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.debug("{}: not currently leader, rejecting request {}", persistenceId(), message);
 436                 throw new NotLeaderException(getSelf());
 437             }
 438
 439             final ABIVersion selectedVersion = selectVersion(message);
 440             final LeaderFrontendState frontend = getFrontend(message.getTarget());
 441             frontend.reconnect();
 442             message.getReplyTo().tell(new ConnectClientSuccess(message.getTarget(), message.getSequence(), getSelf(),
 443                 ImmutableList.of(), store.getDataTree(), CLIENT_MAX_MESSAGES).toVersion(selectedVersion),
 444                 ActorRef.noSender());
 445         } catch (RequestException | RuntimeException e) {
 446             message.getReplyTo().tell(new Failure(e), ActorRef.noSender());
 447         }
 448     }
 449
 450     private @Nullable RequestSuccess<?, ?> handleRequest(final RequestEnvelope envelope, final long now)
 451             throws RequestException {
 452         // We are not the leader, hence we want to fail-fast.
 453         if (!isLeader() || !isLeaderActive()) {
 454             LOG.debug("{}: not currently leader, rejecting request {}", persistenceId(), envelope);
 455             throw new NotLeaderException(getSelf());
 456         }
 457
 458         final Request<?, ?> request = envelope.getMessage();
 459         if (request instanceof TransactionRequest) {
 460             final TransactionRequest<?> txReq = (TransactionRequest<?>)request;
 461             final ClientIdentifier clientId = txReq.getTarget().getHistoryId().getClientId();
 462             return getFrontend(clientId).handleTransactionRequest(txReq, envelope, now);
 463         } else if (request instanceof LocalHistoryRequest) {
 464             final LocalHistoryRequest<?> lhReq = (LocalHistoryRequest<?>)request;
 465             final ClientIdentifier clientId = lhReq.getTarget().getClientId();
 466             return getFrontend(clientId).handleLocalHistoryRequest(lhReq, envelope, now);
 467         } else {
 468             LOG.warn("{}: rejecting unsupported request {}", persistenceId(), request);
 469             throw new UnsupportedRequestException(request);
 470         }
 471     }
 472
 473     private boolean hasLeader() {
 474         return getLeaderId() != null;
 475     }
 476
 477     public int getPendingTxCommitQueueSize() {
 478         return store.getQueueSize();
 479     }
 480
 481     public int getCohortCacheSize() {
 482         return commitCoordinator.getCohortCacheSize();
 483     }
 484
 485     @Override
 486     protected Optional<ActorRef> getRoleChangeNotifier() {
 487         return roleChangeNotifier;
 488     }
 489
 490     @Override
 491     protected LeaderStateChanged newLeaderStateChanged(final String memberId, final String leaderId,
 492             final short leaderPayloadVersion) {
 493         return isLeader() ? new ShardLeaderStateChanged(memberId, leaderId, store.getDataTree(), leaderPayloadVersion)
 494                 : new ShardLeaderStateChanged(memberId, leaderId, leaderPayloadVersion);
 495     }
 496
 497     protected void onDatastoreContext(final DatastoreContext context) {
 498         datastoreContext = context;
 499
 500         setTransactionCommitTimeout();
 501
 502         setPersistence(datastoreContext.isPersistent());
 503
 504         updateConfigParams(datastoreContext.getShardRaftConfig());
 505     }
 506
 507     // applyState() will be invoked once consensus is reached on the payload
 508     void persistPayload(final Identifier id, final Payload payload, final boolean batchHint) {
 509         boolean canSkipPayload = !hasFollowers() && !persistence().isRecoveryApplicable();
 510         if (canSkipPayload) {
 511             applyState(self(), id, payload);
 512         } else {
 513             // We are faking the sender
 514             persistData(self(), id, payload, batchHint);
 515         }
 516     }
 517
 518     private void handleCommitTransaction(final CommitTransaction commit) {
 519         if (isLeader()) {
 520             commitCoordinator.handleCommit(commit.getTransactionId(), getSender(), this);
 521         } else {
 522             ActorSelection leader = getLeader();
 523             if (leader == null) {
 524                 messageRetrySupport.addMessageToRetry(commit, getSender(),
 525                         "Could not commit transaction " + commit.getTransactionId());
 526             } else {
 527                 LOG.debug("{}: Forwarding CommitTransaction to leader {}", persistenceId(), leader);
 528                 leader.forward(commit, getContext());
 529             }
 530         }
 531     }
 532
 533     private void handleCanCommitTransaction(final CanCommitTransaction canCommit) {
 534         LOG.debug("{}: Can committing transaction {}", persistenceId(), canCommit.getTransactionId());
 535
 536         if (isLeader()) {
 537             commitCoordinator.handleCanCommit(canCommit.getTransactionId(), getSender(), this);
 538         } else {
 539             ActorSelection leader = getLeader();
 540             if (leader == null) {
 541                 messageRetrySupport.addMessageToRetry(canCommit, getSender(),
 542                         "Could not canCommit transaction " + canCommit.getTransactionId());
 543             } else {
 544                 LOG.debug("{}: Forwarding CanCommitTransaction to leader {}", persistenceId(), leader);
 545                 leader.forward(canCommit, getContext());
 546             }
 547         }
 548     }
 549
 550     @SuppressWarnings("checkstyle:IllegalCatch")
 551     protected void handleBatchedModificationsLocal(final BatchedModifications batched, final ActorRef sender) {
 552         try {
 553             commitCoordinator.handleBatchedModifications(batched, sender, this);
 554         } catch (Exception e) {
 555             LOG.error("{}: Error handling BatchedModifications for Tx {}", persistenceId(),
 556                     batched.getTransactionId(), e);
 557             sender.tell(new Failure(e), getSelf());
 558         }
 559     }
 560
 561     private void handleBatchedModifications(final BatchedModifications batched) {
 562         // This message is sent to prepare the modifications transaction directly on the Shard as an
 563         // optimization to avoid the extra overhead of a separate ShardTransaction actor. On the last
 564         // BatchedModifications message, the caller sets the ready flag in the message indicating
 565         // modifications are complete. The reply contains the cohort actor path (this actor) for the caller
 566         // to initiate the 3-phase commit. This also avoids the overhead of sending an additional
 567         // ReadyTransaction message.
 568
 569         // If we're not the leader then forward to the leader. This is a safety measure - we shouldn't
 570         // normally get here if we're not the leader as the front-end (TransactionProxy) should determine
 571         // the primary/leader shard. However with timing and caching on the front-end, there's a small
 572         // window where it could have a stale leader during leadership transitions.
 573         //
 574         boolean isLeaderActive = isLeaderActive();
 575         if (isLeader() && isLeaderActive) {
 576             handleBatchedModificationsLocal(batched, getSender());
 577         } else {
 578             ActorSelection leader = getLeader();
 579             if (!isLeaderActive || leader == null) {
 580                 messageRetrySupport.addMessageToRetry(batched, getSender(),
 581                         "Could not commit transaction " + batched.getTransactionId());
 582             } else {
 583                 // If this is not the first batch and leadership changed in between batched messages,
 584                 // we need to reconstruct previous BatchedModifications from the transaction
 585                 // DataTreeModification, honoring the max batched modification count, and forward all the
 586                 // previous BatchedModifications to the new leader.
 587                 Collection<BatchedModifications> newModifications = commitCoordinator
 588                         .createForwardedBatchedModifications(batched,
 589                                 datastoreContext.getShardBatchedModificationCount());
 590
 591                 LOG.debug("{}: Forwarding {} BatchedModifications to leader {}", persistenceId(),
 592                         newModifications.size(), leader);
 593
 594                 for (BatchedModifications bm : newModifications) {
 595                     leader.forward(bm, getContext());
 596                 }
 597             }
 598         }
 599     }
 600
 601     private boolean failIfIsolatedLeader(final ActorRef sender) {
 602         if (isIsolatedLeader()) {
 603             sender.tell(new Failure(new NoShardLeaderException(String.format(
 604                     "Shard %s was the leader but has lost contact with all of its followers. Either all"
 605                     + " other follower nodes are down or this node is isolated by a network partition.",
 606                     persistenceId()))), getSelf());
 607             return true;
 608         }
 609
 610         return false;
 611     }
 612
 613     protected boolean isIsolatedLeader() {
 614         return getRaftState() == RaftState.IsolatedLeader;
 615     }
 616
 617     @SuppressWarnings("checkstyle:IllegalCatch")
 618     private void handleReadyLocalTransaction(final ReadyLocalTransaction message) {
 619         LOG.debug("{}: handleReadyLocalTransaction for {}", persistenceId(), message.getTransactionId());
 620
 621         boolean isLeaderActive = isLeaderActive();
 622         if (isLeader() && isLeaderActive) {
 623             try {
 624                 commitCoordinator.handleReadyLocalTransaction(message, getSender(), this);
 625             } catch (Exception e) {
 626                 LOG.error("{}: Error handling ReadyLocalTransaction for Tx {}", persistenceId(),
 627                         message.getTransactionId(), e);
 628                 getSender().tell(new Failure(e), getSelf());
 629             }
 630         } else {
 631             ActorSelection leader = getLeader();
 632             if (!isLeaderActive || leader == null) {
 633                 messageRetrySupport.addMessageToRetry(message, getSender(),
 634                         "Could not commit transaction " + message.getTransactionId());
 635             } else {
 636                 LOG.debug("{}: Forwarding ReadyLocalTransaction to leader {}", persistenceId(), leader);
 637                 message.setRemoteVersion(getCurrentBehavior().getLeaderPayloadVersion());
 638                 leader.forward(message, getContext());
 639             }
 640         }
 641     }
 642
 643     private void handleForwardedReadyTransaction(final ForwardedReadyTransaction forwardedReady) {
 644         LOG.debug("{}: handleForwardedReadyTransaction for {}", persistenceId(), forwardedReady.getTransactionId());
 645
 646         boolean isLeaderActive = isLeaderActive();
 647         if (isLeader() && isLeaderActive) {
 648             commitCoordinator.handleForwardedReadyTransaction(forwardedReady, getSender(), this);
 649         } else {
 650             ActorSelection leader = getLeader();
 651             if (!isLeaderActive || leader == null) {
 652                 messageRetrySupport.addMessageToRetry(forwardedReady, getSender(),
 653                         "Could not commit transaction " + forwardedReady.getTransactionId());
 654             } else {
 655                 LOG.debug("{}: Forwarding ForwardedReadyTransaction to leader {}", persistenceId(), leader);
 656
 657                 ReadyLocalTransaction readyLocal = new ReadyLocalTransaction(forwardedReady.getTransactionId(),
 658                         forwardedReady.getTransaction().getSnapshot(), forwardedReady.isDoImmediateCommit());
 659                 readyLocal.setRemoteVersion(getCurrentBehavior().getLeaderPayloadVersion());
 660                 leader.forward(readyLocal, getContext());
 661             }
 662         }
 663     }
 664
 665     private void handleAbortTransaction(final AbortTransaction abort) {
 666         doAbortTransaction(abort.getTransactionId(), getSender());
 667     }
 668
 669     void doAbortTransaction(final Identifier transactionID, final ActorRef sender) {
 670         commitCoordinator.handleAbort(transactionID, sender, this);
 671     }
 672
 673     private void handleCreateTransaction(final Object message) {
 674         if (isLeader()) {
 675             createTransaction(CreateTransaction.fromSerializable(message));
 676         } else if (getLeader() != null) {
 677             getLeader().forward(message, getContext());
 678         } else {
 679             getSender().tell(new Failure(new NoShardLeaderException(
 680                     "Could not create a shard transaction", persistenceId())), getSelf());
 681         }
 682     }
 683
 684     private void closeTransactionChain(final CloseTransactionChain closeTransactionChain) {
 685         final LocalHistoryIdentifier id = closeTransactionChain.getIdentifier();
 686         store.closeTransactionChain(id, null);
 687         store.purgeTransactionChain(id, null);
 688     }
 689
 690     @SuppressWarnings("checkstyle:IllegalCatch")
 691     private void createTransaction(final CreateTransaction createTransaction) {
 692         try {
 693             if (TransactionType.fromInt(createTransaction.getTransactionType()) != TransactionType.READ_ONLY
 694                     && failIfIsolatedLeader(getSender())) {
 695                 return;
 696             }
 697
 698             ActorRef transactionActor = createTransaction(createTransaction.getTransactionType(),
 699                 createTransaction.getTransactionId());
 700
 701             getSender().tell(new CreateTransactionReply(Serialization.serializedActorPath(transactionActor),
 702                     createTransaction.getTransactionId(), createTransaction.getVersion()).toSerializable(), getSelf());
 703         } catch (Exception e) {
 704             getSender().tell(new Failure(e), getSelf());
 705         }
 706     }
 707
 708     private ActorRef createTransaction(final int transactionType, final TransactionIdentifier transactionId) {
 709         LOG.debug("{}: Creating transaction : {} ", persistenceId(), transactionId);
 710         return transactionActorFactory.newShardTransaction(TransactionType.fromInt(transactionType),
 711             transactionId);
 712     }
 713
 714     private void updateSchemaContext(final UpdateSchemaContext message) {
 715         updateSchemaContext(message.getSchemaContext());
 716     }
 717
 718     @VisibleForTesting
 719     void updateSchemaContext(final SchemaContext schemaContext) {
 720         store.updateSchemaContext(schemaContext);
 721     }
 722
 723     private boolean isMetricsCaptureEnabled() {
 724         CommonConfig config = new CommonConfig(getContext().system().settings().config());
 725         return config.isMetricCaptureEnabled();
 726     }
 727
 728     @Override
 729     @VisibleForTesting
 730     public RaftActorSnapshotCohort getRaftActorSnapshotCohort() {
 731         return snapshotCohort;
 732     }
 733
 734     @Override
 735     @Nonnull
 736     protected RaftActorRecoveryCohort getRaftActorRecoveryCohort() {
 737         return new ShardRecoveryCoordinator(store,
 738             restoreFromSnapshot != null ? restoreFromSnapshot.getSnapshot() : null, persistenceId(), LOG);
 739     }
 740
 741     @Override
 742     protected void onRecoveryComplete() {
 743         restoreFromSnapshot = null;
 744
 745         //notify shard manager
 746         getContext().parent().tell(new ActorInitialized(), getSelf());
 747
 748         // Being paranoid here - this method should only be called once but just in case...
 749         if (txCommitTimeoutCheckSchedule == null) {
 750             // Schedule a message to be periodically sent to check if the current in-progress
 751             // transaction should be expired and aborted.
 752             FiniteDuration period = Duration.create(transactionCommitTimeout / 3, TimeUnit.MILLISECONDS);
 753             txCommitTimeoutCheckSchedule = getContext().system().scheduler().schedule(
 754                     period, period, getSelf(),
 755                     TX_COMMIT_TIMEOUT_CHECK_MESSAGE, getContext().dispatcher(), ActorRef.noSender());
 756         }
 757     }
 758
 759     @Override
 760     protected void applyState(final ActorRef clientActor, final Identifier identifier, final Object data) {
 761         if (data instanceof Payload) {
 762             try {
 763                 store.applyReplicatedPayload(identifier, (Payload)data);
 764             } catch (DataValidationFailedException | IOException e) {
 765                 LOG.error("{}: Error applying replica {}", persistenceId(), identifier, e);
 766             }
 767         } else {
 768             LOG.error("{}: Unknown state for {} received {}", persistenceId(), identifier, data);
 769         }
 770     }
 771
 772     @Override
 773     protected void onStateChanged() {
 774         boolean isLeader = isLeader();
 775         boolean hasLeader = hasLeader();
 776         changeSupport.onLeadershipChange(isLeader, hasLeader);
 777         treeChangeSupport.onLeadershipChange(isLeader, hasLeader);
 778
 779         // If this actor is no longer the leader close all the transaction chains
 780         if (!isLeader) {
 781             if (LOG.isDebugEnabled()) {
 782                 LOG.debug(
 783                     "{}: onStateChanged: Closing all transaction chains because shard {} is no longer the leader",
 784                     persistenceId(), getId());
 785             }
 786
 787             store.purgeLeaderState();
 788         }
 789
 790         if (hasLeader && !isIsolatedLeader()) {
 791             messageRetrySupport.retryMessages();
 792         }
 793     }
 794
 795     @Override
 796     protected void onLeaderChanged(final String oldLeader, final String newLeader) {
 797         shardMBean.incrementLeadershipChangeCount();
 798
 799         final boolean hasLeader = hasLeader();
 800         if (!hasLeader) {
 801             // No leader implies we are not the leader, lose frontend state if we have any. This also places
 802             // an explicit guard so the map will not get modified accidentally.
 803             if (!knownFrontends.isEmpty()) {
 804                 LOG.debug("{}: removing frontend state for {}", persistenceId(), knownFrontends.keySet());
 805                 knownFrontends = ImmutableMap.of();
 806             }
 807             return;
 808         }
 809
 810         if (!isLeader()) {
 811             // Another leader was elected. If we were the previous leader and had pending transactions, convert
 812             // them to transaction messages and send to the new leader.
 813             ActorSelection leader = getLeader();
 814             if (leader != null) {
 815                 Collection<?> messagesToForward = convertPendingTransactionsToMessages();
 816
 817                 if (!messagesToForward.isEmpty()) {
 818                     LOG.debug("{}: Forwarding {} pending transaction messages to leader {}", persistenceId(),
 819                             messagesToForward.size(), leader);
 820
 821                     for (Object message : messagesToForward) {
 822                         leader.tell(message, self());
 823                     }
 824                 }
 825             } else {
 826                 commitCoordinator.abortPendingTransactions("The transacton was aborted due to inflight leadership "
 827                         + "change and the leader address isn't available.", this);
 828             }
 829         } else {
 830             // We have become the leader, we need to reconstruct frontend state
 831             knownFrontends = Verify.verifyNotNull(frontendMetadata.toLeaderState(this));
 832             LOG.debug("{}: became leader with frontend state for {}", persistenceId(), knownFrontends.keySet());
 833         }
 834
 835         if (!isIsolatedLeader()) {
 836             messageRetrySupport.retryMessages();
 837         }
 838     }
 839
 840     /**
 841      * Clears all pending transactions and converts them to messages to be forwarded to a new leader.
 842      *
 843      * @return the converted messages
 844      */
 845     public Collection<?> convertPendingTransactionsToMessages() {
 846         return commitCoordinator.convertPendingTransactionsToMessages(
 847                 datastoreContext.getShardBatchedModificationCount());
 848     }
 849
 850     @Override
 851     protected void pauseLeader(final Runnable operation) {
 852         LOG.debug("{}: In pauseLeader, operation: {}", persistenceId(), operation);
 853         store.setRunOnPendingTransactionsComplete(operation);
 854     }
 855
 856     @Override
 857     protected OnDemandRaftState.AbstractBuilder<?> newOnDemandRaftStateBuilder() {
 858         return OnDemandShardState.newBuilder().treeChangeListenerActors(treeChangeSupport.getListenerActors())
 859                 .dataChangeListenerActors(changeSupport.getListenerActors())
 860                 .commitCohortActors(store.getCohortActors());
 861     }
 862
 863     @Override
 864     public String persistenceId() {
 865         return this.name;
 866     }
 867
 868     @VisibleForTesting
 869     ShardCommitCoordinator getCommitCoordinator() {
 870         return commitCoordinator;
 871     }
 872
 873     public DatastoreContext getDatastoreContext() {
 874         return datastoreContext;
 875     }
 876
 877     @VisibleForTesting
 878     public ShardDataTree getDataStore() {
 879         return store;
 880     }
 881
 882     @VisibleForTesting
 883     ShardStats getShardMBean() {
 884         return shardMBean;
 885     }
 886
 887     public static Builder builder() {
 888         return new Builder();
 889     }
 890
 891     public abstract static class AbstractBuilder<T extends AbstractBuilder<T, S>, S extends Shard> {
 892         private final Class<S> shardClass;
 893         private ShardIdentifier id;
 894         private Map<String, String> peerAddresses = Collections.emptyMap();
 895         private DatastoreContext datastoreContext;
 896         private SchemaContextProvider schemaContextProvider;
 897         private DatastoreSnapshot.ShardSnapshot restoreFromSnapshot;
 898         private TipProducingDataTree dataTree;
 899         private volatile boolean sealed;
 900
 901         protected AbstractBuilder(final Class<S> shardClass) {
 902             this.shardClass = shardClass;
 903         }
 904
 905         protected void checkSealed() {
 906             Preconditions.checkState(!sealed, "Builder isalready sealed - further modifications are not allowed");
 907         }
 908
 909         @SuppressWarnings("unchecked")
 910         private T self() {
 911             return (T) this;
 912         }
 913
 914         public T id(final ShardIdentifier newId) {
 915             checkSealed();
 916             this.id = newId;
 917             return self();
 918         }
 919
 920         public T peerAddresses(final Map<String, String> newPeerAddresses) {
 921             checkSealed();
 922             this.peerAddresses = newPeerAddresses;
 923             return self();
 924         }
 925
 926         public T datastoreContext(final DatastoreContext newDatastoreContext) {
 927             checkSealed();
 928             this.datastoreContext = newDatastoreContext;
 929             return self();
 930         }
 931
 932         public T schemaContextProvider(final SchemaContextProvider schemaContextProvider) {
 933             checkSealed();
 934             this.schemaContextProvider = Preconditions.checkNotNull(schemaContextProvider);
 935             return self();
 936         }
 937
 938         public T restoreFromSnapshot(final DatastoreSnapshot.ShardSnapshot newRestoreFromSnapshot) {
 939             checkSealed();
 940             this.restoreFromSnapshot = newRestoreFromSnapshot;
 941             return self();
 942         }
 943
 944         public T dataTree(final TipProducingDataTree newDataTree) {
 945             checkSealed();
 946             this.dataTree = newDataTree;
 947             return self();
 948         }
 949
 950         public ShardIdentifier getId() {
 951             return id;
 952         }
 953
 954         public Map<String, String> getPeerAddresses() {
 955             return peerAddresses;
 956         }
 957
 958         public DatastoreContext getDatastoreContext() {
 959             return datastoreContext;
 960         }
 961
 962         public SchemaContext getSchemaContext() {
 963             return Verify.verifyNotNull(schemaContextProvider.getSchemaContext());
 964         }
 965
 966         public DatastoreSnapshot.ShardSnapshot getRestoreFromSnapshot() {
 967             return restoreFromSnapshot;
 968         }
 969
 970         public TipProducingDataTree getDataTree() {
 971             return dataTree;
 972         }
 973
 974         public TreeType getTreeType() {
 975             switch (datastoreContext.getLogicalStoreType()) {
 976                 case CONFIGURATION:
 977                     return TreeType.CONFIGURATION;
 978                 case OPERATIONAL:
 979                     return TreeType.OPERATIONAL;
 980                 default:
 981                     throw new IllegalStateException("Unhandled logical store type "
 982                             + datastoreContext.getLogicalStoreType());
 983             }
 984         }
 985
 986         protected void verify() {
 987             Preconditions.checkNotNull(id, "id should not be null");
 988             Preconditions.checkNotNull(peerAddresses, "peerAddresses should not be null");
 989             Preconditions.checkNotNull(datastoreContext, "dataStoreContext should not be null");
 990             Preconditions.checkNotNull(schemaContextProvider, "schemaContextProvider should not be null");
 991         }
 992
 993         public Props props() {
 994             sealed = true;
 995             verify();
 996             return Props.create(shardClass, this);
 997         }
 998     }
 999
1000     public static class Builder extends AbstractBuilder<Builder, Shard> {
1001         private Builder() {
1002             super(Shard.class);
1003         }
1004     }
1005
1006     Ticker ticker() {
1007         return Ticker.systemTicker();
1008     }
1009
1010     void scheduleNextPendingTransaction() {
1011         self().tell(RESUME_NEXT_PENDING_TRANSACTION, ActorRef.noSender());
1012     }
1013 }