontends.keySet());
874         }
875
876         if (!isIsolatedLeader()) {
877             messageRetrySupport.retryMessages();
878         }
879     }
880
881     /**
882      * Clears all pending transactions and converts them to messages to be forwarded to a new leader.
883      *
884      * @return the converted messages
885      */
886     public Collection<?> convertPendingTransactionsToMessages() {
887         return commitCoordinator.convertPendingTransactionsToMessages(
888                 datastoreContext.getShardBatchedModificationCount());
889     }
890
891     @Override
892     protected void pauseLeader(final Runnable operation) {
893         LOG.debug("{}: In pauseLeader, operation: {}", persistenceId(), operation);
894         paused = true;
895
896         // Tell-based protocol can replay transaction state, so it is safe to blow it up when we are paused.
897         knownFrontends.values().forEach(LeaderFrontendState::retire);
898         knownFrontends = ImmutableMap.of();
899
900         store.setRunOnPendingTransactionsComplete(operation);
901     }
902
903     @Override
904     protected void unpauseLeader() {
905         LOG.debug("{}: In unpauseLeader", persistenceId());
906         paused = false;
907
908         store.setRunOnPendingTransactionsComplete(null);
909
910         // Restore tell-based protocol state as if we were becoming the leader
911         knownFrontends = Verify.verifyNotNull(frontendMetadata.toLeaderState(this));
912     }
913
914     @Override
915     protected OnDemandRaftState.AbstractBuilder<?> newOnDemandRaftStateBuilder() {
916         return OnDemandShardState.newBuilder().treeChangeListenerActors(treeChangeSupport.getListenerActors())
917                 .dataChangeListenerActors(changeSupport.getListenerActors())
918                 .commitCohortActors(store.getCohortActors());
919     }
920
921     @Override
922     public String persistenceId() {
923         return this.name;
924     }
925
926     @VisibleForTesting
927     ShardCommitCoordinator getCommitCoordinator() {
928         return commitCoordinator;
929     }
930
931     public DatastoreContext getDatastoreContext() {
932         return datastoreContext;
933     }
934
935     @VisibleForTesting
936     public ShardDataTree getDataStore() {
937         return store;
938     }
939
940     @VisibleForTesting
941     ShardStats getShardMBean() {
942         return shardMBean;
943     }
944
945     public static Builder builder() {
946         return new Builder();
947     }
948
949     public abstract static class AbstractBuilder<T extends AbstractBuilder<T, S>, S extends Shard> {
950         private final Class<S> shardClass;
951         private ShardIdentifier id;
952         private Map<String, String> peerAddresses = Collections.emptyMap();
953         private DatastoreContext datastoreContext;
954         private SchemaContextProvider schemaContextProvider;
955         private DatastoreSnapshot.ShardSnapshot restoreFromSnapshot;
956         private TipProducingDataTree dataTree;
957         private volatile boolean sealed;
958
959         protected AbstractBuilder(final Class<S> shardClass) {
960             this.shardClass = shardClass;
961         }
962
963         protected void checkSealed() {
964             Preconditions.checkState(!sealed, "Builder isalready sealed - further modifications are not allowed");
965         }
966
967         @SuppressWarnings("unchecked")
968         private T self() {
969             return (T) this;
970         }
971
972         public T id(final ShardIdentifier newId) {
973             checkSealed();
974             this.id = newId;
975             return self();
976         }
977
978         public T peerAddresses(final Map<String, String> newPeerAddresses) {
979             checkSealed();
980             this.peerAddresses = newPeerAddresses;
981             return self();
982         }
983
984         public T datastoreContext(final DatastoreContext newDatastoreContext) {
985             checkSealed();
986             this.datastoreContext = newDatastoreContext;
987             return self();
988         }
989
990         public T schemaContextProvider(final SchemaContextProvider schemaContextProvider) {
991             checkSealed();
992             this.schemaContextProvider = Preconditions.checkNotNull(schemaContextProvider);
993             return self();
994         }
995
996         public T restoreFromSnapshot(final DatastoreSnapshot.ShardSnapshot newRestoreFromSnapshot) {
997             checkSealed();
998             this.restoreFromSnapshot = newRestoreFromSnapshot;
999             return self();
1000         }
1001
1002         public T dataTree(final TipProducingDataTree newDataTree) {
1003             checkSealed();
1004             this.dataTree = newDataTree;
1005             return self();
1006         }
1007
1008         public ShardIdentifier getId() {
1009             return id;
1010         }
1011
1012         public Map<String, String> getPeerAddresses() {
1013             return peerAddresses;
1014         }
1015
1016         public DatastoreContext getDatastoreContext() {
1017             return datastoreContext;
1018         }
1019
1020         public SchemaContext getSchemaContext() {
1021             return Verify.verifyNotNull(schemaContextProvider.getSchemaContext());
1022         }
1023
1024         public DatastoreSnapshot.ShardSnapshot getRestoreFromSnapshot() {
1025             return restoreFromSnapshot;
1026         }
1027
1028         public TipProducingDataTree getDataTree() {
1029             return dataTree;
1030         }
1031
1032         public TreeType getTreeType() {
1033             switch (datastoreContext.getLogicalStoreType()) {
1034                 case CONFIGURATION:
1035                     return TreeType.CONFIGURATION;
1036                 case OPERATIONAL:
1037                     return TreeType.OPERATIONAL;
1038                 default:
1039                     throw new IllegalStateException("Unhandled logical store type "
1040                             + datastoreContext.getLogicalStoreType());
1041             }
1042         }
1043
1044         protected void verify() {
1045             Preconditions.checkNotNull(id, "id should not be null");
1046             Preconditions.checkNotNull(peerAddresses, "peerAddresses should not be null");
1047             Preconditions.checkNotNull(datastoreContext, "dataStoreContext should not be null");
1048             Preconditions.checkNotNull(schemaContextProvider, "schemaContextProvider should not be null");
1049         }
1050
1051         public Props props() {
1052             sealed = true;
1053             verify();
1054             return Props.create(shardClass, this);
1055         }
1056     }
1057
1058     public static class Builder extends AbstractBuilder<Builder, Shard> {
1059         private Builder() {
1060             super(Shard.class);
1061         }
1062     }
1063
1064     Ticker ticker() {
1065         return Ticker.systemTicker();
1066     }
1067
1068     void scheduleNextPendingTransaction() {
1069         self().tell(RESUME_NEXT_PENDING_TRANSACTION, ActorRef.noSender());
1070     }
1071 }