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.japi.Creator;
  16 import akka.persistence.RecoveryFailure;
  17 import akka.serialization.Serialization;
  18 import com.google.common.annotations.VisibleForTesting;
  19 import com.google.common.base.Optional;
  20 import com.google.common.base.Preconditions;
  21 import com.google.common.util.concurrent.FutureCallback;
  22 import com.google.common.util.concurrent.Futures;
  23 import com.google.common.util.concurrent.ListenableFuture;
  24 import java.io.IOException;
  25 import java.util.HashMap;
  26 import java.util.Map;
  27 import java.util.concurrent.TimeUnit;
  28 import javax.annotation.Nonnull;
  29 import org.opendaylight.controller.cluster.common.actor.CommonConfig;
  30 import org.opendaylight.controller.cluster.common.actor.MeteringBehavior;
  31 import org.opendaylight.controller.cluster.datastore.ShardCommitCoordinator.CohortEntry;
  32 import org.opendaylight.controller.cluster.datastore.exceptions.NoShardLeaderException;
  33 import org.opendaylight.controller.cluster.datastore.identifiers.ShardIdentifier;
  34 import org.opendaylight.controller.cluster.datastore.identifiers.ShardTransactionIdentifier;
  35 import org.opendaylight.controller.cluster.datastore.jmx.mbeans.shard.ShardMBeanFactory;
  36 import org.opendaylight.controller.cluster.datastore.jmx.mbeans.shard.ShardStats;
  37 import org.opendaylight.controller.cluster.datastore.messages.AbortTransaction;
  38 import org.opendaylight.controller.cluster.datastore.messages.AbortTransactionReply;
  39 import org.opendaylight.controller.cluster.datastore.messages.ActorInitialized;
  40 import org.opendaylight.controller.cluster.datastore.messages.BatchedModifications;
  41 import org.opendaylight.controller.cluster.datastore.messages.CanCommitTransaction;
  42 import org.opendaylight.controller.cluster.datastore.messages.CloseTransactionChain;
  43 import org.opendaylight.controller.cluster.datastore.messages.CommitTransaction;
  44 import org.opendaylight.controller.cluster.datastore.messages.CommitTransactionReply;
  45 import org.opendaylight.controller.cluster.datastore.messages.CreateTransaction;
  46 import org.opendaylight.controller.cluster.datastore.messages.CreateTransactionReply;
  47 import org.opendaylight.controller.cluster.datastore.messages.ForwardedReadyTransaction;
  48 import org.opendaylight.controller.cluster.datastore.messages.PeerAddressResolved;
  49 import org.opendaylight.controller.cluster.datastore.messages.ReadyLocalTransaction;
  50 import org.opendaylight.controller.cluster.datastore.messages.RegisterChangeListener;
  51 import org.opendaylight.controller.cluster.datastore.messages.RegisterDataTreeChangeListener;
  52 import org.opendaylight.controller.cluster.datastore.messages.ShardLeaderStateChanged;
  53 import org.opendaylight.controller.cluster.datastore.messages.UpdateSchemaContext;
  54 import org.opendaylight.controller.cluster.datastore.modification.Modification;
  55 import org.opendaylight.controller.cluster.datastore.modification.ModificationPayload;
  56 import org.opendaylight.controller.cluster.datastore.modification.MutableCompositeModification;
  57 import org.opendaylight.controller.cluster.datastore.utils.Dispatchers;
  58 import org.opendaylight.controller.cluster.datastore.utils.MessageTracker;
  59 import org.opendaylight.controller.cluster.notifications.LeaderStateChanged;
  60 import org.opendaylight.controller.cluster.notifications.RegisterRoleChangeListener;
  61 import org.opendaylight.controller.cluster.notifications.RoleChangeNotifier;
  62 import org.opendaylight.controller.cluster.raft.RaftActor;
  63 import org.opendaylight.controller.cluster.raft.RaftActorRecoveryCohort;
  64 import org.opendaylight.controller.cluster.raft.RaftActorSnapshotCohort;
  65 import org.opendaylight.controller.cluster.raft.base.messages.FollowerInitialSyncUpStatus;
  66 import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
  67 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.CompositeModificationByteStringPayload;
  68 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.CompositeModificationPayload;
  69 import org.opendaylight.yangtools.yang.data.api.schema.tree.DataTree;
  70 import org.opendaylight.yangtools.yang.data.api.schema.tree.DataTreeCandidate;
  71 import org.opendaylight.yangtools.yang.data.api.schema.tree.DataValidationFailedException;
  72 import org.opendaylight.yangtools.yang.data.api.schema.tree.ModificationType;
  73 import org.opendaylight.yangtools.yang.model.api.SchemaContext;
  74 import scala.concurrent.duration.Duration;
  75 import scala.concurrent.duration.FiniteDuration;
  76
  77 /**
  78  * A Shard represents a portion of the logical data tree <br/>
  79  * <p>
  80  * Our Shard uses InMemoryDataTree as it's internal representation and delegates all requests it
  81  * </p>
  82  */
  83 public class Shard extends RaftActor {
  84
  85     private static final Object TX_COMMIT_TIMEOUT_CHECK_MESSAGE = "txCommitTimeoutCheck";
  86
  87     @VisibleForTesting
  88     static final Object GET_SHARD_MBEAN_MESSAGE = "getShardMBeanMessage";
  89
  90     @VisibleForTesting
  91     static final String DEFAULT_NAME = "default";
  92
  93     // The state of this Shard
  94     private final ShardDataTree store;
  95
  96     /// The name of this shard
  97     private final String name;
  98
  99     private final ShardStats shardMBean;
 100
 101     private DatastoreContext datastoreContext;
 102
 103     private final ShardCommitCoordinator commitCoordinator;
 104
 105     private long transactionCommitTimeout;
 106
 107     private Cancellable txCommitTimeoutCheckSchedule;
 108
 109     private final Optional<ActorRef> roleChangeNotifier;
 110
 111     private final MessageTracker appendEntriesReplyTracker;
 112
 113     private final ShardTransactionActorFactory transactionActorFactory;
 114
 115     private final ShardSnapshotCohort snapshotCohort;
 116
 117     private final DataTreeChangeListenerSupport treeChangeSupport = new DataTreeChangeListenerSupport(this);
 118     private final DataChangeListenerSupport changeSupport = new DataChangeListenerSupport(this);
 119
 120     protected Shard(final ShardIdentifier name, final Map<String, String> peerAddresses,
 121             final DatastoreContext datastoreContext, final SchemaContext schemaContext) {
 122         super(name.toString(), new HashMap<>(peerAddresses), Optional.of(datastoreContext.getShardRaftConfig()),
 123                 DataStoreVersions.CURRENT_VERSION);
 124
 125         this.name = name.toString();
 126         this.datastoreContext = datastoreContext;
 127
 128         setPersistence(datastoreContext.isPersistent());
 129
 130         LOG.info("Shard created : {}, persistent : {}", name, datastoreContext.isPersistent());
 131
 132         store = new ShardDataTree(schemaContext);
 133
 134         shardMBean = ShardMBeanFactory.getShardStatsMBean(name.toString(),
 135                 datastoreContext.getDataStoreMXBeanType());
 136         shardMBean.setShard(this);
 137
 138         if (isMetricsCaptureEnabled()) {
 139             getContext().become(new MeteringBehavior(this));
 140         }
 141
 142         commitCoordinator = new ShardCommitCoordinator(store,
 143                 datastoreContext.getShardCommitQueueExpiryTimeoutInMillis(),
 144                 datastoreContext.getShardTransactionCommitQueueCapacity(), self(), LOG, this.name);
 145
 146         setTransactionCommitTimeout();
 147
 148         // create a notifier actor for each cluster member
 149         roleChangeNotifier = createRoleChangeNotifier(name.toString());
 150
 151         appendEntriesReplyTracker = new MessageTracker(AppendEntriesReply.class,
 152                 getRaftActorContext().getConfigParams().getIsolatedCheckIntervalInMillis());
 153
 154         transactionActorFactory = new ShardTransactionActorFactory(store, datastoreContext,
 155                 new Dispatchers(context().system().dispatchers()).getDispatcherPath(
 156                         Dispatchers.DispatcherType.Transaction), self(), getContext(), shardMBean);
 157
 158         snapshotCohort = new ShardSnapshotCohort(transactionActorFactory, store, LOG, this.name);
 159
 160
 161     }
 162
 163     private void setTransactionCommitTimeout() {
 164         transactionCommitTimeout = TimeUnit.MILLISECONDS.convert(
 165                 datastoreContext.getShardTransactionCommitTimeoutInSeconds(), TimeUnit.SECONDS) / 2;
 166     }
 167
 168     public static Props props(final ShardIdentifier name,
 169         final Map<String, String> peerAddresses,
 170         final DatastoreContext datastoreContext, final SchemaContext schemaContext) {
 171         Preconditions.checkNotNull(name, "name should not be null");
 172         Preconditions.checkNotNull(peerAddresses, "peerAddresses should not be null");
 173         Preconditions.checkNotNull(datastoreContext, "dataStoreContext should not be null");
 174         Preconditions.checkNotNull(schemaContext, "schemaContext should not be null");
 175
 176         return Props.create(new ShardCreator(name, peerAddresses, datastoreContext, schemaContext));
 177     }
 178
 179     private Optional<ActorRef> createRoleChangeNotifier(String shardId) {
 180         ActorRef shardRoleChangeNotifier = this.getContext().actorOf(
 181             RoleChangeNotifier.getProps(shardId), shardId + "-notifier");
 182         return Optional.of(shardRoleChangeNotifier);
 183     }
 184
 185     @Override
 186     public void postStop() {
 187         LOG.info("Stopping Shard {}", persistenceId());
 188
 189         super.postStop();
 190
 191         if(txCommitTimeoutCheckSchedule != null) {
 192             txCommitTimeoutCheckSchedule.cancel();
 193         }
 194
 195         shardMBean.unregisterMBean();
 196     }
 197
 198     @Override
 199     public void onReceiveRecover(final Object message) throws Exception {
 200         if(LOG.isDebugEnabled()) {
 201             LOG.debug("{}: onReceiveRecover: Received message {} from {}", persistenceId(),
 202                 message.getClass().toString(), getSender());
 203         }
 204
 205         if (message instanceof RecoveryFailure){
 206             LOG.error("{}: Recovery failed because of this cause",
 207                     persistenceId(), ((RecoveryFailure) message).cause());
 208
 209             // Even though recovery failed, we still need to finish our recovery, eg send the
 210             // ActorInitialized message and start the txCommitTimeoutCheckSchedule.
 211             onRecoveryComplete();
 212         } else {
 213             super.onReceiveRecover(message);
 214             if(LOG.isTraceEnabled()) {
 215                 appendEntriesReplyTracker.begin();
 216             }
 217         }
 218     }
 219
 220     @Override
 221     public void onReceiveCommand(final Object message) throws Exception {
 222
 223         MessageTracker.Context context = appendEntriesReplyTracker.received(message);
 224
 225         if(context.error().isPresent()){
 226             LOG.trace("{} : AppendEntriesReply failed to arrive at the expected interval {}", persistenceId(),
 227                     context.error());
 228         }
 229
 230         try {
 231             if (CreateTransaction.SERIALIZABLE_CLASS.isInstance(message)) {
 232                 handleCreateTransaction(message);
 233             } else if (BatchedModifications.class.isInstance(message)) {
 234                 handleBatchedModifications((BatchedModifications)message);
 235             } else if (message instanceof ForwardedReadyTransaction) {
 236                 commitCoordinator.handleForwardedReadyTransaction((ForwardedReadyTransaction) message,
 237                         getSender(), this);
 238             } else if (message instanceof ReadyLocalTransaction) {
 239                 handleReadyLocalTransaction((ReadyLocalTransaction)message);
 240             } else if (CanCommitTransaction.SERIALIZABLE_CLASS.isInstance(message)) {
 241                 handleCanCommitTransaction(CanCommitTransaction.fromSerializable(message));
 242             } else if (CommitTransaction.SERIALIZABLE_CLASS.isInstance(message)) {
 243                 handleCommitTransaction(CommitTransaction.fromSerializable(message));
 244             } else if (AbortTransaction.SERIALIZABLE_CLASS.isInstance(message)) {
 245                 handleAbortTransaction(AbortTransaction.fromSerializable(message));
 246             } else if (CloseTransactionChain.SERIALIZABLE_CLASS.isInstance(message)) {
 247                 closeTransactionChain(CloseTransactionChain.fromSerializable(message));
 248             } else if (message instanceof RegisterChangeListener) {
 249                 changeSupport.onMessage((RegisterChangeListener) message, isLeader());
 250             } else if (message instanceof RegisterDataTreeChangeListener) {
 251                 treeChangeSupport.onMessage((RegisterDataTreeChangeListener) message, isLeader());
 252             } else if (message instanceof UpdateSchemaContext) {
 253                 updateSchemaContext((UpdateSchemaContext) message);
 254             } else if (message instanceof PeerAddressResolved) {
 255                 PeerAddressResolved resolved = (PeerAddressResolved) message;
 256                 setPeerAddress(resolved.getPeerId().toString(),
 257                         resolved.getPeerAddress());
 258             } else if (message.equals(TX_COMMIT_TIMEOUT_CHECK_MESSAGE)) {
 259                 handleTransactionCommitTimeoutCheck();
 260             } else if(message instanceof DatastoreContext) {
 261                 onDatastoreContext((DatastoreContext)message);
 262             } else if(message instanceof RegisterRoleChangeListener){
 263                 roleChangeNotifier.get().forward(message, context());
 264             } else if (message instanceof FollowerInitialSyncUpStatus) {
 265                 shardMBean.setFollowerInitialSyncStatus(((FollowerInitialSyncUpStatus) message).isInitialSyncDone());
 266                 context().parent().tell(message, self());
 267             } else if(GET_SHARD_MBEAN_MESSAGE.equals(message)){
 268                 sender().tell(getShardMBean(), self());
 269             } else {
 270                 super.onReceiveCommand(message);
 271             }
 272         } finally {
 273             context.done();
 274         }
 275     }
 276
 277     public int getPendingTxCommitQueueSize() {
 278         return commitCoordinator.getQueueSize();
 279     }
 280
 281     @Override
 282     protected Optional<ActorRef> getRoleChangeNotifier() {
 283         return roleChangeNotifier;
 284     }
 285
 286     @Override
 287     protected LeaderStateChanged newLeaderStateChanged(String memberId, String leaderId, short leaderPayloadVersion) {
 288         return new ShardLeaderStateChanged(memberId, leaderId,
 289                 isLeader() ? Optional.<DataTree>of(store.getDataTree()) : Optional.<DataTree>absent(),
 290                 leaderPayloadVersion);
 291     }
 292
 293     private void onDatastoreContext(DatastoreContext context) {
 294         datastoreContext = context;
 295
 296         commitCoordinator.setQueueCapacity(datastoreContext.getShardTransactionCommitQueueCapacity());
 297
 298         setTransactionCommitTimeout();
 299
 300         if(datastoreContext.isPersistent() && !persistence().isRecoveryApplicable()) {
 301             setPersistence(true);
 302         } else if(!datastoreContext.isPersistent() && persistence().isRecoveryApplicable()) {
 303             setPersistence(false);
 304         }
 305
 306         updateConfigParams(datastoreContext.getShardRaftConfig());
 307     }
 308
 309     private void handleTransactionCommitTimeoutCheck() {
 310         CohortEntry cohortEntry = commitCoordinator.getCurrentCohortEntry();
 311         if(cohortEntry != null) {
 312             if(cohortEntry.isExpired(transactionCommitTimeout)) {
 313                 LOG.warn("{}: Current transaction {} has timed out after {} ms - aborting",
 314                         persistenceId(), cohortEntry.getTransactionID(), transactionCommitTimeout);
 315
 316                 doAbortTransaction(cohortEntry.getTransactionID(), null);
 317             }
 318         }
 319
 320         commitCoordinator.cleanupExpiredCohortEntries();
 321     }
 322
 323     private static boolean isEmptyCommit(final DataTreeCandidate candidate) {
 324         return ModificationType.UNMODIFIED.equals(candidate.getRootNode().getModificationType());
 325     }
 326
 327     void continueCommit(final CohortEntry cohortEntry) throws Exception {
 328         final DataTreeCandidate candidate = cohortEntry.getCohort().getCandidate();
 329
 330         // If we do not have any followers and we are not using persistence
 331         // or if cohortEntry has no modifications
 332         // we can apply modification to the state immediately
 333         if ((!hasFollowers() && !persistence().isRecoveryApplicable()) || isEmptyCommit(candidate)) {
 334             applyModificationToState(cohortEntry.getReplySender(), cohortEntry.getTransactionID(), candidate);
 335         } else {
 336             Shard.this.persistData(cohortEntry.getReplySender(), cohortEntry.getTransactionID(),
 337                 DataTreeCandidatePayload.create(candidate));
 338         }
 339     }
 340
 341     private void handleCommitTransaction(final CommitTransaction commit) {
 342         if(!commitCoordinator.handleCommit(commit.getTransactionID(), getSender(), this)) {
 343             shardMBean.incrementFailedTransactionsCount();
 344         }
 345     }
 346
 347     private void finishCommit(@Nonnull final ActorRef sender, @Nonnull final String transactionID, @Nonnull final CohortEntry cohortEntry) {
 348         LOG.debug("{}: Finishing commit for transaction {}", persistenceId(), cohortEntry.getTransactionID());
 349
 350         try {
 351             // We block on the future here so we don't have to worry about possibly accessing our
 352             // state on a different thread outside of our dispatcher. Also, the data store
 353             // currently uses a same thread executor anyway.
 354             cohortEntry.getCohort().commit().get();
 355
 356             sender.tell(CommitTransactionReply.INSTANCE.toSerializable(), getSelf());
 357
 358             shardMBean.incrementCommittedTransactionCount();
 359             shardMBean.setLastCommittedTransactionTime(System.currentTimeMillis());
 360
 361         } catch (Exception e) {
 362             sender.tell(new akka.actor.Status.Failure(e), getSelf());
 363
 364             LOG.error("{}, An exception occurred while committing transaction {}", persistenceId(),
 365                     transactionID, e);
 366             shardMBean.incrementFailedTransactionsCount();
 367         } finally {
 368             commitCoordinator.currentTransactionComplete(transactionID, true);
 369         }
 370     }
 371
 372     private void finishCommit(@Nonnull final ActorRef sender, final @Nonnull String transactionID) {
 373         // With persistence enabled, this method is called via applyState by the leader strategy
 374         // after the commit has been replicated to a majority of the followers.
 375
 376         CohortEntry cohortEntry = commitCoordinator.getCohortEntryIfCurrent(transactionID);
 377         if (cohortEntry == null) {
 378             // The transaction is no longer the current commit. This can happen if the transaction
 379             // was aborted prior, most likely due to timeout in the front-end. We need to finish
 380             // committing the transaction though since it was successfully persisted and replicated
 381             // however we can't use the original cohort b/c it was already preCommitted and may
 382             // conflict with the current commit or may have been aborted so we commit with a new
 383             // transaction.
 384             cohortEntry = commitCoordinator.getAndRemoveCohortEntry(transactionID);
 385             if(cohortEntry != null) {
 386                 try {
 387                     store.applyForeignCandidate(transactionID, cohortEntry.getCohort().getCandidate());
 388                 } catch (DataValidationFailedException e) {
 389                     shardMBean.incrementFailedTransactionsCount();
 390                     LOG.error("{}: Failed to re-apply transaction {}", persistenceId(), transactionID, e);
 391                 }
 392
 393                 sender.tell(CommitTransactionReply.INSTANCE.toSerializable(), getSelf());
 394             } else {
 395                 // This really shouldn't happen - it likely means that persistence or replication
 396                 // took so long to complete such that the cohort entry was expired from the cache.
 397                 IllegalStateException ex = new IllegalStateException(
 398                         String.format("%s: Could not finish committing transaction %s - no CohortEntry found",
 399                                 persistenceId(), transactionID));
 400                 LOG.error(ex.getMessage());
 401                 sender.tell(new akka.actor.Status.Failure(ex), getSelf());
 402             }
 403         } else {
 404             finishCommit(sender, transactionID, cohortEntry);
 405         }
 406     }
 407
 408     private void handleCanCommitTransaction(final CanCommitTransaction canCommit) {
 409         LOG.debug("{}: Can committing transaction {}", persistenceId(), canCommit.getTransactionID());
 410         commitCoordinator.handleCanCommit(canCommit.getTransactionID(), getSender(), this);
 411     }
 412
 413     private void noLeaderError(Object message) {
 414         // TODO: rather than throwing an immediate exception, we could schedule a timer to try again to make
 415         // it more resilient in case we're in the process of electing a new leader.
 416         getSender().tell(new akka.actor.Status.Failure(new NoShardLeaderException(String.format(
 417             "Could not find the leader for shard %s. This typically happens" +
 418             " when the system is coming up or recovering and a leader is being elected. Try again" +
 419             " later.", persistenceId()))), getSelf());
 420     }
 421
 422     private void handleBatchedModifications(BatchedModifications batched) {
 423         // This message is sent to prepare the modifications transaction directly on the Shard as an
 424         // optimization to avoid the extra overhead of a separate ShardTransaction actor. On the last
 425         // BatchedModifications message, the caller sets the ready flag in the message indicating
 426         // modifications are complete. The reply contains the cohort actor path (this actor) for the caller
 427         // to initiate the 3-phase commit. This also avoids the overhead of sending an additional
 428         // ReadyTransaction message.
 429
 430         // If we're not the leader then forward to the leader. This is a safety measure - we shouldn't
 431         // normally get here if we're not the leader as the front-end (TransactionProxy) should determine
 432         // the primary/leader shard. However with timing and caching on the front-end, there's a small
 433         // window where it could have a stale leader during leadership transitions.
 434         //
 435         if(isLeader()) {
 436             try {
 437                 commitCoordinator.handleBatchedModifications(batched, getSender(), this);
 438             } catch (Exception e) {
 439                 LOG.error("{}: Error handling BatchedModifications for Tx {}", persistenceId(),
 440                         batched.getTransactionID(), e);
 441                 getSender().tell(new akka.actor.Status.Failure(e), getSelf());
 442             }
 443         } else {
 444             ActorSelection leader = getLeader();
 445             if(leader != null) {
 446                 // TODO: what if this is not the first batch and leadership changed in between batched messages?
 447                 // We could check if the commitCoordinator already has a cached entry and forward all the previous
 448                 // batched modifications.
 449                 LOG.debug("{}: Forwarding BatchedModifications to leader {}", persistenceId(), leader);
 450                 leader.forward(batched, getContext());
 451             } else {
 452                 noLeaderError(batched);
 453             }
 454         }
 455     }
 456
 457     private void handleReadyLocalTransaction(final ReadyLocalTransaction message) {
 458         if (isLeader()) {
 459             try {
 460                 commitCoordinator.handleReadyLocalTransaction(message, getSender(), this);
 461             } catch (Exception e) {
 462                 LOG.error("{}: Error handling ReadyLocalTransaction for Tx {}", persistenceId(),
 463                         message.getTransactionID(), e);
 464                 getSender().tell(new akka.actor.Status.Failure(e), getSelf());
 465             }
 466         } else {
 467             ActorSelection leader = getLeader();
 468             if (leader != null) {
 469                 LOG.debug("{}: Forwarding ReadyLocalTransaction to leader {}", persistenceId(), leader);
 470                 message.setRemoteVersion(getCurrentBehavior().getLeaderPayloadVersion());
 471                 leader.forward(message, getContext());
 472             } else {
 473                 noLeaderError(message);
 474             }
 475         }
 476     }
 477
 478     private void handleAbortTransaction(final AbortTransaction abort) {
 479         doAbortTransaction(abort.getTransactionID(), getSender());
 480     }
 481
 482     void doAbortTransaction(final String transactionID, final ActorRef sender) {
 483         final CohortEntry cohortEntry = commitCoordinator.getCohortEntryIfCurrent(transactionID);
 484         if(cohortEntry != null) {
 485             LOG.debug("{}: Aborting transaction {}", persistenceId(), transactionID);
 486
 487             // We don't remove the cached cohort entry here (ie pass false) in case the Tx was
 488             // aborted during replication in which case we may still commit locally if replication
 489             // succeeds.
 490             commitCoordinator.currentTransactionComplete(transactionID, false);
 491
 492             final ListenableFuture<Void> future = cohortEntry.getCohort().abort();
 493             final ActorRef self = getSelf();
 494
 495             Futures.addCallback(future, new FutureCallback<Void>() {
 496                 @Override
 497                 public void onSuccess(final Void v) {
 498                     shardMBean.incrementAbortTransactionsCount();
 499
 500                     if(sender != null) {
 501                         sender.tell(AbortTransactionReply.INSTANCE.toSerializable(), self);
 502                     }
 503                 }
 504
 505                 @Override
 506                 public void onFailure(final Throwable t) {
 507                     LOG.error("{}: An exception happened during abort", persistenceId(), t);
 508
 509                     if(sender != null) {
 510                         sender.tell(new akka.actor.Status.Failure(t), self);
 511                     }
 512                 }
 513             });
 514         }
 515     }
 516
 517     private void handleCreateTransaction(final Object message) {
 518         if (isLeader()) {
 519             createTransaction(CreateTransaction.fromSerializable(message));
 520         } else if (getLeader() != null) {
 521             getLeader().forward(message, getContext());
 522         } else {
 523             getSender().tell(new akka.actor.Status.Failure(new NoShardLeaderException(String.format(
 524                 "Could not find leader for shard %s so transaction cannot be created. This typically happens" +
 525                 " when the system is coming up or recovering and a leader is being elected. Try again" +
 526                 " later.", persistenceId()))), getSelf());
 527         }
 528     }
 529
 530     private void closeTransactionChain(final CloseTransactionChain closeTransactionChain) {
 531         store.closeTransactionChain(closeTransactionChain.getTransactionChainId());
 532     }
 533
 534     private ActorRef createTypedTransactionActor(int transactionType,
 535             ShardTransactionIdentifier transactionId, String transactionChainId,
 536             short clientVersion ) {
 537
 538         return transactionActorFactory.newShardTransaction(TransactionType.fromInt(transactionType),
 539                 transactionId, transactionChainId, clientVersion);
 540     }
 541
 542     private void createTransaction(CreateTransaction createTransaction) {
 543         try {
 544             ActorRef transactionActor = createTransaction(createTransaction.getTransactionType(),
 545                 createTransaction.getTransactionId(), createTransaction.getTransactionChainId(),
 546                 createTransaction.getVersion());
 547
 548             getSender().tell(new CreateTransactionReply(Serialization.serializedActorPath(transactionActor),
 549                     createTransaction.getTransactionId()).toSerializable(), getSelf());
 550         } catch (Exception e) {
 551             getSender().tell(new akka.actor.Status.Failure(e), getSelf());
 552         }
 553     }
 554
 555     private ActorRef createTransaction(int transactionType, String remoteTransactionId,
 556             String transactionChainId, short clientVersion) {
 557
 558
 559         ShardTransactionIdentifier transactionId = new ShardTransactionIdentifier(remoteTransactionId);
 560
 561         if(LOG.isDebugEnabled()) {
 562             LOG.debug("{}: Creating transaction : {} ", persistenceId(), transactionId);
 563         }
 564
 565         ActorRef transactionActor = createTypedTransactionActor(transactionType, transactionId,
 566                 transactionChainId, clientVersion);
 567
 568         return transactionActor;
 569     }
 570
 571     private void commitWithNewTransaction(final Modification modification) {
 572         ReadWriteShardDataTreeTransaction tx = store.newReadWriteTransaction(modification.toString(), null);
 573         modification.apply(tx.getSnapshot());
 574         try {
 575             snapshotCohort.syncCommitTransaction(tx);
 576             shardMBean.incrementCommittedTransactionCount();
 577             shardMBean.setLastCommittedTransactionTime(System.currentTimeMillis());
 578         } catch (Exception e) {
 579             shardMBean.incrementFailedTransactionsCount();
 580             LOG.error("{}: Failed to commit", persistenceId(), e);
 581         }
 582     }
 583
 584     private void updateSchemaContext(final UpdateSchemaContext message) {
 585         updateSchemaContext(message.getSchemaContext());
 586     }
 587
 588     @VisibleForTesting
 589     void updateSchemaContext(final SchemaContext schemaContext) {
 590         store.updateSchemaContext(schemaContext);
 591     }
 592
 593     private boolean isMetricsCaptureEnabled() {
 594         CommonConfig config = new CommonConfig(getContext().system().settings().config());
 595         return config.isMetricCaptureEnabled();
 596     }
 597
 598     @Override
 599     @VisibleForTesting
 600     public RaftActorSnapshotCohort getRaftActorSnapshotCohort() {
 601         return snapshotCohort;
 602     }
 603
 604     @Override
 605     @Nonnull
 606     protected RaftActorRecoveryCohort getRaftActorRecoveryCohort() {
 607         return new ShardRecoveryCoordinator(store, store.getSchemaContext(), persistenceId(), LOG);
 608     }
 609
 610     @Override
 611     protected void onRecoveryComplete() {
 612         //notify shard manager
 613         getContext().parent().tell(new ActorInitialized(), getSelf());
 614
 615         // Being paranoid here - this method should only be called once but just in case...
 616         if(txCommitTimeoutCheckSchedule == null) {
 617             // Schedule a message to be periodically sent to check if the current in-progress
 618             // transaction should be expired and aborted.
 619             FiniteDuration period = Duration.create(transactionCommitTimeout / 3, TimeUnit.MILLISECONDS);
 620             txCommitTimeoutCheckSchedule = getContext().system().scheduler().schedule(
 621                     period, period, getSelf(),
 622                     TX_COMMIT_TIMEOUT_CHECK_MESSAGE, getContext().dispatcher(), ActorRef.noSender());
 623         }
 624     }
 625
 626     @Override
 627     protected void applyState(final ActorRef clientActor, final String identifier, final Object data) {
 628         if (data instanceof DataTreeCandidatePayload) {
 629             if (clientActor == null) {
 630                 // No clientActor indicates a replica coming from the leader
 631                 try {
 632                     store.applyForeignCandidate(identifier, ((DataTreeCandidatePayload)data).getCandidate());
 633                 } catch (DataValidationFailedException | IOException e) {
 634                     LOG.error("{}: Error applying replica {}", persistenceId(), identifier, e);
 635                 }
 636             } else {
 637                 // Replication consensus reached, proceed to commit
 638                 finishCommit(clientActor, identifier);
 639             }
 640         } else if (data instanceof ModificationPayload) {
 641             try {
 642                 applyModificationToState(clientActor, identifier, ((ModificationPayload) data).getModification());
 643             } catch (ClassNotFoundException | IOException e) {
 644                 LOG.error("{}: Error extracting ModificationPayload", persistenceId(), e);
 645             }
 646         } else if (data instanceof CompositeModificationPayload) {
 647             Object modification = ((CompositeModificationPayload) data).getModification();
 648
 649             applyModificationToState(clientActor, identifier, modification);
 650         } else if(data instanceof CompositeModificationByteStringPayload ){
 651             Object modification = ((CompositeModificationByteStringPayload) data).getModification();
 652
 653             applyModificationToState(clientActor, identifier, modification);
 654         } else {
 655             LOG.error("{}: Unknown state received {} Class loader = {} CompositeNodeMod.ClassLoader = {}",
 656                     persistenceId(), data, data.getClass().getClassLoader(),
 657                     CompositeModificationPayload.class.getClassLoader());
 658         }
 659     }
 660
 661     private void applyModificationToState(ActorRef clientActor, String identifier, Object modification) {
 662         if(modification == null) {
 663             LOG.error(
 664                     "{}: modification is null - this is very unexpected, clientActor = {}, identifier = {}",
 665                     persistenceId(), identifier, clientActor != null ? clientActor.path().toString() : null);
 666         } else if(clientActor == null) {
 667             // There's no clientActor to which to send a commit reply so we must be applying
 668             // replicated state from the leader.
 669             commitWithNewTransaction(MutableCompositeModification.fromSerializable(modification));
 670         } else {
 671             // This must be the OK to commit after replication consensus.
 672             finishCommit(clientActor, identifier);
 673         }
 674     }
 675
 676     @Override
 677     protected void onStateChanged() {
 678         boolean isLeader = isLeader();
 679         changeSupport.onLeadershipChange(isLeader);
 680         treeChangeSupport.onLeadershipChange(isLeader);
 681
 682         // If this actor is no longer the leader close all the transaction chains
 683         if (!isLeader) {
 684             if(LOG.isDebugEnabled()) {
 685                 LOG.debug(
 686                     "{}: onStateChanged: Closing all transaction chains because shard {} is no longer the leader",
 687                     persistenceId(), getId());
 688             }
 689
 690             store.closeAllTransactionChains();
 691         }
 692     }
 693
 694     @Override
 695     protected void onLeaderChanged(String oldLeader, String newLeader) {
 696         shardMBean.incrementLeadershipChangeCount();
 697     }
 698
 699     @Override
 700     public String persistenceId() {
 701         return this.name;
 702     }
 703
 704     @VisibleForTesting
 705     ShardCommitCoordinator getCommitCoordinator() {
 706         return commitCoordinator;
 707     }
 708
 709
 710     private static class ShardCreator implements Creator<Shard> {
 711
 712         private static final long serialVersionUID = 1L;
 713
 714         final ShardIdentifier name;
 715         final Map<String, String> peerAddresses;
 716         final DatastoreContext datastoreContext;
 717         final SchemaContext schemaContext;
 718
 719         ShardCreator(final ShardIdentifier name, final Map<String, String> peerAddresses,
 720                 final DatastoreContext datastoreContext, final SchemaContext schemaContext) {
 721             this.name = name;
 722             this.peerAddresses = peerAddresses;
 723             this.datastoreContext = datastoreContext;
 724             this.schemaContext = schemaContext;
 725         }
 726
 727         @Override
 728         public Shard create() throws Exception {
 729             return new Shard(name, peerAddresses, datastoreContext, schemaContext);
 730         }
 731     }
 732
 733     @VisibleForTesting
 734     public ShardDataTree getDataStore() {
 735         return store;
 736     }
 737
 738     @VisibleForTesting
 739     ShardStats getShardMBean() {
 740         return shardMBean;
 741     }
 742 }