opendaylight/md-sal/sal-distributed-datastore/src/main/java/org/opendaylight/controller/cluster/datastore/Shard.java

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