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 com.google.common.base.Throwables;
  20 import java.io.IOException;
  21 import java.util.Collection;
  22 import java.util.Collections;
  23 import java.util.Map;
  24 import java.util.concurrent.ExecutionException;
  25 import java.util.concurrent.TimeUnit;
  26 import javax.annotation.Nonnull;
  27 import org.opendaylight.controller.cluster.access.concepts.TransactionIdentifier;
  28 import org.opendaylight.controller.cluster.common.actor.CommonConfig;
  29 import org.opendaylight.controller.cluster.common.actor.MessageTracker;
  30 import org.opendaylight.controller.cluster.common.actor.MessageTracker.Error;
  31 import org.opendaylight.controller.cluster.common.actor.MeteringBehavior;
  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.jmx.mbeans.shard.ShardMBeanFactory;
  35 import org.opendaylight.controller.cluster.datastore.jmx.mbeans.shard.ShardStats;
  36 import org.opendaylight.controller.cluster.datastore.messages.AbortTransaction;
  37 import org.opendaylight.controller.cluster.datastore.messages.ActorInitialized;
  38 import org.opendaylight.controller.cluster.datastore.messages.BatchedModifications;
  39 import org.opendaylight.controller.cluster.datastore.messages.CanCommitTransaction;
  40 import org.opendaylight.controller.cluster.datastore.messages.CloseTransactionChain;
  41 import org.opendaylight.controller.cluster.datastore.messages.CommitTransaction;
  42 import org.opendaylight.controller.cluster.datastore.messages.CommitTransactionReply;
  43 import org.opendaylight.controller.cluster.datastore.messages.CreateTransaction;
  44 import org.opendaylight.controller.cluster.datastore.messages.CreateTransactionReply;
  45 import org.opendaylight.controller.cluster.datastore.messages.DatastoreSnapshot;
  46 import org.opendaylight.controller.cluster.datastore.messages.DatastoreSnapshot.ShardSnapshot;
  47 import org.opendaylight.controller.cluster.datastore.messages.ForwardedReadyTransaction;
  48 import org.opendaylight.controller.cluster.datastore.messages.GetShardDataTree;
  49 import org.opendaylight.controller.cluster.datastore.messages.PeerAddressResolved;
  50 import org.opendaylight.controller.cluster.datastore.messages.ReadyLocalTransaction;
  51 import org.opendaylight.controller.cluster.datastore.messages.RegisterChangeListener;
  52 import org.opendaylight.controller.cluster.datastore.messages.RegisterDataTreeChangeListener;
  53 import org.opendaylight.controller.cluster.datastore.messages.ShardLeaderStateChanged;
  54 import org.opendaylight.controller.cluster.datastore.messages.UpdateSchemaContext;
  55 import org.opendaylight.controller.cluster.datastore.persisted.CommitTransactionPayload;
  56 import org.opendaylight.controller.cluster.datastore.persisted.DataTreeCandidateSupplier;
  57 import org.opendaylight.controller.cluster.datastore.utils.Dispatchers;
  58 import org.opendaylight.controller.cluster.notifications.LeaderStateChanged;
  59 import org.opendaylight.controller.cluster.notifications.RegisterRoleChangeListener;
  60 import org.opendaylight.controller.cluster.notifications.RoleChangeNotifier;
  61 import org.opendaylight.controller.cluster.raft.RaftActor;
  62 import org.opendaylight.controller.cluster.raft.RaftActorRecoveryCohort;
  63 import org.opendaylight.controller.cluster.raft.RaftActorSnapshotCohort;
  64 import org.opendaylight.controller.cluster.raft.RaftState;
  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.messages.ServerRemoved;
  68 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload;
  69 import org.opendaylight.yangtools.concepts.Identifier;
  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.data.api.schema.tree.TreeType;
  74 import org.opendaylight.yangtools.yang.model.api.SchemaContext;
  75 import scala.concurrent.duration.Duration;
  76 import scala.concurrent.duration.FiniteDuration;
  77
  78 /**
  79  * A Shard represents a portion of the logical data tree <br/>
  80  * <p>
  81  * Our Shard uses InMemoryDataTree as it's internal representation and delegates all requests it
  82  * </p>
  83  */
  84 public class Shard extends RaftActor {
  85
  86     @VisibleForTesting
  87     static final Object TX_COMMIT_TIMEOUT_CHECK_MESSAGE = new Object() {
  88         @Override
  89         public String toString() {
  90             return "txCommitTimeoutCheck";
  91         }
  92     };
  93
  94     @VisibleForTesting
  95     static final Object GET_SHARD_MBEAN_MESSAGE = new Object() {
  96         @Override
  97         public String toString() {
  98             return "getShardMBeanMessage";
  99         }
 100     };
 101
 102     // FIXME: shard names should be encapsulated in their own class and this should be exposed as a constant.
 103     public static final String DEFAULT_NAME = "default";
 104
 105     // The state of this Shard
 106     private final ShardDataTree store;
 107
 108     /// The name of this shard
 109     private final String name;
 110
 111     private final ShardStats shardMBean;
 112
 113     private DatastoreContext datastoreContext;
 114
 115     private final ShardCommitCoordinator commitCoordinator;
 116
 117     private long transactionCommitTimeout;
 118
 119     private Cancellable txCommitTimeoutCheckSchedule;
 120
 121     private final Optional<ActorRef> roleChangeNotifier;
 122
 123     private final MessageTracker appendEntriesReplyTracker;
 124
 125     private final ShardTransactionActorFactory transactionActorFactory;
 126
 127     private final ShardSnapshotCohort snapshotCohort;
 128
 129     private final DataTreeChangeListenerSupport treeChangeSupport = new DataTreeChangeListenerSupport(this);
 130     private final DataChangeListenerSupport changeSupport = new DataChangeListenerSupport(this);
 131
 132
 133     private ShardSnapshot restoreFromSnapshot;
 134
 135     private final ShardTransactionMessageRetrySupport messageRetrySupport;
 136
 137     protected Shard(AbstractBuilder<?, ?> builder) {
 138         super(builder.getId().toString(), builder.getPeerAddresses(),
 139                 Optional.of(builder.getDatastoreContext().getShardRaftConfig()), DataStoreVersions.CURRENT_VERSION);
 140
 141         this.name = builder.getId().toString();
 142         this.datastoreContext = builder.getDatastoreContext();
 143         this.restoreFromSnapshot = builder.getRestoreFromSnapshot();
 144
 145         setPersistence(datastoreContext.isPersistent());
 146
 147         LOG.info("Shard created : {}, persistent : {}", name, datastoreContext.isPersistent());
 148
 149         store = new ShardDataTree(builder.getSchemaContext(), builder.getTreeType(),
 150                 new ShardDataTreeChangeListenerPublisherActorProxy(getContext(), name + "-DTCL-publisher"),
 151                 new ShardDataChangeListenerPublisherActorProxy(getContext(), name + "-DCL-publisher"), name);
 152
 153         shardMBean = ShardMBeanFactory.getShardStatsMBean(name.toString(),
 154                 datastoreContext.getDataStoreMXBeanType());
 155         shardMBean.setShard(this);
 156
 157         if (isMetricsCaptureEnabled()) {
 158             getContext().become(new MeteringBehavior(this));
 159         }
 160
 161         commitCoordinator = new ShardCommitCoordinator(store,
 162                 datastoreContext.getShardCommitQueueExpiryTimeoutInMillis(),
 163                 datastoreContext.getShardTransactionCommitQueueCapacity(), LOG, this.name);
 164
 165         setTransactionCommitTimeout();
 166
 167         // create a notifier actor for each cluster member
 168         roleChangeNotifier = createRoleChangeNotifier(name.toString());
 169
 170         appendEntriesReplyTracker = new MessageTracker(AppendEntriesReply.class,
 171                 getRaftActorContext().getConfigParams().getIsolatedCheckIntervalInMillis());
 172
 173         transactionActorFactory = new ShardTransactionActorFactory(store, datastoreContext,
 174                 new Dispatchers(context().system().dispatchers()).getDispatcherPath(
 175                         Dispatchers.DispatcherType.Transaction), self(), getContext(), shardMBean);
 176
 177         snapshotCohort = ShardSnapshotCohort.create(getContext(), builder.getId().getMemberName(), store, LOG,
 178             this.name);
 179
 180         messageRetrySupport = new ShardTransactionMessageRetrySupport(this);
 181     }
 182
 183     private void setTransactionCommitTimeout() {
 184         transactionCommitTimeout = TimeUnit.MILLISECONDS.convert(
 185                 datastoreContext.getShardTransactionCommitTimeoutInSeconds(), TimeUnit.SECONDS) / 2;
 186     }
 187
 188     private Optional<ActorRef> createRoleChangeNotifier(String shardId) {
 189         ActorRef shardRoleChangeNotifier = this.getContext().actorOf(
 190             RoleChangeNotifier.getProps(shardId), shardId + "-notifier");
 191         return Optional.of(shardRoleChangeNotifier);
 192     }
 193
 194     @Override
 195     public void postStop() {
 196         LOG.info("Stopping Shard {}", persistenceId());
 197
 198         super.postStop();
 199
 200         messageRetrySupport.close();
 201
 202         if(txCommitTimeoutCheckSchedule != null) {
 203             txCommitTimeoutCheckSchedule.cancel();
 204         }
 205
 206         commitCoordinator.abortPendingTransactions("Transaction aborted due to shutdown.", this);
 207
 208         shardMBean.unregisterMBean();
 209     }
 210
 211     @Override
 212     protected void handleRecover(final Object message) {
 213         LOG.debug("{}: onReceiveRecover: Received message {} from {}", persistenceId(), message.getClass(),
 214             getSender());
 215
 216         super.handleRecover(message);
 217         if (LOG.isTraceEnabled()) {
 218             appendEntriesReplyTracker.begin();
 219         }
 220     }
 221
 222     @Override
 223     protected void handleNonRaftCommand(final Object message) {
 224         try (final MessageTracker.Context context = appendEntriesReplyTracker.received(message)) {
 225             final Optional<Error> maybeError = context.error();
 226             if (maybeError.isPresent()) {
 227                 LOG.trace("{} : AppendEntriesReply failed to arrive at the expected interval {}", persistenceId(),
 228                     maybeError.get());
 229             }
 230
 231             if (CreateTransaction.isSerializedType(message)) {
 232                 handleCreateTransaction(message);
 233             } else if (message instanceof BatchedModifications) {
 234                 handleBatchedModifications((BatchedModifications)message);
 235             } else if (message instanceof ForwardedReadyTransaction) {
 236                 handleForwardedReadyTransaction((ForwardedReadyTransaction) message);
 237             } else if (message instanceof ReadyLocalTransaction) {
 238                 handleReadyLocalTransaction((ReadyLocalTransaction)message);
 239             } else if (CanCommitTransaction.isSerializedType(message)) {
 240                 handleCanCommitTransaction(CanCommitTransaction.fromSerializable(message));
 241             } else if (CommitTransaction.isSerializedType(message)) {
 242                 handleCommitTransaction(CommitTransaction.fromSerializable(message));
 243             } else if (AbortTransaction.isSerializedType(message)) {
 244                 handleAbortTransaction(AbortTransaction.fromSerializable(message));
 245             } else if (CloseTransactionChain.isSerializedType(message)) {
 246                 closeTransactionChain(CloseTransactionChain.fromSerializable(message));
 247             } else if (message instanceof RegisterChangeListener) {
 248                 changeSupport.onMessage((RegisterChangeListener) message, isLeader(), hasLeader());
 249             } else if (message instanceof RegisterDataTreeChangeListener) {
 250                 treeChangeSupport.onMessage((RegisterDataTreeChangeListener) message, isLeader(), hasLeader());
 251             } else if (message instanceof UpdateSchemaContext) {
 252                 updateSchemaContext((UpdateSchemaContext) message);
 253             } else if (message instanceof PeerAddressResolved) {
 254                 PeerAddressResolved resolved = (PeerAddressResolved) message;
 255                 setPeerAddress(resolved.getPeerId().toString(),
 256                         resolved.getPeerAddress());
 257             } else if (TX_COMMIT_TIMEOUT_CHECK_MESSAGE.equals(message)) {
 258                 commitCoordinator.checkForExpiredTransactions(transactionCommitTimeout, this);
 259             } else if(message instanceof DatastoreContext) {
 260                 onDatastoreContext((DatastoreContext)message);
 261             } else if(message instanceof RegisterRoleChangeListener){
 262                 roleChangeNotifier.get().forward(message, context());
 263             } else if (message instanceof FollowerInitialSyncUpStatus) {
 264                 shardMBean.setFollowerInitialSyncStatus(((FollowerInitialSyncUpStatus) message).isInitialSyncDone());
 265                 context().parent().tell(message, self());
 266             } else if(GET_SHARD_MBEAN_MESSAGE.equals(message)){
 267                 sender().tell(getShardMBean(), self());
 268             } else if(message instanceof GetShardDataTree) {
 269                 sender().tell(store.getDataTree(), self());
 270             } else if(message instanceof ServerRemoved){
 271                 context().parent().forward(message, context());
 272             } else if(ShardTransactionMessageRetrySupport.TIMER_MESSAGE_CLASS.isInstance(message)) {
 273                 messageRetrySupport.onTimerMessage(message);
 274             } else if (message instanceof DataTreeCohortActorRegistry.CohortRegistryCommand) {
 275                 commitCoordinator.processCohortRegistryCommand(getSender(),
 276                         (DataTreeCohortActorRegistry.CohortRegistryCommand) message);
 277             } else {
 278                 super.handleNonRaftCommand(message);
 279             }
 280         }
 281     }
 282
 283     private boolean hasLeader() {
 284         return getLeaderId() != null;
 285     }
 286
 287     public int getPendingTxCommitQueueSize() {
 288         return commitCoordinator.getQueueSize();
 289     }
 290
 291     public int getCohortCacheSize() {
 292         return commitCoordinator.getCohortCacheSize();
 293     }
 294
 295     @Override
 296     protected Optional<ActorRef> getRoleChangeNotifier() {
 297         return roleChangeNotifier;
 298     }
 299
 300     @Override
 301     protected LeaderStateChanged newLeaderStateChanged(String memberId, String leaderId, short leaderPayloadVersion) {
 302         return isLeader() ? new ShardLeaderStateChanged(memberId, leaderId, store.getDataTree(), leaderPayloadVersion)
 303                 : new ShardLeaderStateChanged(memberId, leaderId, leaderPayloadVersion);
 304     }
 305
 306     protected void onDatastoreContext(DatastoreContext context) {
 307         datastoreContext = context;
 308
 309         commitCoordinator.setQueueCapacity(datastoreContext.getShardTransactionCommitQueueCapacity());
 310
 311         setTransactionCommitTimeout();
 312
 313         if(datastoreContext.isPersistent() && !persistence().isRecoveryApplicable()) {
 314             setPersistence(true);
 315         } else if(!datastoreContext.isPersistent() && persistence().isRecoveryApplicable()) {
 316             setPersistence(false);
 317         }
 318
 319         updateConfigParams(datastoreContext.getShardRaftConfig());
 320     }
 321
 322     private static boolean isEmptyCommit(final DataTreeCandidate candidate) {
 323         return ModificationType.UNMODIFIED.equals(candidate.getRootNode().getModificationType());
 324     }
 325
 326     void continueCommit(final CohortEntry cohortEntry) {
 327         final DataTreeCandidate candidate = cohortEntry.getCandidate();
 328         final TransactionIdentifier transactionId = cohortEntry.getTransactionID();
 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(), transactionId, candidate);
 335             return;
 336         }
 337
 338         final Payload payload;
 339         try {
 340             payload = CommitTransactionPayload.create(transactionId, candidate);
 341         } catch (IOException e) {
 342             LOG.error("{}: failed to encode transaction {} candidate {}", persistenceId(), transactionId, candidate,
 343                 e);
 344             // TODO: do we need to do something smarter here?
 345             throw Throwables.propagate(e);
 346         }
 347
 348         persistData(cohortEntry.getReplySender(), cohortEntry.getTransactionID(), payload);
 349     }
 350
 351     private void handleCommitTransaction(final CommitTransaction commit) {
 352         if (isLeader()) {
 353             if(!commitCoordinator.handleCommit(commit.getTransactionID(), getSender(), this)) {
 354                 shardMBean.incrementFailedTransactionsCount();
 355             }
 356         } else {
 357             ActorSelection leader = getLeader();
 358             if (leader == null) {
 359                 messageRetrySupport.addMessageToRetry(commit, getSender(),
 360                         "Could not commit transaction " + commit.getTransactionID());
 361             } else {
 362                 LOG.debug("{}: Forwarding CommitTransaction to leader {}", persistenceId(), leader);
 363                 leader.forward(commit, getContext());
 364             }
 365         }
 366     }
 367
 368     private void finishCommit(@Nonnull final ActorRef sender, @Nonnull final Identifier transactionID,
 369             @Nonnull final CohortEntry cohortEntry) {
 370         LOG.debug("{}: Finishing commit for transaction {}", persistenceId(), cohortEntry.getTransactionID());
 371
 372         try {
 373             try {
 374             cohortEntry.commit();
 375             } catch(ExecutionException e) {
 376                 // We may get a "store tree and candidate base differ" IllegalStateException from commit under
 377                 // certain edge case scenarios so we'll try to re-apply the candidate from scratch as a last
 378                 // resort. Eg, we're a follower and a tx payload is replicated but the leader goes down before
 379                 // applying it to the state. We then become the leader and a second tx is pre-committed and
 380                 // replicated. When consensus occurs, this will cause the first tx to be applied as a foreign
 381                 // candidate via applyState prior to the second tx. Since the second tx has already been
 382                 // pre-committed, when it gets here to commit it will get an IllegalStateException.
 383
 384                 // FIXME - this is not an ideal way to handle this scenario. This is temporary - a cleaner
 385                 // solution will be forthcoming.
 386                 if(e.getCause() instanceof IllegalStateException) {
 387                     LOG.debug("{}: commit failed for transaction {} - retrying as foreign candidate", persistenceId(),
 388                             transactionID, e);
 389                     store.applyForeignCandidate(transactionID, cohortEntry.getCandidate());
 390                 } else {
 391                     throw e;
 392                 }
 393             }
 394
 395             sender.tell(CommitTransactionReply.instance(cohortEntry.getClientVersion()).toSerializable(), getSelf());
 396
 397             shardMBean.incrementCommittedTransactionCount();
 398             shardMBean.setLastCommittedTransactionTime(System.currentTimeMillis());
 399
 400         } catch (Exception e) {
 401             sender.tell(new akka.actor.Status.Failure(e), getSelf());
 402
 403             LOG.error("{}, An exception occurred while committing transaction {}", persistenceId(),
 404                     transactionID, e);
 405             shardMBean.incrementFailedTransactionsCount();
 406         } finally {
 407             commitCoordinator.currentTransactionComplete(transactionID, true);
 408         }
 409     }
 410
 411     private void finishCommit(@Nonnull final ActorRef sender, final @Nonnull Identifier transactionID) {
 412         // With persistence enabled, this method is called via applyState by the leader strategy
 413         // after the commit has been replicated to a majority of the followers.
 414
 415         CohortEntry cohortEntry = commitCoordinator.getCohortEntryIfCurrent(transactionID);
 416         if (cohortEntry == null) {
 417             // The transaction is no longer the current commit. This can happen if the transaction
 418             // was aborted prior, most likely due to timeout in the front-end. We need to finish
 419             // committing the transaction though since it was successfully persisted and replicated
 420             // however we can't use the original cohort b/c it was already preCommitted and may
 421             // conflict with the current commit or may have been aborted so we commit with a new
 422             // transaction.
 423             cohortEntry = commitCoordinator.getAndRemoveCohortEntry(transactionID);
 424             if(cohortEntry != null) {
 425                 try {
 426                     store.applyForeignCandidate(transactionID, cohortEntry.getCandidate());
 427                 } catch (DataValidationFailedException e) {
 428                     shardMBean.incrementFailedTransactionsCount();
 429                     LOG.error("{}: Failed to re-apply transaction {}", persistenceId(), transactionID, e);
 430                 }
 431
 432                 sender.tell(CommitTransactionReply.instance(cohortEntry.getClientVersion()).toSerializable(),
 433                         getSelf());
 434             } else {
 435                 // This really shouldn't happen - it likely means that persistence or replication
 436                 // took so long to complete such that the cohort entry was expired from the cache.
 437                 IllegalStateException ex = new IllegalStateException(
 438                         String.format("%s: Could not finish committing transaction %s - no CohortEntry found",
 439                                 persistenceId(), transactionID));
 440                 LOG.error(ex.getMessage());
 441                 sender.tell(new akka.actor.Status.Failure(ex), getSelf());
 442             }
 443         } else {
 444             finishCommit(sender, transactionID, cohortEntry);
 445         }
 446     }
 447
 448     private void handleCanCommitTransaction(final CanCommitTransaction canCommit) {
 449         LOG.debug("{}: Can committing transaction {}", persistenceId(), canCommit.getTransactionID());
 450
 451         if (isLeader()) {
 452         commitCoordinator.handleCanCommit(canCommit.getTransactionID(), getSender(), this);
 453         } else {
 454             ActorSelection leader = getLeader();
 455             if (leader == null) {
 456                 messageRetrySupport.addMessageToRetry(canCommit, getSender(),
 457                         "Could not canCommit transaction " + canCommit.getTransactionID());
 458             } else {
 459                 LOG.debug("{}: Forwarding CanCommitTransaction to leader {}", persistenceId(), leader);
 460                 leader.forward(canCommit, getContext());
 461             }
 462         }
 463     }
 464
 465     protected void handleBatchedModificationsLocal(BatchedModifications batched, ActorRef sender) {
 466         try {
 467             commitCoordinator.handleBatchedModifications(batched, sender, this);
 468         } catch (Exception e) {
 469             LOG.error("{}: Error handling BatchedModifications for Tx {}", persistenceId(),
 470                     batched.getTransactionID(), e);
 471             sender.tell(new akka.actor.Status.Failure(e), getSelf());
 472         }
 473     }
 474
 475     private void handleBatchedModifications(BatchedModifications batched) {
 476         // This message is sent to prepare the modifications transaction directly on the Shard as an
 477         // optimization to avoid the extra overhead of a separate ShardTransaction actor. On the last
 478         // BatchedModifications message, the caller sets the ready flag in the message indicating
 479         // modifications are complete. The reply contains the cohort actor path (this actor) for the caller
 480         // to initiate the 3-phase commit. This also avoids the overhead of sending an additional
 481         // ReadyTransaction message.
 482
 483         // If we're not the leader then forward to the leader. This is a safety measure - we shouldn't
 484         // normally get here if we're not the leader as the front-end (TransactionProxy) should determine
 485         // the primary/leader shard. However with timing and caching on the front-end, there's a small
 486         // window where it could have a stale leader during leadership transitions.
 487         //
 488         boolean isLeaderActive = isLeaderActive();
 489         if (isLeader() && isLeaderActive) {
 490             handleBatchedModificationsLocal(batched, getSender());
 491         } else {
 492             ActorSelection leader = getLeader();
 493             if (!isLeaderActive || leader == null) {
 494                 messageRetrySupport.addMessageToRetry(batched, getSender(),
 495                         "Could not commit transaction " + batched.getTransactionID());
 496             } else {
 497                 // If this is not the first batch and leadership changed in between batched messages,
 498                 // we need to reconstruct previous BatchedModifications from the transaction
 499                 // DataTreeModification, honoring the max batched modification count, and forward all the
 500                 // previous BatchedModifications to the new leader.
 501                 Collection<BatchedModifications> newModifications = commitCoordinator.createForwardedBatchedModifications(
 502                         batched, datastoreContext.getShardBatchedModificationCount());
 503
 504                 LOG.debug("{}: Forwarding {} BatchedModifications to leader {}", persistenceId(),
 505                         newModifications.size(), leader);
 506
 507                 for(BatchedModifications bm: newModifications) {
 508                     leader.forward(bm, getContext());
 509                 }
 510             }
 511         }
 512     }
 513
 514     private boolean failIfIsolatedLeader(ActorRef sender) {
 515         if(isIsolatedLeader()) {
 516             sender.tell(new akka.actor.Status.Failure(new NoShardLeaderException(String.format(
 517                     "Shard %s was the leader but has lost contact with all of its followers. Either all" +
 518                     " other follower nodes are down or this node is isolated by a network partition.",
 519                     persistenceId()))), getSelf());
 520             return true;
 521         }
 522
 523         return false;
 524     }
 525
 526     protected boolean isIsolatedLeader() {
 527         return getRaftState() == RaftState.IsolatedLeader;
 528     }
 529
 530     private void handleReadyLocalTransaction(final ReadyLocalTransaction message) {
 531         LOG.debug("{}: handleReadyLocalTransaction for {}", persistenceId(), message.getTransactionID());
 532
 533         boolean isLeaderActive = isLeaderActive();
 534         if (isLeader() && isLeaderActive) {
 535             try {
 536                 commitCoordinator.handleReadyLocalTransaction(message, getSender(), this);
 537             } catch (Exception e) {
 538                 LOG.error("{}: Error handling ReadyLocalTransaction for Tx {}", persistenceId(),
 539                         message.getTransactionID(), e);
 540                 getSender().tell(new akka.actor.Status.Failure(e), getSelf());
 541             }
 542         } else {
 543             ActorSelection leader = getLeader();
 544             if (!isLeaderActive || leader == null) {
 545                 messageRetrySupport.addMessageToRetry(message, getSender(),
 546                         "Could not commit transaction " + message.getTransactionID());
 547             } else {
 548                 LOG.debug("{}: Forwarding ReadyLocalTransaction to leader {}", persistenceId(), leader);
 549                 message.setRemoteVersion(getCurrentBehavior().getLeaderPayloadVersion());
 550                 leader.forward(message, getContext());
 551             }
 552         }
 553     }
 554
 555     private void handleForwardedReadyTransaction(ForwardedReadyTransaction forwardedReady) {
 556         LOG.debug("{}: handleForwardedReadyTransaction for {}", persistenceId(), forwardedReady.getTransactionID());
 557
 558         boolean isLeaderActive = isLeaderActive();
 559         if (isLeader() && isLeaderActive) {
 560             commitCoordinator.handleForwardedReadyTransaction(forwardedReady, getSender(), this,
 561                     store.getSchemaContext());
 562         } else {
 563             ActorSelection leader = getLeader();
 564             if (!isLeaderActive || leader == null) {
 565                 messageRetrySupport.addMessageToRetry(forwardedReady, getSender(),
 566                         "Could not commit transaction " + forwardedReady.getTransactionID());
 567             } else {
 568                 LOG.debug("{}: Forwarding ForwardedReadyTransaction to leader {}", persistenceId(), leader);
 569
 570                 ReadyLocalTransaction readyLocal = new ReadyLocalTransaction(forwardedReady.getTransactionID(),
 571                         forwardedReady.getTransaction().getSnapshot(), forwardedReady.isDoImmediateCommit());
 572                 readyLocal.setRemoteVersion(getCurrentBehavior().getLeaderPayloadVersion());
 573                 leader.forward(readyLocal, getContext());
 574             }
 575         }
 576     }
 577
 578     private void handleAbortTransaction(final AbortTransaction abort) {
 579         doAbortTransaction(abort.getTransactionID(), getSender());
 580     }
 581
 582     void doAbortTransaction(final TransactionIdentifier transactionID, final ActorRef sender) {
 583         commitCoordinator.handleAbort(transactionID, sender, this);
 584     }
 585
 586     private void handleCreateTransaction(final Object message) {
 587         if (isLeader()) {
 588             createTransaction(CreateTransaction.fromSerializable(message));
 589         } else if (getLeader() != null) {
 590             getLeader().forward(message, getContext());
 591         } else {
 592             getSender().tell(new akka.actor.Status.Failure(new NoShardLeaderException(
 593                     "Could not create a shard transaction", persistenceId())), getSelf());
 594         }
 595     }
 596
 597     private void closeTransactionChain(final CloseTransactionChain closeTransactionChain) {
 598         store.closeTransactionChain(closeTransactionChain.getIdentifier());
 599     }
 600
 601     private void createTransaction(CreateTransaction createTransaction) {
 602         try {
 603             if(TransactionType.fromInt(createTransaction.getTransactionType()) != TransactionType.READ_ONLY &&
 604                     failIfIsolatedLeader(getSender())) {
 605                 return;
 606             }
 607
 608             ActorRef transactionActor = createTransaction(createTransaction.getTransactionType(),
 609                 createTransaction.getTransactionId());
 610
 611             getSender().tell(new CreateTransactionReply(Serialization.serializedActorPath(transactionActor),
 612                     createTransaction.getTransactionId(), createTransaction.getVersion()).toSerializable(), getSelf());
 613         } catch (Exception e) {
 614             getSender().tell(new akka.actor.Status.Failure(e), getSelf());
 615         }
 616     }
 617
 618     private ActorRef createTransaction(int transactionType, TransactionIdentifier transactionId) {
 619         LOG.debug("{}: Creating transaction : {} ", persistenceId(), transactionId);
 620         return transactionActorFactory.newShardTransaction(TransactionType.fromInt(transactionType),
 621             transactionId);
 622     }
 623
 624     private void commitWithNewTransaction(final BatchedModifications modification) {
 625         ReadWriteShardDataTreeTransaction tx = store.newReadWriteTransaction(modification.getTransactionID());
 626         modification.apply(tx.getSnapshot());
 627         try {
 628             snapshotCohort.syncCommitTransaction(tx);
 629             shardMBean.incrementCommittedTransactionCount();
 630             shardMBean.setLastCommittedTransactionTime(System.currentTimeMillis());
 631         } catch (Exception e) {
 632             shardMBean.incrementFailedTransactionsCount();
 633             LOG.error("{}: Failed to commit", persistenceId(), e);
 634         }
 635     }
 636
 637     private void updateSchemaContext(final UpdateSchemaContext message) {
 638         updateSchemaContext(message.getSchemaContext());
 639     }
 640
 641     @VisibleForTesting
 642     void updateSchemaContext(final SchemaContext schemaContext) {
 643         store.updateSchemaContext(schemaContext);
 644     }
 645
 646     private boolean isMetricsCaptureEnabled() {
 647         CommonConfig config = new CommonConfig(getContext().system().settings().config());
 648         return config.isMetricCaptureEnabled();
 649     }
 650
 651     @Override
 652     @VisibleForTesting
 653     public RaftActorSnapshotCohort getRaftActorSnapshotCohort() {
 654         return snapshotCohort;
 655     }
 656
 657     @Override
 658     @Nonnull
 659     protected RaftActorRecoveryCohort getRaftActorRecoveryCohort() {
 660         return new ShardRecoveryCoordinator(store,
 661             restoreFromSnapshot != null ? restoreFromSnapshot.getSnapshot() : null, persistenceId(), LOG);
 662     }
 663
 664     @Override
 665     protected void onRecoveryComplete() {
 666         restoreFromSnapshot = null;
 667
 668         //notify shard manager
 669         getContext().parent().tell(new ActorInitialized(), getSelf());
 670
 671         // Being paranoid here - this method should only be called once but just in case...
 672         if(txCommitTimeoutCheckSchedule == null) {
 673             // Schedule a message to be periodically sent to check if the current in-progress
 674             // transaction should be expired and aborted.
 675             FiniteDuration period = Duration.create(transactionCommitTimeout / 3, TimeUnit.MILLISECONDS);
 676             txCommitTimeoutCheckSchedule = getContext().system().scheduler().schedule(
 677                     period, period, getSelf(),
 678                     TX_COMMIT_TIMEOUT_CHECK_MESSAGE, getContext().dispatcher(), ActorRef.noSender());
 679         }
 680     }
 681
 682     @Override
 683     protected void applyState(final ActorRef clientActor, final Identifier identifier, final Object data) {
 684         if (data instanceof DataTreeCandidateSupplier) {
 685             if (clientActor == null) {
 686                 // No clientActor indicates a replica coming from the leader
 687                 try {
 688                     store.applyForeignCandidate(identifier, ((DataTreeCandidateSupplier)data).getCandidate().getValue());
 689                 } catch (DataValidationFailedException | IOException e) {
 690                     LOG.error("{}: Error applying replica {}", persistenceId(), identifier, e);
 691                 }
 692             } else {
 693                 // Replication consensus reached, proceed to commit
 694                 finishCommit(clientActor, identifier);
 695             }
 696         } else {
 697             LOG.error("{}: Unknown state received {} ClassLoader {}", persistenceId(), data,
 698                 data.getClass().getClassLoader());
 699         }
 700     }
 701
 702     private void applyModificationToState(ActorRef clientActor, Identifier identifier, Object modification) {
 703         if(modification == null) {
 704             LOG.error(
 705                     "{}: modification is null - this is very unexpected, clientActor = {}, identifier = {}",
 706                     persistenceId(), identifier, clientActor != null ? clientActor.path().toString() : null);
 707         } else if(clientActor == null) {
 708             // There's no clientActor to which to send a commit reply so we must be applying
 709             // replicated state from the leader.
 710
 711             // The only implementation we know of is BatchedModifications, which also carries a transaction
 712             // identifier -- which we really need that.
 713             Preconditions.checkArgument(modification instanceof BatchedModifications);
 714             commitWithNewTransaction((BatchedModifications)modification);
 715         } else {
 716             // This must be the OK to commit after replication consensus.
 717             finishCommit(clientActor, identifier);
 718         }
 719     }
 720
 721     @Override
 722     protected void onStateChanged() {
 723         boolean isLeader = isLeader();
 724         boolean hasLeader = hasLeader();
 725         changeSupport.onLeadershipChange(isLeader, hasLeader);
 726         treeChangeSupport.onLeadershipChange(isLeader, hasLeader);
 727
 728         // If this actor is no longer the leader close all the transaction chains
 729         if (!isLeader) {
 730             if(LOG.isDebugEnabled()) {
 731                 LOG.debug(
 732                     "{}: onStateChanged: Closing all transaction chains because shard {} is no longer the leader",
 733                     persistenceId(), getId());
 734             }
 735
 736             store.closeAllTransactionChains();
 737         }
 738
 739         if(hasLeader && !isIsolatedLeader()) {
 740             messageRetrySupport.retryMessages();
 741         }
 742     }
 743
 744     @Override
 745     protected void onLeaderChanged(String oldLeader, String newLeader) {
 746         shardMBean.incrementLeadershipChangeCount();
 747
 748         boolean hasLeader = hasLeader();
 749         if(hasLeader && !isLeader()) {
 750             // Another leader was elected. If we were the previous leader and had pending transactions, convert
 751             // them to transaction messages and send to the new leader.
 752             ActorSelection leader = getLeader();
 753             if(leader != null) {
 754                 Collection<Object> messagesToForward = commitCoordinator.convertPendingTransactionsToMessages(
 755                         datastoreContext.getShardBatchedModificationCount());
 756
 757                 if(!messagesToForward.isEmpty()) {
 758                     LOG.debug("{}: Forwarding {} pending transaction messages to leader {}", persistenceId(),
 759                             messagesToForward.size(), leader);
 760
 761                     for(Object message: messagesToForward) {
 762                         leader.tell(message, self());
 763                     }
 764                 }
 765             } else {
 766                 commitCoordinator.abortPendingTransactions(
 767                         "The transacton was aborted due to inflight leadership change and the leader address isn't available.",
 768                         this);
 769             }
 770         }
 771
 772         if(hasLeader && !isIsolatedLeader()) {
 773             messageRetrySupport.retryMessages();
 774         }
 775     }
 776
 777     @Override
 778     protected void pauseLeader(Runnable operation) {
 779         LOG.debug("{}: In pauseLeader, operation: {}", persistenceId(), operation);
 780         commitCoordinator.setRunOnPendingTransactionsComplete(operation);
 781     }
 782
 783     @Override
 784     public String persistenceId() {
 785         return this.name;
 786     }
 787
 788     @VisibleForTesting
 789     ShardCommitCoordinator getCommitCoordinator() {
 790         return commitCoordinator;
 791     }
 792
 793     public DatastoreContext getDatastoreContext() {
 794         return datastoreContext;
 795     }
 796
 797     @VisibleForTesting
 798     public ShardDataTree getDataStore() {
 799         return store;
 800     }
 801
 802     @VisibleForTesting
 803     ShardStats getShardMBean() {
 804         return shardMBean;
 805     }
 806
 807     public static Builder builder() {
 808         return new Builder();
 809     }
 810
 811     public static abstract class AbstractBuilder<T extends AbstractBuilder<T, S>, S extends Shard> {
 812         private final Class<S> shardClass;
 813         private ShardIdentifier id;
 814         private Map<String, String> peerAddresses = Collections.emptyMap();
 815         private DatastoreContext datastoreContext;
 816         private SchemaContext schemaContext;
 817         private DatastoreSnapshot.ShardSnapshot restoreFromSnapshot;
 818         private volatile boolean sealed;
 819
 820         protected AbstractBuilder(Class<S> shardClass) {
 821             this.shardClass = shardClass;
 822         }
 823
 824         protected void checkSealed() {
 825             Preconditions.checkState(!sealed, "Builder isalready sealed - further modifications are not allowed");
 826         }
 827
 828         @SuppressWarnings("unchecked")
 829         private T self() {
 830             return (T) this;
 831         }
 832
 833         public T id(ShardIdentifier id) {
 834             checkSealed();
 835             this.id = id;
 836             return self();
 837         }
 838
 839         public T peerAddresses(Map<String, String> peerAddresses) {
 840             checkSealed();
 841             this.peerAddresses = peerAddresses;
 842             return self();
 843         }
 844
 845         public T datastoreContext(DatastoreContext datastoreContext) {
 846             checkSealed();
 847             this.datastoreContext = datastoreContext;
 848             return self();
 849         }
 850
 851         public T schemaContext(SchemaContext schemaContext) {
 852             checkSealed();
 853             this.schemaContext = schemaContext;
 854             return self();
 855         }
 856
 857         public T restoreFromSnapshot(DatastoreSnapshot.ShardSnapshot restoreFromSnapshot) {
 858             checkSealed();
 859             this.restoreFromSnapshot = restoreFromSnapshot;
 860             return self();
 861         }
 862
 863         public ShardIdentifier getId() {
 864             return id;
 865         }
 866
 867         public Map<String, String> getPeerAddresses() {
 868             return peerAddresses;
 869         }
 870
 871         public DatastoreContext getDatastoreContext() {
 872             return datastoreContext;
 873         }
 874
 875         public SchemaContext getSchemaContext() {
 876             return schemaContext;
 877         }
 878
 879         public DatastoreSnapshot.ShardSnapshot getRestoreFromSnapshot() {
 880             return restoreFromSnapshot;
 881         }
 882
 883         public TreeType getTreeType() {
 884             switch (datastoreContext.getLogicalStoreType()) {
 885             case CONFIGURATION:
 886                 return TreeType.CONFIGURATION;
 887             case OPERATIONAL:
 888                 return TreeType.OPERATIONAL;
 889             }
 890
 891             throw new IllegalStateException("Unhandled logical store type " + datastoreContext.getLogicalStoreType());
 892         }
 893
 894         protected void verify() {
 895             Preconditions.checkNotNull(id, "id should not be null");
 896             Preconditions.checkNotNull(peerAddresses, "peerAddresses should not be null");
 897             Preconditions.checkNotNull(datastoreContext, "dataStoreContext should not be null");
 898             Preconditions.checkNotNull(schemaContext, "schemaContext should not be null");
 899         }
 900
 901         public Props props() {
 902             sealed = true;
 903             verify();
 904             return Props.create(shardClass, this);
 905         }
 906     }
 907
 908     public static class Builder extends AbstractBuilder<Builder, Shard> {
 909         private Builder() {
 910             super(Shard.class);
 911         }
 912     }
 913 }