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.serialization.Serialization;
  16 import com.google.common.annotations.VisibleForTesting;
  17 import com.google.common.base.Optional;
  18 import com.google.common.base.Preconditions;
  19 import java.io.IOException;
  20 import java.util.Collection;
  21 import java.util.Collections;
  22 import java.util.Map;
  23 import java.util.concurrent.ExecutionException;
  24 import java.util.concurrent.TimeUnit;
  25 import javax.annotation.Nonnull;
  26 import org.opendaylight.controller.cluster.common.actor.CommonConfig;
  27 import org.opendaylight.controller.cluster.common.actor.MessageTracker;
  28 import org.opendaylight.controller.cluster.common.actor.MessageTracker.Error;
  29 import org.opendaylight.controller.cluster.common.actor.MeteringBehavior;
  30 import org.opendaylight.controller.cluster.datastore.exceptions.NoShardLeaderException;
  31 import org.opendaylight.controller.cluster.datastore.identifiers.ShardIdentifier;
  32 import org.opendaylight.controller.cluster.datastore.jmx.mbeans.shard.ShardMBeanFactory;
  33 import org.opendaylight.controller.cluster.datastore.jmx.mbeans.shard.ShardStats;
  34 import org.opendaylight.controller.cluster.datastore.messages.AbortTransaction;
  35 import org.opendaylight.controller.cluster.datastore.messages.ActorInitialized;
  36 import org.opendaylight.controller.cluster.datastore.messages.BatchedModifications;
  37 import org.opendaylight.controller.cluster.datastore.messages.CanCommitTransaction;
  38 import org.opendaylight.controller.cluster.datastore.messages.CloseTransactionChain;
  39 import org.opendaylight.controller.cluster.datastore.messages.CommitTransaction;
  40 import org.opendaylight.controller.cluster.datastore.messages.CommitTransactionReply;
  41 import org.opendaylight.controller.cluster.datastore.messages.CreateTransaction;
  42 import org.opendaylight.controller.cluster.datastore.messages.CreateTransactionReply;
  43 import org.opendaylight.controller.cluster.datastore.messages.DatastoreSnapshot;
  44 import org.opendaylight.controller.cluster.datastore.messages.DatastoreSnapshot.ShardSnapshot;
  45 import org.opendaylight.controller.cluster.datastore.messages.ForwardedReadyTransaction;
  46 import org.opendaylight.controller.cluster.datastore.messages.GetShardDataTree;
  47 import org.opendaylight.controller.cluster.datastore.messages.PeerAddressResolved;
  48 import org.opendaylight.controller.cluster.datastore.messages.ReadyLocalTransaction;
  49 import org.opendaylight.controller.cluster.datastore.messages.RegisterChangeListener;
  50 import org.opendaylight.controller.cluster.datastore.messages.RegisterDataTreeChangeListener;
  51 import org.opendaylight.controller.cluster.datastore.messages.ShardLeaderStateChanged;
  52 import org.opendaylight.controller.cluster.datastore.messages.UpdateSchemaContext;
  53 import org.opendaylight.controller.cluster.datastore.modification.Modification;
  54 import org.opendaylight.controller.cluster.datastore.modification.MutableCompositeModification;
  55 import org.opendaylight.controller.cluster.datastore.utils.Dispatchers;
  56 import org.opendaylight.controller.cluster.notifications.LeaderStateChanged;
  57 import org.opendaylight.controller.cluster.notifications.RegisterRoleChangeListener;
  58 import org.opendaylight.controller.cluster.notifications.RoleChangeNotifier;
  59 import org.opendaylight.controller.cluster.raft.RaftActor;
  60 import org.opendaylight.controller.cluster.raft.RaftActorRecoveryCohort;
  61 import org.opendaylight.controller.cluster.raft.RaftActorSnapshotCohort;
  62 import org.opendaylight.controller.cluster.raft.RaftState;
  63 import org.opendaylight.controller.cluster.raft.base.messages.FollowerInitialSyncUpStatus;
  64 import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
  65 import org.opendaylight.controller.cluster.raft.messages.ServerRemoved;
  66 import org.opendaylight.yangtools.concepts.Identifier;
  67 import org.opendaylight.yangtools.util.StringIdentifier;
  68 import org.opendaylight.yangtools.yang.data.api.schema.tree.DataTreeCandidate;
  69 import org.opendaylight.yangtools.yang.data.api.schema.tree.DataValidationFailedException;
  70 import org.opendaylight.yangtools.yang.data.api.schema.tree.ModificationType;
  71 import org.opendaylight.yangtools.yang.data.api.schema.tree.TreeType;
  72 import org.opendaylight.yangtools.yang.model.api.SchemaContext;
  73 import scala.concurrent.duration.Duration;
  74 import scala.concurrent.duration.FiniteDuration;
  75
  76 /**
  77  * A Shard represents a portion of the logical data tree <br/>
  78  * <p>
  79  * Our Shard uses InMemoryDataTree as it's internal representation and delegates all requests it
  80  * </p>
  81  */
  82 public class Shard extends RaftActor {
  83
  84     @VisibleForTesting
  85     static final Object TX_COMMIT_TIMEOUT_CHECK_MESSAGE = new Object() {
  86         @Override
  87         public String toString() {
  88             return "txCommitTimeoutCheck";
  89         }
  90     };
  91
  92     @VisibleForTesting
  93     static final Object GET_SHARD_MBEAN_MESSAGE = new Object() {
  94         @Override
  95         public String toString() {
  96             return "getShardMBeanMessage";
  97         }
  98     };
  99
 100     // FIXME: shard names should be encapsulated in their own class and this should be exposed as a constant.
 101     public static final String DEFAULT_NAME = "default";
 102
 103     // The state of this Shard
 104     private final ShardDataTree store;
 105
 106     /// The name of this shard
 107     private final String name;
 108
 109     private final ShardStats shardMBean;
 110
 111     private DatastoreContext datastoreContext;
 112
 113     private final ShardCommitCoordinator commitCoordinator;
 114
 115     private long transactionCommitTimeout;
 116
 117     private Cancellable txCommitTimeoutCheckSchedule;
 118
 119     private final Optional<ActorRef> roleChangeNotifier;
 120
 121     private final MessageTracker appendEntriesReplyTracker;
 122
 123     private final ShardTransactionActorFactory transactionActorFactory;
 124
 125     private final ShardSnapshotCohort snapshotCohort;
 126
 127     private final DataTreeChangeListenerSupport treeChangeSupport = new DataTreeChangeListenerSupport(this);
 128     private final DataChangeListenerSupport changeSupport = new DataChangeListenerSupport(this);
 129
 130
 131     private ShardSnapshot restoreFromSnapshot;
 132
 133     private final ShardTransactionMessageRetrySupport messageRetrySupport;
 134
 135     protected Shard(AbstractBuilder<?, ?> builder) {
 136         super(builder.getId().toString(), builder.getPeerAddresses(),
 137                 Optional.of(builder.getDatastoreContext().getShardRaftConfig()), DataStoreVersions.CURRENT_VERSION);
 138
 139         this.name = builder.getId().toString();
 140         this.datastoreContext = builder.getDatastoreContext();
 141         this.restoreFromSnapshot = builder.getRestoreFromSnapshot();
 142
 143         setPersistence(datastoreContext.isPersistent());
 144
 145         LOG.info("Shard created : {}, persistent : {}", name, datastoreContext.isPersistent());
 146
 147         store = new ShardDataTree(builder.getSchemaContext(), builder.getTreeType(),
 148                 new ShardDataTreeChangeListenerPublisherActorProxy(getContext(), name + "-DTCL-publisher"),
 149                 new ShardDataChangeListenerPublisherActorProxy(getContext(), name + "-DCL-publisher"), name);
 150
 151         shardMBean = ShardMBeanFactory.getShardStatsMBean(name.toString(),
 152                 datastoreContext.getDataStoreMXBeanType());
 153         shardMBean.setShard(this);
 154
 155         if (isMetricsCaptureEnabled()) {
 156             getContext().become(new MeteringBehavior(this));
 157         }
 158
 159         commitCoordinator = new ShardCommitCoordinator(store,
 160                 datastoreContext.getShardCommitQueueExpiryTimeoutInMillis(),
 161                 datastoreContext.getShardTransactionCommitQueueCapacity(), LOG, this.name);
 162
 163         setTransactionCommitTimeout();
 164
 165         // create a notifier actor for each cluster member
 166         roleChangeNotifier = createRoleChangeNotifier(name.toString());
 167
 168         appendEntriesReplyTracker = new MessageTracker(AppendEntriesReply.class,
 169                 getRaftActorContext().getConfigParams().getIsolatedCheckIntervalInMillis());
 170
 171         transactionActorFactory = new ShardTransactionActorFactory(store, datastoreContext,
 172                 new Dispatchers(context().system().dispatchers()).getDispatcherPath(
 173                         Dispatchers.DispatcherType.Transaction), self(), getContext(), shardMBean);
 174
 175         snapshotCohort = new ShardSnapshotCohort(transactionActorFactory, store, LOG, this.name);
 176
 177         messageRetrySupport = new ShardTransactionMessageRetrySupport(this);
 178     }
 179
 180     private void setTransactionCommitTimeout() {
 181         transactionCommitTimeout = TimeUnit.MILLISECONDS.convert(
 182                 datastoreContext.getShardTransactionCommitTimeoutInSeconds(), TimeUnit.SECONDS) / 2;
 183     }
 184
 185     private Optional<ActorRef> createRoleChangeNotifier(String shardId) {
 186         ActorRef shardRoleChangeNotifier = this.getContext().actorOf(
 187             RoleChangeNotifier.getProps(shardId), shardId + "-notifier");
 188         return Optional.of(shardRoleChangeNotifier);
 189     }
 190
 191     @Override
 192     public void postStop() {
 193         LOG.info("Stopping Shard {}", persistenceId());
 194
 195         super.postStop();
 196
 197         messageRetrySupport.close();
 198
 199         if(txCommitTimeoutCheckSchedule != null) {
 200             txCommitTimeoutCheckSchedule.cancel();
 201         }
 202
 203         commitCoordinator.abortPendingTransactions("Transaction aborted due to shutdown.", this);
 204
 205         shardMBean.unregisterMBean();
 206     }
 207
 208     @Override
 209     protected void handleRecover(final Object message) {
 210         LOG.debug("{}: onReceiveRecover: Received message {} from {}", persistenceId(), message.getClass(),
 211             getSender());
 212
 213         super.handleRecover(message);
 214         if (LOG.isTraceEnabled()) {
 215             appendEntriesReplyTracker.begin();
 216         }
 217     }
 218
 219     @Override
 220     protected void handleNonRaftCommand(final Object message) {
 221         try (final MessageTracker.Context context = appendEntriesReplyTracker.received(message)) {
 222             final Optional<Error> maybeError = context.error();
 223             if (maybeError.isPresent()) {
 224                 LOG.trace("{} : AppendEntriesReply failed to arrive at the expected interval {}", persistenceId(),
 225                     maybeError.get());
 226             }
 227
 228             if (CreateTransaction.isSerializedType(message)) {
 229                 handleCreateTransaction(message);
 230             } else if (message instanceof BatchedModifications) {
 231                 handleBatchedModifications((BatchedModifications)message);
 232             } else if (message instanceof ForwardedReadyTransaction) {
 233                 handleForwardedReadyTransaction((ForwardedReadyTransaction) message);
 234             } else if (message instanceof ReadyLocalTransaction) {
 235                 handleReadyLocalTransaction((ReadyLocalTransaction)message);
 236             } else if (CanCommitTransaction.isSerializedType(message)) {
 237                 handleCanCommitTransaction(CanCommitTransaction.fromSerializable(message));
 238             } else if (CommitTransaction.isSerializedType(message)) {
 239                 handleCommitTransaction(CommitTransaction.fromSerializable(message));
 240             } else if (AbortTransaction.isSerializedType(message)) {
 241                 handleAbortTransaction(AbortTransaction.fromSerializable(message));
 242             } else if (CloseTransactionChain.isSerializedType(message)) {
 243                 closeTransactionChain(CloseTransactionChain.fromSerializable(message));
 244             } else if (message instanceof RegisterChangeListener) {
 245                 changeSupport.onMessage((RegisterChangeListener) message, isLeader(), hasLeader());
 246             } else if (message instanceof RegisterDataTreeChangeListener) {
 247                 treeChangeSupport.onMessage((RegisterDataTreeChangeListener) message, isLeader(), hasLeader());
 248             } else if (message instanceof UpdateSchemaContext) {
 249                 updateSchemaContext((UpdateSchemaContext) message);
 250             } else if (message instanceof PeerAddressResolved) {
 251                 PeerAddressResolved resolved = (PeerAddressResolved) message;
 252                 setPeerAddress(resolved.getPeerId().toString(),
 253                         resolved.getPeerAddress());
 254             } else if (TX_COMMIT_TIMEOUT_CHECK_MESSAGE.equals(message)) {
 255                 commitCoordinator.checkForExpiredTransactions(transactionCommitTimeout, this);
 256             } else if(message instanceof DatastoreContext) {
 257                 onDatastoreContext((DatastoreContext)message);
 258             } else if(message instanceof RegisterRoleChangeListener){
 259                 roleChangeNotifier.get().forward(message, context());
 260             } else if (message instanceof FollowerInitialSyncUpStatus) {
 261                 shardMBean.setFollowerInitialSyncStatus(((FollowerInitialSyncUpStatus) message).isInitialSyncDone());
 262                 context().parent().tell(message, self());
 263             } else if(GET_SHARD_MBEAN_MESSAGE.equals(message)){
 264                 sender().tell(getShardMBean(), self());
 265             } else if(message instanceof GetShardDataTree) {
 266                 sender().tell(store.getDataTree(), self());
 267             } else if(message instanceof ServerRemoved){
 268                 context().parent().forward(message, context());
 269             } else if(ShardTransactionMessageRetrySupport.TIMER_MESSAGE_CLASS.isInstance(message)) {
 270                 messageRetrySupport.onTimerMessage(message);
 271             } else if (message instanceof DataTreeCohortActorRegistry.CohortRegistryCommand) {
 272                 commitCoordinator.processCohortRegistryCommand(getSender(),
 273                         (DataTreeCohortActorRegistry.CohortRegistryCommand) message);
 274             } else {
 275                 super.handleNonRaftCommand(message);
 276             }
 277         }
 278     }
 279
 280     private boolean hasLeader() {
 281         return getLeaderId() != null;
 282     }
 283
 284     public int getPendingTxCommitQueueSize() {
 285         return commitCoordinator.getQueueSize();
 286     }
 287
 288     public int getCohortCacheSize() {
 289         return commitCoordinator.getCohortCacheSize();
 290     }
 291
 292     @Override
 293     protected Optional<ActorRef> getRoleChangeNotifier() {
 294         return roleChangeNotifier;
 295     }
 296
 297     @Override
 298     protected LeaderStateChanged newLeaderStateChanged(String memberId, String leaderId, short leaderPayloadVersion) {
 299         return isLeader() ? new ShardLeaderStateChanged(memberId, leaderId, store.getDataTree(), leaderPayloadVersion)
 300                 : new ShardLeaderStateChanged(memberId, leaderId, leaderPayloadVersion);
 301     }
 302
 303     protected void onDatastoreContext(DatastoreContext context) {
 304         datastoreContext = context;
 305
 306         commitCoordinator.setQueueCapacity(datastoreContext.getShardTransactionCommitQueueCapacity());
 307
 308         setTransactionCommitTimeout();
 309
 310         if(datastoreContext.isPersistent() && !persistence().isRecoveryApplicable()) {
 311             setPersistence(true);
 312         } else if(!datastoreContext.isPersistent() && persistence().isRecoveryApplicable()) {
 313             setPersistence(false);
 314         }
 315
 316         updateConfigParams(datastoreContext.getShardRaftConfig());
 317     }
 318
 319     private static boolean isEmptyCommit(final DataTreeCandidate candidate) {
 320         return ModificationType.UNMODIFIED.equals(candidate.getRootNode().getModificationType());
 321     }
 322
 323     void continueCommit(final CohortEntry cohortEntry) {
 324         final DataTreeCandidate candidate = cohortEntry.getCandidate();
 325
 326         // If we do not have any followers and we are not using persistence
 327         // or if cohortEntry has no modifications
 328         // we can apply modification to the state immediately
 329         if ((!hasFollowers() && !persistence().isRecoveryApplicable()) || isEmptyCommit(candidate)) {
 330             applyModificationToState(cohortEntry.getReplySender(), cohortEntry.getTransactionID(), candidate);
 331         } else {
 332             persistData(cohortEntry.getReplySender(), cohortEntry.getTransactionID(),
 333                     DataTreeCandidatePayload.create(candidate));
 334         }
 335     }
 336
 337     private void handleCommitTransaction(final CommitTransaction commit) {
 338         if (isLeader()) {
 339             if(!commitCoordinator.handleCommit(new StringIdentifier(commit.getTransactionID()), getSender(), this)) {
 340                 shardMBean.incrementFailedTransactionsCount();
 341             }
 342         } else {
 343             ActorSelection leader = getLeader();
 344             if (leader == null) {
 345                 messageRetrySupport.addMessageToRetry(commit, getSender(),
 346                         "Could not commit transaction " + commit.getTransactionID());
 347             } else {
 348                 LOG.debug("{}: Forwarding CommitTransaction to leader {}", persistenceId(), leader);
 349                 leader.forward(commit, getContext());
 350             }
 351         }
 352     }
 353
 354     private void finishCommit(@Nonnull final ActorRef sender, @Nonnull final Identifier transactionID,
 355             @Nonnull final CohortEntry cohortEntry) {
 356         LOG.debug("{}: Finishing commit for transaction {}", persistenceId(), cohortEntry.getTransactionID());
 357
 358         try {
 359             try {
 360             cohortEntry.commit();
 361             } catch(ExecutionException e) {
 362                 // We may get a "store tree and candidate base differ" IllegalStateException from commit under
 363                 // certain edge case scenarios so we'll try to re-apply the candidate from scratch as a last
 364                 // resort. Eg, we're a follower and a tx payload is replicated but the leader goes down before
 365                 // applying it to the state. We then become the leader and a second tx is pre-committed and
 366                 // replicated. When consensus occurs, this will cause the first tx to be applied as a foreign
 367                 // candidate via applyState prior to the second tx. Since the second tx has already been
 368                 // pre-committed, when it gets here to commit it will get an IllegalStateException.
 369
 370                 // FIXME - this is not an ideal way to handle this scenario. This is temporary - a cleaner
 371                 // solution will be forthcoming.
 372                 if(e.getCause() instanceof IllegalStateException) {
 373                     LOG.debug("{}: commit failed for transaction {} - retrying as foreign candidate", persistenceId(),
 374                             transactionID, e);
 375                     store.applyForeignCandidate(transactionID, cohortEntry.getCandidate());
 376                 } else {
 377                     throw e;
 378                 }
 379             }
 380
 381             sender.tell(CommitTransactionReply.instance(cohortEntry.getClientVersion()).toSerializable(), getSelf());
 382
 383             shardMBean.incrementCommittedTransactionCount();
 384             shardMBean.setLastCommittedTransactionTime(System.currentTimeMillis());
 385
 386         } catch (Exception e) {
 387             sender.tell(new akka.actor.Status.Failure(e), getSelf());
 388
 389             LOG.error("{}, An exception occurred while committing transaction {}", persistenceId(),
 390                     transactionID, e);
 391             shardMBean.incrementFailedTransactionsCount();
 392         } finally {
 393             commitCoordinator.currentTransactionComplete(transactionID, true);
 394         }
 395     }
 396
 397     private void finishCommit(@Nonnull final ActorRef sender, final @Nonnull Identifier transactionID) {
 398         // With persistence enabled, this method is called via applyState by the leader strategy
 399         // after the commit has been replicated to a majority of the followers.
 400
 401         CohortEntry cohortEntry = commitCoordinator.getCohortEntryIfCurrent(transactionID);
 402         if (cohortEntry == null) {
 403             // The transaction is no longer the current commit. This can happen if the transaction
 404             // was aborted prior, most likely due to timeout in the front-end. We need to finish
 405             // committing the transaction though since it was successfully persisted and replicated
 406             // however we can't use the original cohort b/c it was already preCommitted and may
 407             // conflict with the current commit or may have been aborted so we commit with a new
 408             // transaction.
 409             cohortEntry = commitCoordinator.getAndRemoveCohortEntry(transactionID);
 410             if(cohortEntry != null) {
 411                 try {
 412                     store.applyForeignCandidate(transactionID, cohortEntry.getCandidate());
 413                 } catch (DataValidationFailedException e) {
 414                     shardMBean.incrementFailedTransactionsCount();
 415                     LOG.error("{}: Failed to re-apply transaction {}", persistenceId(), transactionID, e);
 416                 }
 417
 418                 sender.tell(CommitTransactionReply.instance(cohortEntry.getClientVersion()).toSerializable(),
 419                         getSelf());
 420             } else {
 421                 // This really shouldn't happen - it likely means that persistence or replication
 422                 // took so long to complete such that the cohort entry was expired from the cache.
 423                 IllegalStateException ex = new IllegalStateException(
 424                         String.format("%s: Could not finish committing transaction %s - no CohortEntry found",
 425                                 persistenceId(), transactionID));
 426                 LOG.error(ex.getMessage());
 427                 sender.tell(new akka.actor.Status.Failure(ex), getSelf());
 428             }
 429         } else {
 430             finishCommit(sender, transactionID, cohortEntry);
 431         }
 432     }
 433
 434     private void handleCanCommitTransaction(final CanCommitTransaction canCommit) {
 435         LOG.debug("{}: Can committing transaction {}", persistenceId(), canCommit.getTransactionID());
 436
 437         if (isLeader()) {
 438         commitCoordinator.handleCanCommit(new StringIdentifier(canCommit.getTransactionID()), getSender(), this);
 439         } else {
 440             ActorSelection leader = getLeader();
 441             if (leader == null) {
 442                 messageRetrySupport.addMessageToRetry(canCommit, getSender(),
 443                         "Could not canCommit transaction " + canCommit.getTransactionID());
 444             } else {
 445                 LOG.debug("{}: Forwarding CanCommitTransaction to leader {}", persistenceId(), leader);
 446                 leader.forward(canCommit, getContext());
 447             }
 448         }
 449     }
 450
 451     protected void handleBatchedModificationsLocal(BatchedModifications batched, ActorRef sender) {
 452         try {
 453             commitCoordinator.handleBatchedModifications(batched, sender, this, store.getSchemaContext());
 454         } catch (Exception e) {
 455             LOG.error("{}: Error handling BatchedModifications for Tx {}", persistenceId(),
 456                     batched.getTransactionID(), e);
 457             sender.tell(new akka.actor.Status.Failure(e), getSelf());
 458         }
 459     }
 460
 461     private void handleBatchedModifications(BatchedModifications batched) {
 462         // This message is sent to prepare the modifications transaction directly on the Shard as an
 463         // optimization to avoid the extra overhead of a separate ShardTransaction actor. On the last
 464         // BatchedModifications message, the caller sets the ready flag in the message indicating
 465         // modifications are complete. The reply contains the cohort actor path (this actor) for the caller
 466         // to initiate the 3-phase commit. This also avoids the overhead of sending an additional
 467         // ReadyTransaction message.
 468
 469         // If we're not the leader then forward to the leader. This is a safety measure - we shouldn't
 470         // normally get here if we're not the leader as the front-end (TransactionProxy) should determine
 471         // the primary/leader shard. However with timing and caching on the front-end, there's a small
 472         // window where it could have a stale leader during leadership transitions.
 473         //
 474         boolean isLeaderActive = isLeaderActive();
 475         if (isLeader() && isLeaderActive) {
 476             handleBatchedModificationsLocal(batched, getSender());
 477         } else {
 478             ActorSelection leader = getLeader();
 479             if (!isLeaderActive || leader == null) {
 480                 messageRetrySupport.addMessageToRetry(batched, getSender(),
 481                         "Could not commit transaction " + batched.getTransactionID());
 482             } else {
 483                 // If this is not the first batch and leadership changed in between batched messages,
 484                 // we need to reconstruct previous BatchedModifications from the transaction
 485                 // DataTreeModification, honoring the max batched modification count, and forward all the
 486                 // previous BatchedModifications to the new leader.
 487                 Collection<BatchedModifications> newModifications = commitCoordinator.createForwardedBatchedModifications(
 488                         batched, datastoreContext.getShardBatchedModificationCount());
 489
 490                 LOG.debug("{}: Forwarding {} BatchedModifications to leader {}", persistenceId(),
 491                         newModifications.size(), leader);
 492
 493                 for(BatchedModifications bm: newModifications) {
 494                     leader.forward(bm, getContext());
 495                 }
 496             }
 497         }
 498     }
 499
 500     private boolean failIfIsolatedLeader(ActorRef sender) {
 501         if(isIsolatedLeader()) {
 502             sender.tell(new akka.actor.Status.Failure(new NoShardLeaderException(String.format(
 503                     "Shard %s was the leader but has lost contact with all of its followers. Either all" +
 504                     " other follower nodes are down or this node is isolated by a network partition.",
 505                     persistenceId()))), getSelf());
 506             return true;
 507         }
 508
 509         return false;
 510     }
 511
 512     protected boolean isIsolatedLeader() {
 513         return getRaftState() == RaftState.IsolatedLeader;
 514     }
 515
 516     private void handleReadyLocalTransaction(final ReadyLocalTransaction message) {
 517         LOG.debug("{}: handleReadyLocalTransaction for {}", persistenceId(), message.getTransactionID());
 518
 519         boolean isLeaderActive = isLeaderActive();
 520         if (isLeader() && isLeaderActive) {
 521             try {
 522                 commitCoordinator.handleReadyLocalTransaction(message, getSender(), this, store.getSchemaContext());
 523             } catch (Exception e) {
 524                 LOG.error("{}: Error handling ReadyLocalTransaction for Tx {}", persistenceId(),
 525                         message.getTransactionID(), e);
 526                 getSender().tell(new akka.actor.Status.Failure(e), getSelf());
 527             }
 528         } else {
 529             ActorSelection leader = getLeader();
 530             if (!isLeaderActive || leader == null) {
 531                 messageRetrySupport.addMessageToRetry(message, getSender(),
 532                         "Could not commit transaction " + message.getTransactionID());
 533             } else {
 534                 LOG.debug("{}: Forwarding ReadyLocalTransaction to leader {}", persistenceId(), leader);
 535                 message.setRemoteVersion(getCurrentBehavior().getLeaderPayloadVersion());
 536                 leader.forward(message, getContext());
 537             }
 538         }
 539     }
 540
 541     private void handleForwardedReadyTransaction(ForwardedReadyTransaction forwardedReady) {
 542         LOG.debug("{}: handleForwardedReadyTransaction for {}", persistenceId(), forwardedReady.getTransactionID());
 543
 544         boolean isLeaderActive = isLeaderActive();
 545         if (isLeader() && isLeaderActive) {
 546             commitCoordinator.handleForwardedReadyTransaction(forwardedReady, getSender(), this,
 547                     store.getSchemaContext());
 548         } else {
 549             ActorSelection leader = getLeader();
 550             if (!isLeaderActive || leader == null) {
 551                 messageRetrySupport.addMessageToRetry(forwardedReady, getSender(),
 552                         "Could not commit transaction " + forwardedReady.getTransactionID());
 553             } else {
 554                 LOG.debug("{}: Forwarding ForwardedReadyTransaction to leader {}", persistenceId(), leader);
 555
 556                 ReadyLocalTransaction readyLocal = new ReadyLocalTransaction(forwardedReady.getTransactionID(),
 557                         forwardedReady.getTransaction().getSnapshot(), forwardedReady.isDoImmediateCommit());
 558                 readyLocal.setRemoteVersion(getCurrentBehavior().getLeaderPayloadVersion());
 559                 leader.forward(readyLocal, getContext());
 560             }
 561         }
 562     }
 563
 564     private void handleAbortTransaction(final AbortTransaction abort) {
 565         doAbortTransaction(abort.getTransactionID(), getSender());
 566     }
 567
 568     void doAbortTransaction(final String transactionID, final ActorRef sender) {
 569         commitCoordinator.handleAbort(new StringIdentifier(transactionID), sender, this);
 570     }
 571
 572     private void handleCreateTransaction(final Object message) {
 573         if (isLeader()) {
 574             createTransaction(CreateTransaction.fromSerializable(message));
 575         } else if (getLeader() != null) {
 576             getLeader().forward(message, getContext());
 577         } else {
 578             getSender().tell(new akka.actor.Status.Failure(new NoShardLeaderException(
 579                     "Could not create a shard transaction", persistenceId())), getSelf());
 580         }
 581     }
 582
 583     private void closeTransactionChain(final CloseTransactionChain closeTransactionChain) {
 584         store.closeTransactionChain(closeTransactionChain.getTransactionChainId());
 585     }
 586
 587     private void createTransaction(CreateTransaction createTransaction) {
 588         try {
 589             if(TransactionType.fromInt(createTransaction.getTransactionType()) != TransactionType.READ_ONLY &&
 590                     failIfIsolatedLeader(getSender())) {
 591                 return;
 592             }
 593
 594             ActorRef transactionActor = createTransaction(createTransaction.getTransactionType(),
 595                 createTransaction.getTransactionId(), createTransaction.getTransactionChainId());
 596
 597             getSender().tell(new CreateTransactionReply(Serialization.serializedActorPath(transactionActor),
 598                     createTransaction.getTransactionId(), createTransaction.getVersion()).toSerializable(), getSelf());
 599         } catch (Exception e) {
 600             getSender().tell(new akka.actor.Status.Failure(e), getSelf());
 601         }
 602     }
 603
 604     private ActorRef createTransaction(int transactionType, String transactionId, String transactionChainId) {
 605         LOG.debug("{}: Creating transaction : {} ", persistenceId(), transactionId);
 606         return transactionActorFactory.newShardTransaction(TransactionType.fromInt(transactionType),
 607             transactionId, transactionChainId);
 608     }
 609
 610     private void commitWithNewTransaction(final Modification modification) {
 611         ReadWriteShardDataTreeTransaction tx = store.newReadWriteTransaction(modification.toString(), null);
 612         modification.apply(tx.getSnapshot());
 613         try {
 614             snapshotCohort.syncCommitTransaction(tx);
 615             shardMBean.incrementCommittedTransactionCount();
 616             shardMBean.setLastCommittedTransactionTime(System.currentTimeMillis());
 617         } catch (Exception e) {
 618             shardMBean.incrementFailedTransactionsCount();
 619             LOG.error("{}: Failed to commit", persistenceId(), e);
 620         }
 621     }
 622
 623     private void updateSchemaContext(final UpdateSchemaContext message) {
 624         updateSchemaContext(message.getSchemaContext());
 625     }
 626
 627     @VisibleForTesting
 628     void updateSchemaContext(final SchemaContext schemaContext) {
 629         store.updateSchemaContext(schemaContext);
 630     }
 631
 632     private boolean isMetricsCaptureEnabled() {
 633         CommonConfig config = new CommonConfig(getContext().system().settings().config());
 634         return config.isMetricCaptureEnabled();
 635     }
 636
 637     @Override
 638     @VisibleForTesting
 639     public RaftActorSnapshotCohort getRaftActorSnapshotCohort() {
 640         return snapshotCohort;
 641     }
 642
 643     @Override
 644     @Nonnull
 645     protected RaftActorRecoveryCohort getRaftActorRecoveryCohort() {
 646         return new ShardRecoveryCoordinator(store, store.getSchemaContext(),
 647                 restoreFromSnapshot != null ? restoreFromSnapshot.getSnapshot() : null, persistenceId(), LOG);
 648     }
 649
 650     @Override
 651     protected void onRecoveryComplete() {
 652         restoreFromSnapshot = null;
 653
 654         //notify shard manager
 655         getContext().parent().tell(new ActorInitialized(), getSelf());
 656
 657         // Being paranoid here - this method should only be called once but just in case...
 658         if(txCommitTimeoutCheckSchedule == null) {
 659             // Schedule a message to be periodically sent to check if the current in-progress
 660             // transaction should be expired and aborted.
 661             FiniteDuration period = Duration.create(transactionCommitTimeout / 3, TimeUnit.MILLISECONDS);
 662             txCommitTimeoutCheckSchedule = getContext().system().scheduler().schedule(
 663                     period, period, getSelf(),
 664                     TX_COMMIT_TIMEOUT_CHECK_MESSAGE, getContext().dispatcher(), ActorRef.noSender());
 665         }
 666     }
 667
 668     @Override
 669     protected void applyState(final ActorRef clientActor, final Identifier identifier, final Object data) {
 670         if (data instanceof DataTreeCandidatePayload) {
 671             if (clientActor == null) {
 672                 // No clientActor indicates a replica coming from the leader
 673                 try {
 674                     store.applyForeignCandidate(identifier, ((DataTreeCandidatePayload)data).getCandidate());
 675                 } catch (DataValidationFailedException | IOException e) {
 676                     LOG.error("{}: Error applying replica {}", persistenceId(), identifier, e);
 677                 }
 678             } else {
 679                 // Replication consensus reached, proceed to commit
 680                 finishCommit(clientActor, identifier);
 681             }
 682         } else {
 683             LOG.error("{}: Unknown state received {} ClassLoader {}", persistenceId(), data,
 684                 data.getClass().getClassLoader());
 685         }
 686     }
 687
 688     private void applyModificationToState(ActorRef clientActor, Identifier identifier, Object modification) {
 689         if(modification == null) {
 690             LOG.error(
 691                     "{}: modification is null - this is very unexpected, clientActor = {}, identifier = {}",
 692                     persistenceId(), identifier, clientActor != null ? clientActor.path().toString() : null);
 693         } else if(clientActor == null) {
 694             // There's no clientActor to which to send a commit reply so we must be applying
 695             // replicated state from the leader.
 696             commitWithNewTransaction(MutableCompositeModification.fromSerializable(modification));
 697         } else {
 698             // This must be the OK to commit after replication consensus.
 699             finishCommit(clientActor, identifier);
 700         }
 701     }
 702
 703     @Override
 704     protected void onStateChanged() {
 705         boolean isLeader = isLeader();
 706         boolean hasLeader = hasLeader();
 707         changeSupport.onLeadershipChange(isLeader, hasLeader);
 708         treeChangeSupport.onLeadershipChange(isLeader, hasLeader);
 709
 710         // If this actor is no longer the leader close all the transaction chains
 711         if (!isLeader) {
 712             if(LOG.isDebugEnabled()) {
 713                 LOG.debug(
 714                     "{}: onStateChanged: Closing all transaction chains because shard {} is no longer the leader",
 715                     persistenceId(), getId());
 716             }
 717
 718             store.closeAllTransactionChains();
 719         }
 720
 721         if(hasLeader && !isIsolatedLeader()) {
 722             messageRetrySupport.retryMessages();
 723         }
 724     }
 725
 726     @Override
 727     protected void onLeaderChanged(String oldLeader, String newLeader) {
 728         shardMBean.incrementLeadershipChangeCount();
 729
 730         boolean hasLeader = hasLeader();
 731         if(hasLeader && !isLeader()) {
 732             // Another leader was elected. If we were the previous leader and had pending transactions, convert
 733             // them to transaction messages and send to the new leader.
 734             ActorSelection leader = getLeader();
 735             if(leader != null) {
 736                 Collection<Object> messagesToForward = commitCoordinator.convertPendingTransactionsToMessages(
 737                         datastoreContext.getShardBatchedModificationCount());
 738
 739                 if(!messagesToForward.isEmpty()) {
 740                     LOG.debug("{}: Forwarding {} pending transaction messages to leader {}", persistenceId(),
 741                             messagesToForward.size(), leader);
 742
 743                     for(Object message: messagesToForward) {
 744                         leader.tell(message, self());
 745                     }
 746                 }
 747             } else {
 748                 commitCoordinator.abortPendingTransactions(
 749                         "The transacton was aborted due to inflight leadership change and the leader address isn't available.",
 750                         this);
 751             }
 752         }
 753
 754         if(hasLeader && !isIsolatedLeader()) {
 755             messageRetrySupport.retryMessages();
 756         }
 757     }
 758
 759     @Override
 760     protected void pauseLeader(Runnable operation) {
 761         LOG.debug("{}: In pauseLeader, operation: {}", persistenceId(), operation);
 762         commitCoordinator.setRunOnPendingTransactionsComplete(operation);
 763     }
 764
 765     @Override
 766     public String persistenceId() {
 767         return this.name;
 768     }
 769
 770     @VisibleForTesting
 771     ShardCommitCoordinator getCommitCoordinator() {
 772         return commitCoordinator;
 773     }
 774
 775     public DatastoreContext getDatastoreContext() {
 776         return datastoreContext;
 777     }
 778
 779     @VisibleForTesting
 780     public ShardDataTree getDataStore() {
 781         return store;
 782     }
 783
 784     @VisibleForTesting
 785     ShardStats getShardMBean() {
 786         return shardMBean;
 787     }
 788
 789     public static Builder builder() {
 790         return new Builder();
 791     }
 792
 793     public static abstract class AbstractBuilder<T extends AbstractBuilder<T, S>, S extends Shard> {
 794         private final Class<S> shardClass;
 795         private ShardIdentifier id;
 796         private Map<String, String> peerAddresses = Collections.emptyMap();
 797         private DatastoreContext datastoreContext;
 798         private SchemaContext schemaContext;
 799         private DatastoreSnapshot.ShardSnapshot restoreFromSnapshot;
 800         private volatile boolean sealed;
 801
 802         protected AbstractBuilder(Class<S> shardClass) {
 803             this.shardClass = shardClass;
 804         }
 805
 806         protected void checkSealed() {
 807             Preconditions.checkState(!sealed, "Builder isalready sealed - further modifications are not allowed");
 808         }
 809
 810         @SuppressWarnings("unchecked")
 811         private T self() {
 812             return (T) this;
 813         }
 814
 815         public T id(ShardIdentifier id) {
 816             checkSealed();
 817             this.id = id;
 818             return self();
 819         }
 820
 821         public T peerAddresses(Map<String, String> peerAddresses) {
 822             checkSealed();
 823             this.peerAddresses = peerAddresses;
 824             return self();
 825         }
 826
 827         public T datastoreContext(DatastoreContext datastoreContext) {
 828             checkSealed();
 829             this.datastoreContext = datastoreContext;
 830             return self();
 831         }
 832
 833         public T schemaContext(SchemaContext schemaContext) {
 834             checkSealed();
 835             this.schemaContext = schemaContext;
 836             return self();
 837         }
 838
 839         public T restoreFromSnapshot(DatastoreSnapshot.ShardSnapshot restoreFromSnapshot) {
 840             checkSealed();
 841             this.restoreFromSnapshot = restoreFromSnapshot;
 842             return self();
 843         }
 844
 845         public ShardIdentifier getId() {
 846             return id;
 847         }
 848
 849         public Map<String, String> getPeerAddresses() {
 850             return peerAddresses;
 851         }
 852
 853         public DatastoreContext getDatastoreContext() {
 854             return datastoreContext;
 855         }
 856
 857         public SchemaContext getSchemaContext() {
 858             return schemaContext;
 859         }
 860
 861         public DatastoreSnapshot.ShardSnapshot getRestoreFromSnapshot() {
 862             return restoreFromSnapshot;
 863         }
 864
 865         public TreeType getTreeType() {
 866             switch (datastoreContext.getLogicalStoreType()) {
 867             case CONFIGURATION:
 868                 return TreeType.CONFIGURATION;
 869             case OPERATIONAL:
 870                 return TreeType.OPERATIONAL;
 871             }
 872
 873             throw new IllegalStateException("Unhandled logical store type " + datastoreContext.getLogicalStoreType());
 874         }
 875
 876         protected void verify() {
 877             Preconditions.checkNotNull(id, "id should not be null");
 878             Preconditions.checkNotNull(peerAddresses, "peerAddresses should not be null");
 879             Preconditions.checkNotNull(datastoreContext, "dataStoreContext should not be null");
 880             Preconditions.checkNotNull(schemaContext, "schemaContext should not be null");
 881         }
 882
 883         public Props props() {
 884             sealed = true;
 885             verify();
 886             return Props.create(shardClass, this);
 887         }
 888     }
 889
 890     public static class Builder extends AbstractBuilder<Builder, Shard> {
 891         private Builder() {
 892             super(Shard.class);
 893         }
 894     }
 895 }