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