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