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 package org.opendaylight.controller.cluster.datastore;
   9
  10 import static com.google.common.base.Preconditions.checkState;
  11 import static com.google.common.base.Verify.verify;
  12 import static com.google.common.base.Verify.verifyNotNull;
  13 import static java.util.Objects.requireNonNull;
  14
  15 import akka.actor.ActorRef;
  16 import akka.actor.ActorSelection;
  17 import akka.actor.Cancellable;
  18 import akka.actor.ExtendedActorSystem;
  19 import akka.actor.PoisonPill;
  20 import akka.actor.Props;
  21 import akka.actor.Status;
  22 import akka.actor.Status.Failure;
  23 import akka.persistence.RecoveryCompleted;
  24 import akka.persistence.SnapshotOffer;
  25 import akka.serialization.JavaSerializer;
  26 import akka.serialization.Serialization;
  27 import com.google.common.annotations.VisibleForTesting;
  28 import com.google.common.base.Ticker;
  29 import com.google.common.collect.ImmutableList;
  30 import com.google.common.collect.ImmutableMap;
  31 import com.google.common.collect.ImmutableSet;
  32 import com.google.common.collect.Range;
  33 import java.io.IOException;
  34 import java.util.Arrays;
  35 import java.util.Collection;
  36 import java.util.Collections;
  37 import java.util.HashMap;
  38 import java.util.Map;
  39 import java.util.Optional;
  40 import java.util.OptionalLong;
  41 import java.util.concurrent.TimeUnit;
  42 import org.eclipse.jdt.annotation.NonNull;
  43 import org.eclipse.jdt.annotation.Nullable;
  44 import org.opendaylight.controller.cluster.access.ABIVersion;
  45 import org.opendaylight.controller.cluster.access.commands.ConnectClientRequest;
  46 import org.opendaylight.controller.cluster.access.commands.ConnectClientSuccess;
  47 import org.opendaylight.controller.cluster.access.commands.LocalHistoryRequest;
  48 import org.opendaylight.controller.cluster.access.commands.NotLeaderException;
  49 import org.opendaylight.controller.cluster.access.commands.OutOfSequenceEnvelopeException;
  50 import org.opendaylight.controller.cluster.access.commands.TransactionRequest;
  51 import org.opendaylight.controller.cluster.access.concepts.ClientIdentifier;
  52 import org.opendaylight.controller.cluster.access.concepts.FrontendIdentifier;
  53 import org.opendaylight.controller.cluster.access.concepts.LocalHistoryIdentifier;
  54 import org.opendaylight.controller.cluster.access.concepts.Request;
  55 import org.opendaylight.controller.cluster.access.concepts.RequestEnvelope;
  56 import org.opendaylight.controller.cluster.access.concepts.RequestException;
  57 import org.opendaylight.controller.cluster.access.concepts.RequestSuccess;
  58 import org.opendaylight.controller.cluster.access.concepts.RetiredGenerationException;
  59 import org.opendaylight.controller.cluster.access.concepts.RuntimeRequestException;
  60 import org.opendaylight.controller.cluster.access.concepts.SliceableMessage;
  61 import org.opendaylight.controller.cluster.access.concepts.TransactionIdentifier;
  62 import org.opendaylight.controller.cluster.access.concepts.UnsupportedRequestException;
  63 import org.opendaylight.controller.cluster.common.actor.CommonConfig;
  64 import org.opendaylight.controller.cluster.common.actor.Dispatchers;
  65 import org.opendaylight.controller.cluster.common.actor.Dispatchers.DispatcherType;
  66 import org.opendaylight.controller.cluster.common.actor.MessageTracker;
  67 import org.opendaylight.controller.cluster.common.actor.MessageTracker.Error;
  68 import org.opendaylight.controller.cluster.common.actor.MeteringBehavior;
  69 import org.opendaylight.controller.cluster.datastore.actors.JsonExportActor;
  70 import org.opendaylight.controller.cluster.datastore.exceptions.NoShardLeaderException;
  71 import org.opendaylight.controller.cluster.datastore.identifiers.ShardIdentifier;
  72 import org.opendaylight.controller.cluster.datastore.messages.AbortTransaction;
  73 import org.opendaylight.controller.cluster.datastore.messages.ActorInitialized;
  74 import org.opendaylight.controller.cluster.datastore.messages.BatchedModifications;
  75 import org.opendaylight.controller.cluster.datastore.messages.CanCommitTransaction;
  76 import org.opendaylight.controller.cluster.datastore.messages.CloseTransactionChain;
  77 import org.opendaylight.controller.cluster.datastore.messages.CommitTransaction;
  78 import org.opendaylight.controller.cluster.datastore.messages.CreateTransaction;
  79 import org.opendaylight.controller.cluster.datastore.messages.CreateTransactionReply;
  80 import org.opendaylight.controller.cluster.datastore.messages.DataTreeChangedReply;
  81 import org.opendaylight.controller.cluster.datastore.messages.ForwardedReadyTransaction;
  82 import org.opendaylight.controller.cluster.datastore.messages.GetKnownClients;
  83 import org.opendaylight.controller.cluster.datastore.messages.GetKnownClientsReply;
  84 import org.opendaylight.controller.cluster.datastore.messages.GetShardDataTree;
  85 import org.opendaylight.controller.cluster.datastore.messages.MakeLeaderLocal;
  86 import org.opendaylight.controller.cluster.datastore.messages.OnDemandShardState;
  87 import org.opendaylight.controller.cluster.datastore.messages.PeerAddressResolved;
  88 import org.opendaylight.controller.cluster.datastore.messages.ReadyLocalTransaction;
  89 import org.opendaylight.controller.cluster.datastore.messages.RegisterDataTreeChangeListener;
  90 import org.opendaylight.controller.cluster.datastore.messages.ShardLeaderStateChanged;
  91 import org.opendaylight.controller.cluster.datastore.messages.UpdateSchemaContext;
  92 import org.opendaylight.controller.cluster.datastore.persisted.DatastoreSnapshot;
  93 import org.opendaylight.controller.cluster.datastore.persisted.DatastoreSnapshot.ShardSnapshot;
  94 import org.opendaylight.controller.cluster.datastore.persisted.DisableTrackingPayload;
  95 import org.opendaylight.controller.cluster.messaging.MessageAssembler;
  96 import org.opendaylight.controller.cluster.messaging.MessageSlicer;
  97 import org.opendaylight.controller.cluster.messaging.SliceOptions;
  98 import org.opendaylight.controller.cluster.notifications.LeaderStateChanged;
  99 import org.opendaylight.controller.cluster.notifications.RegisterRoleChangeListener;
 100 import org.opendaylight.controller.cluster.notifications.RoleChangeNotifier;
 101 import org.opendaylight.controller.cluster.raft.LeadershipTransferFailedException;
 102 import org.opendaylight.controller.cluster.raft.RaftActor;
 103 import org.opendaylight.controller.cluster.raft.RaftActorRecoveryCohort;
 104 import org.opendaylight.controller.cluster.raft.RaftActorSnapshotCohort;
 105 import org.opendaylight.controller.cluster.raft.RaftState;
 106 import org.opendaylight.controller.cluster.raft.ReplicatedLogEntry;
 107 import org.opendaylight.controller.cluster.raft.base.messages.FollowerInitialSyncUpStatus;
 108 import org.opendaylight.controller.cluster.raft.client.messages.OnDemandRaftState;
 109 import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
 110 import org.opendaylight.controller.cluster.raft.messages.RequestLeadership;
 111 import org.opendaylight.controller.cluster.raft.messages.ServerRemoved;
 112 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload;
 113 import org.opendaylight.yang.gen.v1.urn.opendaylight.params.xml.ns.yang.controller.config.distributed.datastore.provider.rev140612.DataStoreProperties.ExportOnRecovery;
 114 import org.opendaylight.yangtools.concepts.Identifier;
 115 import org.opendaylight.yangtools.yang.data.api.schema.tree.DataTree;
 116 import org.opendaylight.yangtools.yang.data.api.schema.tree.DataValidationFailedException;
 117 import org.opendaylight.yangtools.yang.data.api.schema.tree.TreeType;
 118 import org.opendaylight.yangtools.yang.model.api.EffectiveModelContext;
 119 import org.opendaylight.yangtools.yang.model.api.EffectiveModelContextProvider;
 120 import scala.concurrent.duration.FiniteDuration;
 121
 122 /**
 123  * A Shard represents a portion of the logical data tree.
 124  *
 125  * <p>
 126  * Our Shard uses InMemoryDataTree as it's internal representation and delegates all requests it
 127  */
 128 // FIXME: non-final for testing?
 129 public class Shard extends RaftActor {
 130
 131     @VisibleForTesting
 132     static final Object TX_COMMIT_TIMEOUT_CHECK_MESSAGE = new Object() {
 133         @Override
 134         public String toString() {
 135             return "txCommitTimeoutCheck";
 136         }
 137     };
 138
 139     @VisibleForTesting
 140     static final Object GET_SHARD_MBEAN_MESSAGE = new Object() {
 141         @Override
 142         public String toString() {
 143             return "getShardMBeanMessage";
 144         }
 145     };
 146
 147     static final Object RESUME_NEXT_PENDING_TRANSACTION = new Object() {
 148         @Override
 149         public String toString() {
 150             return "resumeNextPendingTransaction";
 151         }
 152     };
 153
 154     // FIXME: shard names should be encapsulated in their own class and this should be exposed as a constant.
 155     public static final String DEFAULT_NAME = "default";
 156
 157     private static final Collection<ABIVersion> SUPPORTED_ABIVERSIONS;
 158
 159     // Make sure to keep this in sync with the journal configuration in factory-akka.conf
 160     public static final String NON_PERSISTENT_JOURNAL_ID = "akka.persistence.non-persistent.journal";
 161
 162     static {
 163         final ABIVersion[] values = ABIVersion.values();
 164         final ABIVersion[] real = Arrays.copyOfRange(values, 1, values.length - 1);
 165         SUPPORTED_ABIVERSIONS = ImmutableList.copyOf(real).reverse();
 166     }
 167
 168     // FIXME: make this a dynamic property based on mailbox size and maximum number of clients
 169     private static final int CLIENT_MAX_MESSAGES = 1000;
 170
 171     // The state of this Shard
 172     private final ShardDataTree store;
 173
 174     /// The name of this shard
 175     private final String name;
 176
 177     private final String shardName;
 178
 179     private final ShardStats shardMBean;
 180
 181     private final ShardDataTreeListenerInfoMXBeanImpl listenerInfoMXBean;
 182
 183     private DatastoreContext datastoreContext;
 184
 185     private final ShardCommitCoordinator commitCoordinator;
 186
 187     private long transactionCommitTimeout;
 188
 189     private Cancellable txCommitTimeoutCheckSchedule;
 190
 191     private final Optional<ActorRef> roleChangeNotifier;
 192
 193     private final MessageTracker appendEntriesReplyTracker;
 194
 195     private final ShardTransactionActorFactory transactionActorFactory;
 196
 197     private final ShardSnapshotCohort snapshotCohort;
 198
 199     private final DataTreeChangeListenerSupport treeChangeSupport = new DataTreeChangeListenerSupport(this);
 200
 201     private ShardSnapshot restoreFromSnapshot;
 202
 203     private final ShardTransactionMessageRetrySupport messageRetrySupport;
 204
 205     @VisibleForTesting
 206     final FrontendMetadata frontendMetadata;
 207
 208     private Map<FrontendIdentifier, LeaderFrontendState> knownFrontends = ImmutableMap.of();
 209     private boolean paused;
 210
 211     private final MessageSlicer responseMessageSlicer;
 212     private final Dispatchers dispatchers;
 213
 214     private final MessageAssembler requestMessageAssembler;
 215
 216     private final ExportOnRecovery exportOnRecovery;
 217
 218     private final ActorRef exportActor;
 219
 220     Shard(final AbstractBuilder<?, ?> builder) {
 221         super(builder.getId().toString(), builder.getPeerAddresses(),
 222                 Optional.of(builder.getDatastoreContext().getShardRaftConfig()), DataStoreVersions.CURRENT_VERSION);
 223
 224         name = builder.getId().toString();
 225         shardName = builder.getId().getShardName();
 226         datastoreContext = builder.getDatastoreContext();
 227         restoreFromSnapshot = builder.getRestoreFromSnapshot();
 228         frontendMetadata = new FrontendMetadata(name);
 229         exportOnRecovery = datastoreContext.getExportOnRecovery();
 230
 231         switch (exportOnRecovery) {
 232             case Json:
 233                 exportActor = getContext().actorOf(JsonExportActor.props(builder.getSchemaContext(),
 234                         datastoreContext.getRecoveryExportBaseDir()));
 235                 break;
 236             case Off:
 237             default:
 238                 exportActor = null;
 239                 break;
 240         }
 241
 242         setPersistence(datastoreContext.isPersistent());
 243
 244         LOG.info("Shard created : {}, persistent : {}", name, datastoreContext.isPersistent());
 245
 246         ShardDataTreeChangeListenerPublisherActorProxy treeChangeListenerPublisher =
 247                 new ShardDataTreeChangeListenerPublisherActorProxy(getContext(), name + "-DTCL-publisher", name);
 248         if (builder.getDataTree() != null) {
 249             store = new ShardDataTree(this, builder.getSchemaContext(), builder.getDataTree(),
 250                     treeChangeListenerPublisher, name,
 251                     frontendMetadata);
 252         } else {
 253             store = new ShardDataTree(this, builder.getSchemaContext(), builder.getTreeType(),
 254                     builder.getDatastoreContext().getStoreRoot(), treeChangeListenerPublisher, name,
 255                     frontendMetadata);
 256         }
 257
 258         shardMBean = ShardStats.create(name, datastoreContext.getDataStoreMXBeanType(), this);
 259
 260         if (isMetricsCaptureEnabled()) {
 261             getContext().become(new MeteringBehavior(this));
 262         }
 263
 264         commitCoordinator = new ShardCommitCoordinator(store, LOG, name);
 265
 266         setTransactionCommitTimeout();
 267
 268         // create a notifier actor for each cluster member
 269         roleChangeNotifier = createRoleChangeNotifier(name);
 270
 271         appendEntriesReplyTracker = new MessageTracker(AppendEntriesReply.class,
 272                 getRaftActorContext().getConfigParams().getIsolatedCheckIntervalInMillis());
 273
 274         dispatchers = new Dispatchers(context().system().dispatchers());
 275         transactionActorFactory = new ShardTransactionActorFactory(store, datastoreContext,
 276             dispatchers.getDispatcherPath(Dispatchers.DispatcherType.Transaction),
 277                 self(), getContext(), shardMBean, builder.getId().getShardName());
 278
 279         snapshotCohort = ShardSnapshotCohort.create(getContext(), builder.getId().getMemberName(), store, LOG,
 280             name, datastoreContext);
 281
 282         messageRetrySupport = new ShardTransactionMessageRetrySupport(this);
 283
 284         responseMessageSlicer = MessageSlicer.builder().logContext(name)
 285                 .messageSliceSize(datastoreContext.getMaximumMessageSliceSize())
 286                 .fileBackedStreamFactory(getRaftActorContext().getFileBackedOutputStreamFactory())
 287                 .expireStateAfterInactivity(2, TimeUnit.MINUTES).build();
 288
 289         requestMessageAssembler = MessageAssembler.builder().logContext(name)
 290                 .fileBackedStreamFactory(getRaftActorContext().getFileBackedOutputStreamFactory())
 291                 .assembledMessageCallback((message, sender) -> self().tell(message, sender))
 292                 .expireStateAfterInactivity(datastoreContext.getRequestTimeout(), TimeUnit.NANOSECONDS).build();
 293
 294         listenerInfoMXBean = new ShardDataTreeListenerInfoMXBeanImpl(name, datastoreContext.getDataStoreMXBeanType(),
 295                 self());
 296         listenerInfoMXBean.register();
 297     }
 298
 299     private void setTransactionCommitTimeout() {
 300         transactionCommitTimeout = TimeUnit.MILLISECONDS.convert(
 301                 datastoreContext.getShardTransactionCommitTimeoutInSeconds(), TimeUnit.SECONDS) / 2;
 302     }
 303
 304     private Optional<ActorRef> createRoleChangeNotifier(final String shardId) {
 305         ActorRef shardRoleChangeNotifier = this.getContext().actorOf(
 306             RoleChangeNotifier.getProps(shardId), shardId + "-notifier");
 307         return Optional.of(shardRoleChangeNotifier);
 308     }
 309
 310     @Override
 311     public final void postStop() throws Exception {
 312         LOG.info("Stopping Shard {}", persistenceId());
 313
 314         super.postStop();
 315
 316         messageRetrySupport.close();
 317
 318         if (txCommitTimeoutCheckSchedule != null) {
 319             txCommitTimeoutCheckSchedule.cancel();
 320         }
 321
 322         commitCoordinator.abortPendingTransactions("Transaction aborted due to shutdown.", this);
 323
 324         shardMBean.unregisterMBean();
 325         listenerInfoMXBean.unregister();
 326     }
 327
 328     @Override
 329     protected final void handleRecover(final Object message) {
 330         LOG.debug("{}: onReceiveRecover: Received message {} from {}", persistenceId(), message.getClass(),
 331             getSender());
 332
 333         super.handleRecover(message);
 334
 335         switch (exportOnRecovery) {
 336             case Json:
 337                 if (message instanceof SnapshotOffer) {
 338                     exportActor.tell(new JsonExportActor.ExportSnapshot(store.readCurrentData().get(), name),
 339                             ActorRef.noSender());
 340                 } else if (message instanceof ReplicatedLogEntry) {
 341                     exportActor.tell(new JsonExportActor.ExportJournal((ReplicatedLogEntry) message),
 342                             ActorRef.noSender());
 343                 } else if (message instanceof RecoveryCompleted) {
 344                     exportActor.tell(new JsonExportActor.FinishExport(name), ActorRef.noSender());
 345                     exportActor.tell(PoisonPill.getInstance(), ActorRef.noSender());
 346                 }
 347                 break;
 348             case Off:
 349             default:
 350                 break;
 351         }
 352
 353         if (LOG.isTraceEnabled()) {
 354             appendEntriesReplyTracker.begin();
 355         }
 356     }
 357
 358     @Override
 359     // non-final for TestShard
 360     protected void handleNonRaftCommand(final Object message) {
 361         try (MessageTracker.Context context = appendEntriesReplyTracker.received(message)) {
 362             final Optional<Error> maybeError = context.error();
 363             if (maybeError.isPresent()) {
 364                 LOG.trace("{} : AppendEntriesReply failed to arrive at the expected interval {}", persistenceId(),
 365                     maybeError.get());
 366             }
 367
 368             store.resetTransactionBatch();
 369
 370             if (message instanceof RequestEnvelope) {
 371                 handleRequestEnvelope((RequestEnvelope)message);
 372             } else if (MessageAssembler.isHandledMessage(message)) {
 373                 handleRequestAssemblerMessage(message);
 374             } else if (message instanceof ConnectClientRequest) {
 375                 handleConnectClient((ConnectClientRequest)message);
 376             } else if (CreateTransaction.isSerializedType(message)) {
 377                 handleCreateTransaction(message);
 378             } else if (message instanceof BatchedModifications) {
 379                 handleBatchedModifications((BatchedModifications)message);
 380             } else if (message instanceof ForwardedReadyTransaction) {
 381                 handleForwardedReadyTransaction((ForwardedReadyTransaction) message);
 382             } else if (message instanceof ReadyLocalTransaction) {
 383                 handleReadyLocalTransaction((ReadyLocalTransaction)message);
 384             } else if (CanCommitTransaction.isSerializedType(message)) {
 385                 handleCanCommitTransaction(CanCommitTransaction.fromSerializable(message));
 386             } else if (CommitTransaction.isSerializedType(message)) {
 387                 handleCommitTransaction(CommitTransaction.fromSerializable(message));
 388             } else if (AbortTransaction.isSerializedType(message)) {
 389                 handleAbortTransaction(AbortTransaction.fromSerializable(message));
 390             } else if (CloseTransactionChain.isSerializedType(message)) {
 391                 closeTransactionChain(CloseTransactionChain.fromSerializable(message));
 392             } else if (message instanceof DataTreeChangedReply) {
 393                 // Ignore reply
 394             } else if (message instanceof RegisterDataTreeChangeListener) {
 395                 treeChangeSupport.onMessage((RegisterDataTreeChangeListener) message, isLeader(), hasLeader());
 396             } else if (message instanceof UpdateSchemaContext) {
 397                 updateSchemaContext((UpdateSchemaContext) message);
 398             } else if (message instanceof PeerAddressResolved) {
 399                 PeerAddressResolved resolved = (PeerAddressResolved) message;
 400                 setPeerAddress(resolved.getPeerId(), resolved.getPeerAddress());
 401             } else if (TX_COMMIT_TIMEOUT_CHECK_MESSAGE.equals(message)) {
 402                 commitTimeoutCheck();
 403             } else if (message instanceof DatastoreContext) {
 404                 onDatastoreContext((DatastoreContext)message);
 405             } else if (message instanceof RegisterRoleChangeListener) {
 406                 roleChangeNotifier.get().forward(message, context());
 407             } else if (message instanceof FollowerInitialSyncUpStatus) {
 408                 shardMBean.setFollowerInitialSyncStatus(((FollowerInitialSyncUpStatus) message).isInitialSyncDone());
 409                 context().parent().tell(message, self());
 410             } else if (GET_SHARD_MBEAN_MESSAGE.equals(message)) {
 411                 sender().tell(getShardMBean(), self());
 412             } else if (message instanceof GetShardDataTree) {
 413                 sender().tell(store.getDataTree(), self());
 414             } else if (message instanceof ServerRemoved) {
 415                 context().parent().forward(message, context());
 416             } else if (ShardTransactionMessageRetrySupport.TIMER_MESSAGE_CLASS.isInstance(message)) {
 417                 messageRetrySupport.onTimerMessage(message);
 418             } else if (message instanceof DataTreeCohortActorRegistry.CohortRegistryCommand) {
 419                 store.processCohortRegistryCommand(getSender(),
 420                         (DataTreeCohortActorRegistry.CohortRegistryCommand) message);
 421             } else if (message instanceof MakeLeaderLocal) {
 422                 onMakeLeaderLocal();
 423             } else if (RESUME_NEXT_PENDING_TRANSACTION.equals(message)) {
 424                 store.resumeNextPendingTransaction();
 425             } else if (GetKnownClients.INSTANCE.equals(message)) {
 426                 handleGetKnownClients();
 427             } else if (!responseMessageSlicer.handleMessage(message)) {
 428                 super.handleNonRaftCommand(message);
 429             }
 430         }
 431     }
 432
 433     private void handleRequestAssemblerMessage(final Object message) {
 434         dispatchers.getDispatcher(DispatcherType.Serialization).execute(() -> {
 435             JavaSerializer.currentSystem().value_$eq((ExtendedActorSystem) context().system());
 436             requestMessageAssembler.handleMessage(message, self());
 437         });
 438     }
 439
 440     @SuppressWarnings("checkstyle:IllegalCatch")
 441     private void handleRequestEnvelope(final RequestEnvelope envelope) {
 442         final long now = ticker().read();
 443         try {
 444             final RequestSuccess<?, ?> success = handleRequest(envelope, now);
 445             if (success != null) {
 446                 final long executionTimeNanos = ticker().read() - now;
 447                 if (success instanceof SliceableMessage) {
 448                     dispatchers.getDispatcher(DispatcherType.Serialization).execute(() ->
 449                         responseMessageSlicer.slice(SliceOptions.builder().identifier(success.getTarget())
 450                             .message(envelope.newSuccessEnvelope(success, executionTimeNanos))
 451                             .sendTo(envelope.getMessage().getReplyTo()).replyTo(self())
 452                             .onFailureCallback(t -> LOG.warn("Error slicing response {}", success, t)).build()));
 453                 } else {
 454                     envelope.sendSuccess(success, executionTimeNanos);
 455                 }
 456             }
 457         } catch (RequestException e) {
 458             LOG.debug("{}: request {} failed", persistenceId(), envelope, e);
 459             envelope.sendFailure(e, ticker().read() - now);
 460         } catch (Exception e) {
 461             LOG.debug("{}: request {} caused failure", persistenceId(), envelope, e);
 462             envelope.sendFailure(new RuntimeRequestException("Request failed to process", e),
 463                 ticker().read() - now);
 464         }
 465     }
 466
 467     private void commitTimeoutCheck() {
 468         store.checkForExpiredTransactions(transactionCommitTimeout, this::updateAccess);
 469         commitCoordinator.checkForExpiredTransactions(transactionCommitTimeout, this);
 470         requestMessageAssembler.checkExpiredAssembledMessageState();
 471     }
 472
 473     private OptionalLong updateAccess(final SimpleShardDataTreeCohort cohort) {
 474         final FrontendIdentifier frontend = cohort.getIdentifier().getHistoryId().getClientId().getFrontendId();
 475         final LeaderFrontendState state = knownFrontends.get(frontend);
 476         if (state == null) {
 477             // Not tell-based protocol, do nothing
 478             return OptionalLong.empty();
 479         }
 480
 481         if (isIsolatedLeader()) {
 482             // We are isolated and no new request can come through until we emerge from it. We are still updating
 483             // liveness of frontend when we see it attempting to communicate. Use the last access timer.
 484             return OptionalLong.of(state.getLastSeenTicks());
 485         }
 486
 487         // If this frontend has freshly connected, give it some time to catch up before killing its transactions.
 488         return OptionalLong.of(state.getLastConnectTicks());
 489     }
 490
 491     private void disableTracking(final DisableTrackingPayload payload) {
 492         final ClientIdentifier clientId = payload.getIdentifier();
 493         LOG.debug("{}: disabling tracking of {}", persistenceId(), clientId);
 494         frontendMetadata.disableTracking(clientId);
 495
 496         if (isLeader()) {
 497             final FrontendIdentifier frontendId = clientId.getFrontendId();
 498             final LeaderFrontendState frontend = knownFrontends.get(frontendId);
 499             if (frontend != null) {
 500                 if (clientId.equals(frontend.getIdentifier())) {
 501                     if (!(frontend instanceof LeaderFrontendState.Disabled)) {
 502                         verify(knownFrontends.replace(frontendId, frontend,
 503                             new LeaderFrontendState.Disabled(persistenceId(), clientId, store)));
 504                         LOG.debug("{}: leader state for {} disabled", persistenceId(), clientId);
 505                     } else {
 506                         LOG.debug("{}: leader state {} is already disabled", persistenceId(), frontend);
 507                     }
 508                 } else {
 509                     LOG.debug("{}: leader state {} does not match {}", persistenceId(), frontend, clientId);
 510                 }
 511             } else {
 512                 LOG.debug("{}: leader state for {} not found", persistenceId(), clientId);
 513                 knownFrontends.put(frontendId, new LeaderFrontendState.Disabled(persistenceId(), clientId,
 514                     getDataStore()));
 515             }
 516         }
 517     }
 518
 519     private void onMakeLeaderLocal() {
 520         LOG.debug("{}: onMakeLeaderLocal received", persistenceId());
 521         if (isLeader()) {
 522             getSender().tell(new Status.Success(null), getSelf());
 523             return;
 524         }
 525
 526         final ActorSelection leader = getLeader();
 527
 528         if (leader == null) {
 529             // Leader is not present. The cluster is most likely trying to
 530             // elect a leader and we should let that run its normal course
 531
 532             // TODO we can wait for the election to complete and retry the
 533             // request. We can also let the caller retry by sending a flag
 534             // in the response indicating the request is "reTryable".
 535             getSender().tell(new Failure(
 536                     new LeadershipTransferFailedException("We cannot initiate leadership transfer to local node. "
 537                             + "Currently there is no leader for " + persistenceId())),
 538                     getSelf());
 539             return;
 540         }
 541
 542         leader.tell(new RequestLeadership(getId(), getSender()), getSelf());
 543     }
 544
 545     // Acquire our frontend tracking handle and verify generation matches
 546     private @Nullable LeaderFrontendState findFrontend(final ClientIdentifier clientId) throws RequestException {
 547         final LeaderFrontendState existing = knownFrontends.get(clientId.getFrontendId());
 548         if (existing != null) {
 549             final int cmp = Long.compareUnsigned(existing.getIdentifier().getGeneration(), clientId.getGeneration());
 550             if (cmp == 0) {
 551                 existing.touch();
 552                 return existing;
 553             }
 554             if (cmp > 0) {
 555                 LOG.debug("{}: rejecting request from outdated client {}", persistenceId(), clientId);
 556                 throw new RetiredGenerationException(clientId.getGeneration(),
 557                     existing.getIdentifier().getGeneration());
 558             }
 559
 560             LOG.info("{}: retiring state {}, outdated by request from client {}", persistenceId(), existing, clientId);
 561             existing.retire();
 562             knownFrontends.remove(clientId.getFrontendId());
 563         } else {
 564             LOG.debug("{}: client {} is not yet known", persistenceId(), clientId);
 565         }
 566
 567         return null;
 568     }
 569
 570     private LeaderFrontendState getFrontend(final ClientIdentifier clientId) throws RequestException {
 571         final LeaderFrontendState ret = findFrontend(clientId);
 572         if (ret != null) {
 573             return ret;
 574         }
 575
 576         // TODO: a dedicated exception would be better, but this is technically true, too
 577         throw new OutOfSequenceEnvelopeException(0);
 578     }
 579
 580     private static @NonNull ABIVersion selectVersion(final ConnectClientRequest message) {
 581         final Range<ABIVersion> clientRange = Range.closed(message.getMinVersion(), message.getMaxVersion());
 582         for (ABIVersion v : SUPPORTED_ABIVERSIONS) {
 583             if (clientRange.contains(v)) {
 584                 return v;
 585             }
 586         }
 587
 588         throw new IllegalArgumentException(String.format(
 589             "No common version between backend versions %s and client versions %s", SUPPORTED_ABIVERSIONS,
 590             clientRange));
 591     }
 592
 593     @SuppressWarnings("checkstyle:IllegalCatch")
 594     private void handleConnectClient(final ConnectClientRequest message) {
 595         try {
 596             final ClientIdentifier clientId = message.getTarget();
 597             final LeaderFrontendState existing = findFrontend(clientId);
 598             if (existing != null) {
 599                 existing.touch();
 600             }
 601
 602             if (!isLeader() || !isLeaderActive()) {
 603                 LOG.info("{}: not currently leader, rejecting request {}. isLeader: {}, isLeaderActive: {},"
 604                                 + "isLeadershipTransferInProgress: {}.",
 605                         persistenceId(), message, isLeader(), isLeaderActive(), isLeadershipTransferInProgress());
 606                 throw new NotLeaderException(getSelf());
 607             }
 608
 609             final ABIVersion selectedVersion = selectVersion(message);
 610             final LeaderFrontendState frontend;
 611             if (existing == null) {
 612                 frontend = new LeaderFrontendState.Enabled(persistenceId(), clientId, store);
 613                 knownFrontends.put(clientId.getFrontendId(), frontend);
 614                 LOG.debug("{}: created state {} for client {}", persistenceId(), frontend, clientId);
 615             } else {
 616                 frontend = existing;
 617             }
 618
 619             frontend.reconnect();
 620             message.getReplyTo().tell(new ConnectClientSuccess(message.getTarget(), message.getSequence(), getSelf(),
 621                 ImmutableList.of(), store.getDataTree(), CLIENT_MAX_MESSAGES).toVersion(selectedVersion),
 622                 ActorRef.noSender());
 623         } catch (RequestException | RuntimeException e) {
 624             message.getReplyTo().tell(new Failure(e), ActorRef.noSender());
 625         }
 626     }
 627
 628     private @Nullable RequestSuccess<?, ?> handleRequest(final RequestEnvelope envelope, final long now)
 629             throws RequestException {
 630         // We are not the leader, hence we want to fail-fast.
 631         if (!isLeader() || paused || !isLeaderActive()) {
 632             LOG.debug("{}: not currently active leader, rejecting request {}. isLeader: {}, isLeaderActive: {},"
 633                             + "isLeadershipTransferInProgress: {}, paused: {}",
 634                     persistenceId(), envelope, isLeader(), isLeaderActive(), isLeadershipTransferInProgress(), paused);
 635             throw new NotLeaderException(getSelf());
 636         }
 637
 638         final Request<?, ?> request = envelope.getMessage();
 639         if (request instanceof TransactionRequest) {
 640             final TransactionRequest<?> txReq = (TransactionRequest<?>)request;
 641             final ClientIdentifier clientId = txReq.getTarget().getHistoryId().getClientId();
 642             return getFrontend(clientId).handleTransactionRequest(txReq, envelope, now);
 643         } else if (request instanceof LocalHistoryRequest) {
 644             final LocalHistoryRequest<?> lhReq = (LocalHistoryRequest<?>)request;
 645             final ClientIdentifier clientId = lhReq.getTarget().getClientId();
 646             return getFrontend(clientId).handleLocalHistoryRequest(lhReq, envelope, now);
 647         } else {
 648             LOG.warn("{}: rejecting unsupported request {}", persistenceId(), request);
 649             throw new UnsupportedRequestException(request);
 650         }
 651     }
 652
 653     private void handleGetKnownClients() {
 654         final ImmutableSet<ClientIdentifier> clients;
 655         if (isLeader()) {
 656             clients = knownFrontends.values().stream()
 657                     .map(LeaderFrontendState::getIdentifier)
 658                     .collect(ImmutableSet.toImmutableSet());
 659         } else {
 660             clients = frontendMetadata.getClients();
 661         }
 662         sender().tell(new GetKnownClientsReply(clients), self());
 663     }
 664
 665     private boolean hasLeader() {
 666         return getLeaderId() != null;
 667     }
 668
 669     final int getPendingTxCommitQueueSize() {
 670         return store.getQueueSize();
 671     }
 672
 673     final int getCohortCacheSize() {
 674         return commitCoordinator.getCohortCacheSize();
 675     }
 676
 677     @Override
 678     protected final Optional<ActorRef> getRoleChangeNotifier() {
 679         return roleChangeNotifier;
 680     }
 681
 682     final String getShardName() {
 683         return shardName;
 684     }
 685
 686     @Override
 687     protected final LeaderStateChanged newLeaderStateChanged(final String memberId, final String leaderId,
 688             final short leaderPayloadVersion) {
 689         return isLeader() ? new ShardLeaderStateChanged(memberId, leaderId, store.getDataTree(), leaderPayloadVersion)
 690                 : new ShardLeaderStateChanged(memberId, leaderId, leaderPayloadVersion);
 691     }
 692
 693     private void onDatastoreContext(final DatastoreContext context) {
 694         datastoreContext = verifyNotNull(context);
 695
 696         setTransactionCommitTimeout();
 697
 698         setPersistence(datastoreContext.isPersistent());
 699
 700         updateConfigParams(datastoreContext.getShardRaftConfig());
 701     }
 702
 703     // applyState() will be invoked once consensus is reached on the payload
 704     // non-final for mocking
 705     void persistPayload(final Identifier id, final Payload payload, final boolean batchHint) {
 706         final boolean canSkipPayload = !hasFollowers() && !persistence().isRecoveryApplicable();
 707         if (canSkipPayload) {
 708             applyState(self(), id, payload);
 709         } else {
 710             // We are faking the sender
 711             persistData(self(), id, payload, batchHint);
 712         }
 713     }
 714
 715     private void handleCommitTransaction(final CommitTransaction commit) {
 716         final TransactionIdentifier txId = commit.getTransactionId();
 717         if (isLeader()) {
 718             askProtocolEncountered(txId);
 719             commitCoordinator.handleCommit(txId, getSender(), this);
 720         } else {
 721             ActorSelection leader = getLeader();
 722             if (leader == null) {
 723                 messageRetrySupport.addMessageToRetry(commit, getSender(), "Could not commit transaction " + txId);
 724             } else {
 725                 LOG.debug("{}: Forwarding CommitTransaction to leader {}", persistenceId(), leader);
 726                 leader.forward(commit, getContext());
 727             }
 728         }
 729     }
 730
 731     private void handleCanCommitTransaction(final CanCommitTransaction canCommit) {
 732         final TransactionIdentifier txId = canCommit.getTransactionId();
 733         LOG.debug("{}: Can committing transaction {}", persistenceId(), txId);
 734
 735         if (isLeader()) {
 736             askProtocolEncountered(txId);
 737             commitCoordinator.handleCanCommit(txId, getSender(), this);
 738         } else {
 739             ActorSelection leader = getLeader();
 740             if (leader == null) {
 741                 messageRetrySupport.addMessageToRetry(canCommit, getSender(),
 742                         "Could not canCommit transaction " + txId);
 743             } else {
 744                 LOG.debug("{}: Forwarding CanCommitTransaction to leader {}", persistenceId(), leader);
 745                 leader.forward(canCommit, getContext());
 746             }
 747         }
 748     }
 749
 750     @SuppressWarnings("checkstyle:IllegalCatch")
 751     private void handleBatchedModificationsLocal(final BatchedModifications batched, final ActorRef sender) {
 752         askProtocolEncountered(batched.getTransactionId());
 753
 754         try {
 755             commitCoordinator.handleBatchedModifications(batched, sender, this);
 756         } catch (Exception e) {
 757             LOG.error("{}: Error handling BatchedModifications for Tx {}", persistenceId(),
 758                     batched.getTransactionId(), e);
 759             sender.tell(new Failure(e), getSelf());
 760         }
 761     }
 762
 763     private void handleBatchedModifications(final BatchedModifications batched) {
 764         // This message is sent to prepare the modifications transaction directly on the Shard as an
 765         // optimization to avoid the extra overhead of a separate ShardTransaction actor. On the last
 766         // BatchedModifications message, the caller sets the ready flag in the message indicating
 767         // modifications are complete. The reply contains the cohort actor path (this actor) for the caller
 768         // to initiate the 3-phase commit. This also avoids the overhead of sending an additional
 769         // ReadyTransaction message.
 770
 771         // If we're not the leader then forward to the leader. This is a safety measure - we shouldn't
 772         // normally get here if we're not the leader as the front-end (TransactionProxy) should determine
 773         // the primary/leader shard. However with timing and caching on the front-end, there's a small
 774         // window where it could have a stale leader during leadership transitions.
 775         //
 776         boolean isLeaderActive = isLeaderActive();
 777         if (isLeader() && isLeaderActive) {
 778             handleBatchedModificationsLocal(batched, getSender());
 779         } else {
 780             ActorSelection leader = getLeader();
 781             if (!isLeaderActive || leader == null) {
 782                 messageRetrySupport.addMessageToRetry(batched, getSender(),
 783                         "Could not process BatchedModifications " + batched.getTransactionId());
 784             } else {
 785                 // If this is not the first batch and leadership changed in between batched messages,
 786                 // we need to reconstruct previous BatchedModifications from the transaction
 787                 // DataTreeModification, honoring the max batched modification count, and forward all the
 788                 // previous BatchedModifications to the new leader.
 789                 Collection<BatchedModifications> newModifications = commitCoordinator
 790                         .createForwardedBatchedModifications(batched,
 791                                 datastoreContext.getShardBatchedModificationCount());
 792
 793                 LOG.debug("{}: Forwarding {} BatchedModifications to leader {}", persistenceId(),
 794                         newModifications.size(), leader);
 795
 796                 for (BatchedModifications bm : newModifications) {
 797                     leader.forward(bm, getContext());
 798                 }
 799             }
 800         }
 801     }
 802
 803     private boolean failIfIsolatedLeader(final ActorRef sender) {
 804         if (isIsolatedLeader()) {
 805             sender.tell(new Failure(new NoShardLeaderException(String.format(
 806                     "Shard %s was the leader but has lost contact with all of its followers. Either all"
 807                     + " other follower nodes are down or this node is isolated by a network partition.",
 808                     persistenceId()))), getSelf());
 809             return true;
 810         }
 811
 812         return false;
 813     }
 814
 815     protected boolean isIsolatedLeader() {
 816         return getRaftState() == RaftState.IsolatedLeader;
 817     }
 818
 819     @SuppressWarnings("checkstyle:IllegalCatch")
 820     private void handleReadyLocalTransaction(final ReadyLocalTransaction message) {
 821         final TransactionIdentifier txId = message.getTransactionId();
 822         LOG.debug("{}: handleReadyLocalTransaction for {}", persistenceId(), txId);
 823
 824         boolean isLeaderActive = isLeaderActive();
 825         if (isLeader() && isLeaderActive) {
 826             askProtocolEncountered(txId);
 827             try {
 828                 commitCoordinator.handleReadyLocalTransaction(message, getSender(), this);
 829             } catch (Exception e) {
 830                 LOG.error("{}: Error handling ReadyLocalTransaction for Tx {}", persistenceId(), txId, e);
 831                 getSender().tell(new Failure(e), getSelf());
 832             }
 833         } else {
 834             ActorSelection leader = getLeader();
 835             if (!isLeaderActive || leader == null) {
 836                 messageRetrySupport.addMessageToRetry(message, getSender(),
 837                         "Could not process ready local transaction " + txId);
 838             } else {
 839                 LOG.debug("{}: Forwarding ReadyLocalTransaction to leader {}", persistenceId(), leader);
 840                 message.setRemoteVersion(getCurrentBehavior().getLeaderPayloadVersion());
 841                 leader.forward(message, getContext());
 842             }
 843         }
 844     }
 845
 846     private void handleForwardedReadyTransaction(final ForwardedReadyTransaction forwardedReady) {
 847         LOG.debug("{}: handleForwardedReadyTransaction for {}", persistenceId(), forwardedReady.getTransactionId());
 848
 849         boolean isLeaderActive = isLeaderActive();
 850         if (isLeader() && isLeaderActive) {
 851             askProtocolEncountered(forwardedReady.getTransactionId());
 852             commitCoordinator.handleForwardedReadyTransaction(forwardedReady, getSender(), this);
 853         } else {
 854             ActorSelection leader = getLeader();
 855             if (!isLeaderActive || leader == null) {
 856                 messageRetrySupport.addMessageToRetry(forwardedReady, getSender(),
 857                         "Could not process forwarded ready transaction " + forwardedReady.getTransactionId());
 858             } else {
 859                 LOG.debug("{}: Forwarding ForwardedReadyTransaction to leader {}", persistenceId(), leader);
 860
 861                 ReadyLocalTransaction readyLocal = new ReadyLocalTransaction(forwardedReady.getTransactionId(),
 862                         forwardedReady.getTransaction().getSnapshot(), forwardedReady.isDoImmediateCommit(),
 863                         forwardedReady.getParticipatingShardNames());
 864                 readyLocal.setRemoteVersion(getCurrentBehavior().getLeaderPayloadVersion());
 865                 leader.forward(readyLocal, getContext());
 866             }
 867         }
 868     }
 869
 870     private void handleAbortTransaction(final AbortTransaction abort) {
 871         final TransactionIdentifier transactionId = abort.getTransactionId();
 872         askProtocolEncountered(transactionId);
 873         doAbortTransaction(transactionId, getSender());
 874     }
 875
 876     final void doAbortTransaction(final Identifier transactionID, final ActorRef sender) {
 877         commitCoordinator.handleAbort(transactionID, sender, this);
 878     }
 879
 880     private void handleCreateTransaction(final Object message) {
 881         if (isLeader()) {
 882             createTransaction(CreateTransaction.fromSerializable(message));
 883         } else if (getLeader() != null) {
 884             getLeader().forward(message, getContext());
 885         } else {
 886             getSender().tell(new Failure(new NoShardLeaderException(
 887                     "Could not create a shard transaction", persistenceId())), getSelf());
 888         }
 889     }
 890
 891     private void closeTransactionChain(final CloseTransactionChain closeTransactionChain) {
 892         if (isLeader()) {
 893             final LocalHistoryIdentifier id = closeTransactionChain.getIdentifier();
 894             askProtocolEncountered(id.getClientId());
 895             store.closeTransactionChain(id);
 896         } else if (getLeader() != null) {
 897             getLeader().forward(closeTransactionChain, getContext());
 898         } else {
 899             LOG.warn("{}: Could not close transaction {}", persistenceId(), closeTransactionChain.getIdentifier());
 900         }
 901     }
 902
 903     @SuppressWarnings("checkstyle:IllegalCatch")
 904     private void createTransaction(final CreateTransaction createTransaction) {
 905         askProtocolEncountered(createTransaction.getTransactionId());
 906
 907         try {
 908             if (TransactionType.fromInt(createTransaction.getTransactionType()) != TransactionType.READ_ONLY
 909                     && failIfIsolatedLeader(getSender())) {
 910                 return;
 911             }
 912
 913             ActorRef transactionActor = createTransaction(createTransaction.getTransactionType(),
 914                 createTransaction.getTransactionId());
 915
 916             getSender().tell(new CreateTransactionReply(Serialization.serializedActorPath(transactionActor),
 917                     createTransaction.getTransactionId(), createTransaction.getVersion()).toSerializable(), getSelf());
 918         } catch (Exception e) {
 919             getSender().tell(new Failure(e), getSelf());
 920         }
 921     }
 922
 923     private ActorRef createTransaction(final int transactionType, final TransactionIdentifier transactionId) {
 924         LOG.debug("{}: Creating transaction : {} ", persistenceId(), transactionId);
 925         return transactionActorFactory.newShardTransaction(TransactionType.fromInt(transactionType),
 926             transactionId);
 927     }
 928
 929     // Called on leader only
 930     private void askProtocolEncountered(final TransactionIdentifier transactionId) {
 931         askProtocolEncountered(transactionId.getHistoryId().getClientId());
 932     }
 933
 934     // Called on leader only
 935     private void askProtocolEncountered(final ClientIdentifier clientId) {
 936         final FrontendIdentifier frontend = clientId.getFrontendId();
 937         final LeaderFrontendState state = knownFrontends.get(frontend);
 938         if (!(state instanceof LeaderFrontendState.Disabled)) {
 939             LOG.debug("{}: encountered ask-based client {}, disabling transaction tracking", persistenceId(), clientId);
 940             if (knownFrontends.isEmpty()) {
 941                 knownFrontends = new HashMap<>();
 942             }
 943             knownFrontends.put(frontend, new LeaderFrontendState.Disabled(persistenceId(), clientId, getDataStore()));
 944
 945             persistPayload(clientId, DisableTrackingPayload.create(clientId,
 946                 datastoreContext.getInitialPayloadSerializedBufferCapacity()), false);
 947         }
 948     }
 949
 950     private void updateSchemaContext(final UpdateSchemaContext message) {
 951         updateSchemaContext(message.getEffectiveModelContext());
 952     }
 953
 954     @VisibleForTesting
 955     void updateSchemaContext(final @NonNull EffectiveModelContext schemaContext) {
 956         store.updateSchemaContext(schemaContext);
 957     }
 958
 959     private boolean isMetricsCaptureEnabled() {
 960         CommonConfig config = new CommonConfig(getContext().system().settings().config());
 961         return config.isMetricCaptureEnabled();
 962     }
 963
 964     @Override
 965     protected final RaftActorSnapshotCohort getRaftActorSnapshotCohort() {
 966         return snapshotCohort;
 967     }
 968
 969     @Override
 970     protected final RaftActorRecoveryCohort getRaftActorRecoveryCohort() {
 971         if (restoreFromSnapshot == null) {
 972             return ShardRecoveryCoordinator.create(store, persistenceId(), LOG);
 973         }
 974
 975         return ShardRecoveryCoordinator.forSnapshot(store, persistenceId(), LOG, restoreFromSnapshot.getSnapshot());
 976     }
 977
 978     @Override
 979     // non-final for testing
 980     protected void onRecoveryComplete() {
 981         restoreFromSnapshot = null;
 982
 983         //notify shard manager
 984         getContext().parent().tell(new ActorInitialized(), getSelf());
 985
 986         // Being paranoid here - this method should only be called once but just in case...
 987         if (txCommitTimeoutCheckSchedule == null) {
 988             // Schedule a message to be periodically sent to check if the current in-progress
 989             // transaction should be expired and aborted.
 990             FiniteDuration period = FiniteDuration.create(transactionCommitTimeout / 3, TimeUnit.MILLISECONDS);
 991             txCommitTimeoutCheckSchedule = getContext().system().scheduler().schedule(
 992                     period, period, getSelf(),
 993                     TX_COMMIT_TIMEOUT_CHECK_MESSAGE, getContext().dispatcher(), ActorRef.noSender());
 994         }
 995     }
 996
 997     @Override
 998     protected final void applyState(final ActorRef clientActor, final Identifier identifier, final Object data) {
 999         if (data instanceof Payload) {
1000             if (data instanceof DisableTrackingPayload) {
1001                 disableTracking((DisableTrackingPayload) data);
1002                 return;
1003             }
1004
1005             try {
1006                 store.applyReplicatedPayload(identifier, (Payload)data);
1007             } catch (DataValidationFailedException | IOException e) {
1008                 LOG.error("{}: Error applying replica {}", persistenceId(), identifier, e);
1009             }
1010         } else {
1011             LOG.error("{}: Unknown state for {} received {}", persistenceId(), identifier, data);
1012         }
1013     }
1014
1015     @Override
1016     protected final void onStateChanged() {
1017         boolean isLeader = isLeader();
1018         boolean hasLeader = hasLeader();
1019         treeChangeSupport.onLeadershipChange(isLeader, hasLeader);
1020
1021         // If this actor is no longer the leader close all the transaction chains
1022         if (!isLeader) {
1023             if (LOG.isDebugEnabled()) {
1024                 LOG.debug(
1025                     "{}: onStateChanged: Closing all transaction chains because shard {} is no longer the leader",
1026                     persistenceId(), getId());
1027             }
1028
1029             paused = false;
1030             store.purgeLeaderState();
1031         }
1032
1033         if (hasLeader && !isIsolatedLeader()) {
1034             messageRetrySupport.retryMessages();
1035         }
1036     }
1037
1038     @Override
1039     protected final void onLeaderChanged(final String oldLeader, final String newLeader) {
1040         shardMBean.incrementLeadershipChangeCount();
1041         paused = false;
1042
1043         if (!isLeader()) {
1044             if (!knownFrontends.isEmpty()) {
1045                 LOG.debug("{}: removing frontend state for {}", persistenceId(), knownFrontends.keySet());
1046                 knownFrontends = ImmutableMap.of();
1047             }
1048
1049             requestMessageAssembler.close();
1050
1051             if (!hasLeader()) {
1052                 // No leader anywhere, nothing else to do
1053                 return;
1054             }
1055
1056             // Another leader was elected. If we were the previous leader and had pending transactions, convert
1057             // them to transaction messages and send to the new leader.
1058             ActorSelection leader = getLeader();
1059             if (leader != null) {
1060                 // Clears all pending transactions and converts them to messages to be forwarded to a new leader.
1061                 Collection<?> messagesToForward = commitCoordinator.convertPendingTransactionsToMessages(
1062                     datastoreContext.getShardBatchedModificationCount());
1063
1064                 if (!messagesToForward.isEmpty()) {
1065                     LOG.debug("{}: Forwarding {} pending transaction messages to leader {}", persistenceId(),
1066                             messagesToForward.size(), leader);
1067
1068                     for (Object message : messagesToForward) {
1069                         LOG.debug("{}: Forwarding pending transaction message {}", persistenceId(), message);
1070
1071                         leader.tell(message, self());
1072                     }
1073                 }
1074             } else {
1075                 commitCoordinator.abortPendingTransactions("The transacton was aborted due to inflight leadership "
1076                         + "change and the leader address isn't available.", this);
1077             }
1078         } else {
1079             // We have become the leader, we need to reconstruct frontend state
1080             knownFrontends = verifyNotNull(frontendMetadata.toLeaderState(this));
1081             LOG.debug("{}: became leader with frontend state for {}", persistenceId(), knownFrontends.keySet());
1082         }
1083
1084         if (!isIsolatedLeader()) {
1085             messageRetrySupport.retryMessages();
1086         }
1087     }
1088
1089     @Override
1090     protected final void pauseLeader(final Runnable operation) {
1091         LOG.debug("{}: In pauseLeader, operation: {}", persistenceId(), operation);
1092         paused = true;
1093
1094         // Tell-based protocol can replay transaction state, so it is safe to blow it up when we are paused.
1095         if (datastoreContext.isUseTellBasedProtocol()) {
1096             knownFrontends.values().forEach(LeaderFrontendState::retire);
1097             knownFrontends = ImmutableMap.of();
1098         }
1099
1100         store.setRunOnPendingTransactionsComplete(operation);
1101     }
1102
1103     @Override
1104     protected final void unpauseLeader() {
1105         LOG.debug("{}: In unpauseLeader", persistenceId());
1106         paused = false;
1107
1108         store.setRunOnPendingTransactionsComplete(null);
1109
1110         // Restore tell-based protocol state as if we were becoming the leader
1111         knownFrontends = verifyNotNull(frontendMetadata.toLeaderState(this));
1112     }
1113
1114     @Override
1115     protected final OnDemandRaftState.AbstractBuilder<?, ?> newOnDemandRaftStateBuilder() {
1116         return OnDemandShardState.newBuilder()
1117             .treeChangeListenerActors(treeChangeSupport.getListenerActors())
1118             .commitCohortActors(store.getCohortActors());
1119     }
1120
1121     @Override
1122     public final String persistenceId() {
1123         return name;
1124     }
1125
1126     @Override
1127     public final String journalPluginId() {
1128         // This method may be invoked from super constructor (wonderful), hence we also need to handle the case of
1129         // the field being uninitialized because our constructor is not finished.
1130         if (datastoreContext != null && !datastoreContext.isPersistent()) {
1131             return NON_PERSISTENT_JOURNAL_ID;
1132         }
1133         return super.journalPluginId();
1134     }
1135
1136     @VisibleForTesting
1137     final ShardCommitCoordinator getCommitCoordinator() {
1138         return commitCoordinator;
1139     }
1140
1141     // non-final for mocking
1142     DatastoreContext getDatastoreContext() {
1143         return datastoreContext;
1144     }
1145
1146     @VisibleForTesting
1147     final ShardDataTree getDataStore() {
1148         return store;
1149     }
1150
1151     @VisibleForTesting
1152     // non-final for mocking
1153     ShardStats getShardMBean() {
1154         return shardMBean;
1155     }
1156
1157     public static Builder builder() {
1158         return new Builder();
1159     }
1160
1161     public abstract static class AbstractBuilder<T extends AbstractBuilder<T, S>, S extends Shard> {
1162         private final Class<? extends S> shardClass;
1163         private ShardIdentifier id;
1164         private Map<String, String> peerAddresses = Collections.emptyMap();
1165         private DatastoreContext datastoreContext;
1166         private EffectiveModelContextProvider schemaContextProvider;
1167         private DatastoreSnapshot.ShardSnapshot restoreFromSnapshot;
1168         private DataTree dataTree;
1169
1170         private volatile boolean sealed;
1171
1172         AbstractBuilder(final Class<? extends S> shardClass) {
1173             this.shardClass = shardClass;
1174         }
1175
1176         final void checkSealed() {
1177             checkState(!sealed, "Builder is already sealed - further modifications are not allowed");
1178         }
1179
1180         @SuppressWarnings("unchecked")
1181         private T self() {
1182             return (T) this;
1183         }
1184
1185         public T id(final ShardIdentifier newId) {
1186             checkSealed();
1187             this.id = newId;
1188             return self();
1189         }
1190
1191         public T peerAddresses(final Map<String, String> newPeerAddresses) {
1192             checkSealed();
1193             this.peerAddresses = newPeerAddresses;
1194             return self();
1195         }
1196
1197         public T datastoreContext(final DatastoreContext newDatastoreContext) {
1198             checkSealed();
1199             this.datastoreContext = newDatastoreContext;
1200             return self();
1201         }
1202
1203         public T schemaContextProvider(final EffectiveModelContextProvider newSchemaContextProvider) {
1204             checkSealed();
1205             this.schemaContextProvider = requireNonNull(newSchemaContextProvider);
1206             return self();
1207         }
1208
1209         public T restoreFromSnapshot(final DatastoreSnapshot.ShardSnapshot newRestoreFromSnapshot) {
1210             checkSealed();
1211             this.restoreFromSnapshot = newRestoreFromSnapshot;
1212             return self();
1213         }
1214
1215         public T dataTree(final DataTree newDataTree) {
1216             checkSealed();
1217             this.dataTree = newDataTree;
1218             return self();
1219         }
1220
1221         public ShardIdentifier getId() {
1222             return id;
1223         }
1224
1225         public Map<String, String> getPeerAddresses() {
1226             return peerAddresses;
1227         }
1228
1229         public DatastoreContext getDatastoreContext() {
1230             return datastoreContext;
1231         }
1232
1233         public EffectiveModelContext getSchemaContext() {
1234             return verifyNotNull(schemaContextProvider.getEffectiveModelContext());
1235         }
1236
1237         public DatastoreSnapshot.ShardSnapshot getRestoreFromSnapshot() {
1238             return restoreFromSnapshot;
1239         }
1240
1241         public DataTree getDataTree() {
1242             return dataTree;
1243         }
1244
1245         public TreeType getTreeType() {
1246             switch (datastoreContext.getLogicalStoreType()) {
1247                 case CONFIGURATION:
1248                     return TreeType.CONFIGURATION;
1249                 case OPERATIONAL:
1250                     return TreeType.OPERATIONAL;
1251                 default:
1252                     throw new IllegalStateException("Unhandled logical store type "
1253                             + datastoreContext.getLogicalStoreType());
1254             }
1255         }
1256
1257         protected void verify() {
1258             requireNonNull(id, "id should not be null");
1259             requireNonNull(peerAddresses, "peerAddresses should not be null");
1260             requireNonNull(datastoreContext, "dataStoreContext should not be null");
1261             requireNonNull(schemaContextProvider, "schemaContextProvider should not be null");
1262         }
1263
1264         public Props props() {
1265             sealed = true;
1266             verify();
1267             return Props.create(shardClass, this);
1268         }
1269     }
1270
1271     public static class Builder extends AbstractBuilder<Builder, Shard> {
1272         Builder() {
1273             this(Shard.class);
1274         }
1275
1276         Builder(final Class<? extends Shard> shardClass) {
1277             super(shardClass);
1278         }
1279     }
1280
1281     Ticker ticker() {
1282         return Ticker.systemTicker();
1283     }
1284
1285     void scheduleNextPendingTransaction() {
1286         self().tell(RESUME_NEXT_PENDING_TRANSACTION, ActorRef.noSender());
1287     }
1288 }