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         this.name = builder.getId().toString();
 225         this.shardName = builder.getId().getShardName();
 226         this.datastoreContext = builder.getDatastoreContext();
 227         this.restoreFromSnapshot = builder.getRestoreFromSnapshot();
 228         this.frontendMetadata = new FrontendMetadata(name);
 229         this.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, this.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             this.name, datastoreContext);
 281
 282         messageRetrySupport = new ShardTransactionMessageRetrySupport(this);
 283
 284         responseMessageSlicer = MessageSlicer.builder().logContext(this.name)
 285                 .messageSliceSize(datastoreContext.getMaximumMessageSliceSize())
 286                 .fileBackedStreamFactory(getRaftActorContext().getFileBackedOutputStreamFactory())
 287                 .expireStateAfterInactivity(2, TimeUnit.MINUTES).build();
 288
 289         requestMessageAssembler = MessageAssembler.builder().logContext(this.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         LOG.debug("{}: handleReadyLocalTransaction for {}", persistenceId(), message.getTransactionId());
 822
 823         boolean isLeaderActive = isLeaderActive();
 824         if (isLeader() && isLeaderActive) {
 825             try {
 826                 commitCoordinator.handleReadyLocalTransaction(message, getSender(), this);
 827             } catch (Exception e) {
 828                 LOG.error("{}: Error handling ReadyLocalTransaction for Tx {}", persistenceId(),
 829                         message.getTransactionId(), e);
 830                 getSender().tell(new Failure(e), getSelf());
 831             }
 832         } else {
 833             ActorSelection leader = getLeader();
 834             if (!isLeaderActive || leader == null) {
 835                 messageRetrySupport.addMessageToRetry(message, getSender(),
 836                         "Could not process ready local transaction " + message.getTransactionId());
 837             } else {
 838                 LOG.debug("{}: Forwarding ReadyLocalTransaction to leader {}", persistenceId(), leader);
 839                 message.setRemoteVersion(getCurrentBehavior().getLeaderPayloadVersion());
 840                 leader.forward(message, getContext());
 841             }
 842         }
 843     }
 844
 845     private void handleForwardedReadyTransaction(final ForwardedReadyTransaction forwardedReady) {
 846         LOG.debug("{}: handleForwardedReadyTransaction for {}", persistenceId(), forwardedReady.getTransactionId());
 847
 848         boolean isLeaderActive = isLeaderActive();
 849         if (isLeader() && isLeaderActive) {
 850             askProtocolEncountered(forwardedReady.getTransactionId());
 851             commitCoordinator.handleForwardedReadyTransaction(forwardedReady, getSender(), this);
 852         } else {
 853             ActorSelection leader = getLeader();
 854             if (!isLeaderActive || leader == null) {
 855                 messageRetrySupport.addMessageToRetry(forwardedReady, getSender(),
 856                         "Could not process forwarded ready transaction " + forwardedReady.getTransactionId());
 857             } else {
 858                 LOG.debug("{}: Forwarding ForwardedReadyTransaction to leader {}", persistenceId(), leader);
 859
 860                 ReadyLocalTransaction readyLocal = new ReadyLocalTransaction(forwardedReady.getTransactionId(),
 861                         forwardedReady.getTransaction().getSnapshot(), forwardedReady.isDoImmediateCommit(),
 862                         forwardedReady.getParticipatingShardNames());
 863                 readyLocal.setRemoteVersion(getCurrentBehavior().getLeaderPayloadVersion());
 864                 leader.forward(readyLocal, getContext());
 865             }
 866         }
 867     }
 868
 869     private void handleAbortTransaction(final AbortTransaction abort) {
 870         final TransactionIdentifier transactionId = abort.getTransactionId();
 871         askProtocolEncountered(transactionId);
 872         doAbortTransaction(transactionId, getSender());
 873     }
 874
 875     final void doAbortTransaction(final Identifier transactionID, final ActorRef sender) {
 876         commitCoordinator.handleAbort(transactionID, sender, this);
 877     }
 878
 879     private void handleCreateTransaction(final Object message) {
 880         if (isLeader()) {
 881             createTransaction(CreateTransaction.fromSerializable(message));
 882         } else if (getLeader() != null) {
 883             getLeader().forward(message, getContext());
 884         } else {
 885             getSender().tell(new Failure(new NoShardLeaderException(
 886                     "Could not create a shard transaction", persistenceId())), getSelf());
 887         }
 888     }
 889
 890     private void closeTransactionChain(final CloseTransactionChain closeTransactionChain) {
 891         if (isLeader()) {
 892             final LocalHistoryIdentifier id = closeTransactionChain.getIdentifier();
 893             askProtocolEncountered(id.getClientId());
 894             store.closeTransactionChain(id);
 895         } else if (getLeader() != null) {
 896             getLeader().forward(closeTransactionChain, getContext());
 897         } else {
 898             LOG.warn("{}: Could not close transaction {}", persistenceId(), closeTransactionChain.getIdentifier());
 899         }
 900     }
 901
 902     @SuppressWarnings("checkstyle:IllegalCatch")
 903     private void createTransaction(final CreateTransaction createTransaction) {
 904         askProtocolEncountered(createTransaction.getTransactionId());
 905
 906         try {
 907             if (TransactionType.fromInt(createTransaction.getTransactionType()) != TransactionType.READ_ONLY
 908                     && failIfIsolatedLeader(getSender())) {
 909                 return;
 910             }
 911
 912             ActorRef transactionActor = createTransaction(createTransaction.getTransactionType(),
 913                 createTransaction.getTransactionId());
 914
 915             getSender().tell(new CreateTransactionReply(Serialization.serializedActorPath(transactionActor),
 916                     createTransaction.getTransactionId(), createTransaction.getVersion()).toSerializable(), getSelf());
 917         } catch (Exception e) {
 918             getSender().tell(new Failure(e), getSelf());
 919         }
 920     }
 921
 922     private ActorRef createTransaction(final int transactionType, final TransactionIdentifier transactionId) {
 923         LOG.debug("{}: Creating transaction : {} ", persistenceId(), transactionId);
 924         return transactionActorFactory.newShardTransaction(TransactionType.fromInt(transactionType),
 925             transactionId);
 926     }
 927
 928     // Called on leader only
 929     private void askProtocolEncountered(final TransactionIdentifier transactionId) {
 930         askProtocolEncountered(transactionId.getHistoryId().getClientId());
 931     }
 932
 933     // Called on leader only
 934     private void askProtocolEncountered(final ClientIdentifier clientId) {
 935         final FrontendIdentifier frontend = clientId.getFrontendId();
 936         final LeaderFrontendState state = knownFrontends.get(frontend);
 937         if (!(state instanceof LeaderFrontendState.Disabled)) {
 938             LOG.debug("{}: encountered ask-based client {}, disabling transaction tracking", persistenceId(), clientId);
 939             if (knownFrontends.isEmpty()) {
 940                 knownFrontends = new HashMap<>();
 941             }
 942             knownFrontends.put(frontend, new LeaderFrontendState.Disabled(persistenceId(), clientId, getDataStore()));
 943
 944             persistPayload(clientId, DisableTrackingPayload.create(clientId,
 945                 datastoreContext.getInitialPayloadSerializedBufferCapacity()), false);
 946         }
 947     }
 948
 949     private void updateSchemaContext(final UpdateSchemaContext message) {
 950         updateSchemaContext(message.getEffectiveModelContext());
 951     }
 952
 953     @VisibleForTesting
 954     void updateSchemaContext(final @NonNull EffectiveModelContext schemaContext) {
 955         store.updateSchemaContext(schemaContext);
 956     }
 957
 958     private boolean isMetricsCaptureEnabled() {
 959         CommonConfig config = new CommonConfig(getContext().system().settings().config());
 960         return config.isMetricCaptureEnabled();
 961     }
 962
 963     @Override
 964     protected final RaftActorSnapshotCohort getRaftActorSnapshotCohort() {
 965         return snapshotCohort;
 966     }
 967
 968     @Override
 969     protected final RaftActorRecoveryCohort getRaftActorRecoveryCohort() {
 970         if (restoreFromSnapshot == null) {
 971             return ShardRecoveryCoordinator.create(store, persistenceId(), LOG);
 972         }
 973
 974         return ShardRecoveryCoordinator.forSnapshot(store, persistenceId(), LOG, restoreFromSnapshot.getSnapshot());
 975     }
 976
 977     @Override
 978     // non-final for testing
 979     protected void onRecoveryComplete() {
 980         restoreFromSnapshot = null;
 981
 982         //notify shard manager
 983         getContext().parent().tell(new ActorInitialized(), getSelf());
 984
 985         // Being paranoid here - this method should only be called once but just in case...
 986         if (txCommitTimeoutCheckSchedule == null) {
 987             // Schedule a message to be periodically sent to check if the current in-progress
 988             // transaction should be expired and aborted.
 989             FiniteDuration period = FiniteDuration.create(transactionCommitTimeout / 3, TimeUnit.MILLISECONDS);
 990             txCommitTimeoutCheckSchedule = getContext().system().scheduler().schedule(
 991                     period, period, getSelf(),
 992                     TX_COMMIT_TIMEOUT_CHECK_MESSAGE, getContext().dispatcher(), ActorRef.noSender());
 993         }
 994     }
 995
 996     @Override
 997     protected final void applyState(final ActorRef clientActor, final Identifier identifier, final Object data) {
 998         if (data instanceof Payload) {
 999             if (data instanceof DisableTrackingPayload) {
1000                 disableTracking((DisableTrackingPayload) data);
1001                 return;
1002             }
1003
1004             try {
1005                 store.applyReplicatedPayload(identifier, (Payload)data);
1006             } catch (DataValidationFailedException | IOException e) {
1007                 LOG.error("{}: Error applying replica {}", persistenceId(), identifier, e);
1008             }
1009         } else {
1010             LOG.error("{}: Unknown state for {} received {}", persistenceId(), identifier, data);
1011         }
1012     }
1013
1014     @Override
1015     protected final void onStateChanged() {
1016         boolean isLeader = isLeader();
1017         boolean hasLeader = hasLeader();
1018         treeChangeSupport.onLeadershipChange(isLeader, hasLeader);
1019
1020         // If this actor is no longer the leader close all the transaction chains
1021         if (!isLeader) {
1022             if (LOG.isDebugEnabled()) {
1023                 LOG.debug(
1024                     "{}: onStateChanged: Closing all transaction chains because shard {} is no longer the leader",
1025                     persistenceId(), getId());
1026             }
1027
1028             paused = false;
1029             store.purgeLeaderState();
1030         }
1031
1032         if (hasLeader && !isIsolatedLeader()) {
1033             messageRetrySupport.retryMessages();
1034         }
1035     }
1036
1037     @Override
1038     protected final void onLeaderChanged(final String oldLeader, final String newLeader) {
1039         shardMBean.incrementLeadershipChangeCount();
1040         paused = false;
1041
1042         if (!isLeader()) {
1043             if (!knownFrontends.isEmpty()) {
1044                 LOG.debug("{}: removing frontend state for {}", persistenceId(), knownFrontends.keySet());
1045                 knownFrontends = ImmutableMap.of();
1046             }
1047
1048             requestMessageAssembler.close();
1049
1050             if (!hasLeader()) {
1051                 // No leader anywhere, nothing else to do
1052                 return;
1053             }
1054
1055             // Another leader was elected. If we were the previous leader and had pending transactions, convert
1056             // them to transaction messages and send to the new leader.
1057             ActorSelection leader = getLeader();
1058             if (leader != null) {
1059                 // Clears all pending transactions and converts them to messages to be forwarded to a new leader.
1060                 Collection<?> messagesToForward = commitCoordinator.convertPendingTransactionsToMessages(
1061                     datastoreContext.getShardBatchedModificationCount());
1062
1063                 if (!messagesToForward.isEmpty()) {
1064                     LOG.debug("{}: Forwarding {} pending transaction messages to leader {}", persistenceId(),
1065                             messagesToForward.size(), leader);
1066
1067                     for (Object message : messagesToForward) {
1068                         LOG.debug("{}: Forwarding pending transaction message {}", persistenceId(), message);
1069
1070                         leader.tell(message, self());
1071                     }
1072                 }
1073             } else {
1074                 commitCoordinator.abortPendingTransactions("The transacton was aborted due to inflight leadership "
1075                         + "change and the leader address isn't available.", this);
1076             }
1077         } else {
1078             // We have become the leader, we need to reconstruct frontend state
1079             knownFrontends = verifyNotNull(frontendMetadata.toLeaderState(this));
1080             LOG.debug("{}: became leader with frontend state for {}", persistenceId(), knownFrontends.keySet());
1081         }
1082
1083         if (!isIsolatedLeader()) {
1084             messageRetrySupport.retryMessages();
1085         }
1086     }
1087
1088     @Override
1089     protected final void pauseLeader(final Runnable operation) {
1090         LOG.debug("{}: In pauseLeader, operation: {}", persistenceId(), operation);
1091         paused = true;
1092
1093         // Tell-based protocol can replay transaction state, so it is safe to blow it up when we are paused.
1094         if (datastoreContext.isUseTellBasedProtocol()) {
1095             knownFrontends.values().forEach(LeaderFrontendState::retire);
1096             knownFrontends = ImmutableMap.of();
1097         }
1098
1099         store.setRunOnPendingTransactionsComplete(operation);
1100     }
1101
1102     @Override
1103     protected final void unpauseLeader() {
1104         LOG.debug("{}: In unpauseLeader", persistenceId());
1105         paused = false;
1106
1107         store.setRunOnPendingTransactionsComplete(null);
1108
1109         // Restore tell-based protocol state as if we were becoming the leader
1110         knownFrontends = verifyNotNull(frontendMetadata.toLeaderState(this));
1111     }
1112
1113     @Override
1114     protected final OnDemandRaftState.AbstractBuilder<?, ?> newOnDemandRaftStateBuilder() {
1115         return OnDemandShardState.newBuilder()
1116             .treeChangeListenerActors(treeChangeSupport.getListenerActors())
1117             .commitCohortActors(store.getCohortActors());
1118     }
1119
1120     @Override
1121     public final String persistenceId() {
1122         return this.name;
1123     }
1124
1125     @Override
1126     public final String journalPluginId() {
1127         // This method may be invoked from super constructor (wonderful), hence we also need to handle the case of
1128         // the field being uninitialized because our constructor is not finished.
1129         if (datastoreContext != null && !datastoreContext.isPersistent()) {
1130             return NON_PERSISTENT_JOURNAL_ID;
1131         }
1132         return super.journalPluginId();
1133     }
1134
1135     @VisibleForTesting
1136     final ShardCommitCoordinator getCommitCoordinator() {
1137         return commitCoordinator;
1138     }
1139
1140     // non-final for mocking
1141     DatastoreContext getDatastoreContext() {
1142         return datastoreContext;
1143     }
1144
1145     @VisibleForTesting
1146     final ShardDataTree getDataStore() {
1147         return store;
1148     }
1149
1150     @VisibleForTesting
1151     // non-final for mocking
1152     ShardStats getShardMBean() {
1153         return shardMBean;
1154     }
1155
1156     public static Builder builder() {
1157         return new Builder();
1158     }
1159
1160     public abstract static class AbstractBuilder<T extends AbstractBuilder<T, S>, S extends Shard> {
1161         private final Class<? extends S> shardClass;
1162         private ShardIdentifier id;
1163         private Map<String, String> peerAddresses = Collections.emptyMap();
1164         private DatastoreContext datastoreContext;
1165         private EffectiveModelContextProvider schemaContextProvider;
1166         private DatastoreSnapshot.ShardSnapshot restoreFromSnapshot;
1167         private DataTree dataTree;
1168
1169         private volatile boolean sealed;
1170
1171         AbstractBuilder(final Class<? extends S> shardClass) {
1172             this.shardClass = shardClass;
1173         }
1174
1175         final void checkSealed() {
1176             checkState(!sealed, "Builder is already sealed - further modifications are not allowed");
1177         }
1178
1179         @SuppressWarnings("unchecked")
1180         private T self() {
1181             return (T) this;
1182         }
1183
1184         public T id(final ShardIdentifier newId) {
1185             checkSealed();
1186             this.id = newId;
1187             return self();
1188         }
1189
1190         public T peerAddresses(final Map<String, String> newPeerAddresses) {
1191             checkSealed();
1192             this.peerAddresses = newPeerAddresses;
1193             return self();
1194         }
1195
1196         public T datastoreContext(final DatastoreContext newDatastoreContext) {
1197             checkSealed();
1198             this.datastoreContext = newDatastoreContext;
1199             return self();
1200         }
1201
1202         public T schemaContextProvider(final EffectiveModelContextProvider newSchemaContextProvider) {
1203             checkSealed();
1204             this.schemaContextProvider = requireNonNull(newSchemaContextProvider);
1205             return self();
1206         }
1207
1208         public T restoreFromSnapshot(final DatastoreSnapshot.ShardSnapshot newRestoreFromSnapshot) {
1209             checkSealed();
1210             this.restoreFromSnapshot = newRestoreFromSnapshot;
1211             return self();
1212         }
1213
1214         public T dataTree(final DataTree newDataTree) {
1215             checkSealed();
1216             this.dataTree = newDataTree;
1217             return self();
1218         }
1219
1220         public ShardIdentifier getId() {
1221             return id;
1222         }
1223
1224         public Map<String, String> getPeerAddresses() {
1225             return peerAddresses;
1226         }
1227
1228         public DatastoreContext getDatastoreContext() {
1229             return datastoreContext;
1230         }
1231
1232         public EffectiveModelContext getSchemaContext() {
1233             return verifyNotNull(schemaContextProvider.getEffectiveModelContext());
1234         }
1235
1236         public DatastoreSnapshot.ShardSnapshot getRestoreFromSnapshot() {
1237             return restoreFromSnapshot;
1238         }
1239
1240         public DataTree getDataTree() {
1241             return dataTree;
1242         }
1243
1244         public TreeType getTreeType() {
1245             switch (datastoreContext.getLogicalStoreType()) {
1246                 case CONFIGURATION:
1247                     return TreeType.CONFIGURATION;
1248                 case OPERATIONAL:
1249                     return TreeType.OPERATIONAL;
1250                 default:
1251                     throw new IllegalStateException("Unhandled logical store type "
1252                             + datastoreContext.getLogicalStoreType());
1253             }
1254         }
1255
1256         protected void verify() {
1257             requireNonNull(id, "id should not be null");
1258             requireNonNull(peerAddresses, "peerAddresses should not be null");
1259             requireNonNull(datastoreContext, "dataStoreContext should not be null");
1260             requireNonNull(schemaContextProvider, "schemaContextProvider should not be null");
1261         }
1262
1263         public Props props() {
1264             sealed = true;
1265             verify();
1266             return Props.create(shardClass, this);
1267         }
1268     }
1269
1270     public static class Builder extends AbstractBuilder<Builder, Shard> {
1271         Builder() {
1272             this(Shard.class);
1273         }
1274
1275         Builder(final Class<? extends Shard> shardClass) {
1276             super(shardClass);
1277         }
1278     }
1279
1280     Ticker ticker() {
1281         return Ticker.systemTicker();
1282     }
1283
1284     void scheduleNextPendingTransaction() {
1285         self().tell(RESUME_NEXT_PENDING_TRANSACTION, ActorRef.noSender());
1286     }
1287 }