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