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.ShardCommitCoordinator.CohortEntry;
  31 import org.opendaylight.controller.cluster.datastore.exceptions.NoShardLeaderException;
  32 import org.opendaylight.controller.cluster.datastore.identifiers.ShardIdentifier;
  33 import org.opendaylight.controller.cluster.datastore.identifiers.ShardTransactionIdentifier;
  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.modification.Modification;
  56 import org.opendaylight.controller.cluster.datastore.modification.MutableCompositeModification;
  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.CompositeModificationByteStringPayload;
  69 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.CompositeModificationPayload;
  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     protected static final Object TX_COMMIT_TIMEOUT_CHECK_MESSAGE = "txCommitTimeoutCheck";
  87
  88     @VisibleForTesting
  89     static final Object GET_SHARD_MBEAN_MESSAGE = "getShardMBeanMessage";
  90
  91     // FIXME: shard names should be encapsulated in their own class and this should be exposed as a constant.
  92     public static final String DEFAULT_NAME = "default";
  93
  94     // The state of this Shard
  95     private final ShardDataTree store;
  96
  97     /// The name of this shard
  98     private final String name;
  99
 100     private final ShardStats shardMBean;
 101
 102     private DatastoreContext datastoreContext;
 103
 104     private final ShardCommitCoordinator commitCoordinator;
 105
 106     private long transactionCommitTimeout;
 107
 108     private Cancellable txCommitTimeoutCheckSchedule;
 109
 110     private final Optional<ActorRef> roleChangeNotifier;
 111
 112     private final MessageTracker appendEntriesReplyTracker;
 113
 114     private final ShardTransactionActorFactory transactionActorFactory;
 115
 116     private final ShardSnapshotCohort snapshotCohort;
 117
 118     private final DataTreeChangeListenerSupport treeChangeSupport = new DataTreeChangeListenerSupport(this);
 119     private final DataChangeListenerSupport changeSupport = new DataChangeListenerSupport(this);
 120
 121
 122     private ShardSnapshot restoreFromSnapshot;
 123
 124     private final ShardTransactionMessageRetrySupport messageRetrySupport;
 125
 126     protected Shard(AbstractBuilder<?, ?> builder) {
 127         super(builder.getId().toString(), builder.getPeerAddresses(),
 128                 Optional.of(builder.getDatastoreContext().getShardRaftConfig()), DataStoreVersions.CURRENT_VERSION);
 129
 130         this.name = builder.getId().toString();
 131         this.datastoreContext = builder.getDatastoreContext();
 132         this.restoreFromSnapshot = builder.getRestoreFromSnapshot();
 133
 134         setPersistence(datastoreContext.isPersistent());
 135
 136         LOG.info("Shard created : {}, persistent : {}", name, datastoreContext.isPersistent());
 137
 138         store = new ShardDataTree(builder.getSchemaContext(), builder.getTreeType(),
 139                 new ShardDataTreeChangeListenerPublisherActorProxy(getContext(), name + "-DTCL-publisher"),
 140                 new ShardDataChangeListenerPublisherActorProxy(getContext(), name + "-DCL-publisher"), name);
 141
 142         shardMBean = ShardMBeanFactory.getShardStatsMBean(name.toString(),
 143                 datastoreContext.getDataStoreMXBeanType());
 144         shardMBean.setShard(this);
 145
 146         if (isMetricsCaptureEnabled()) {
 147             getContext().become(new MeteringBehavior(this));
 148         }
 149
 150         commitCoordinator = new ShardCommitCoordinator(store,
 151                 datastoreContext.getShardCommitQueueExpiryTimeoutInMillis(),
 152                 datastoreContext.getShardTransactionCommitQueueCapacity(), LOG, this.name);
 153
 154         setTransactionCommitTimeout();
 155
 156         // create a notifier actor for each cluster member
 157         roleChangeNotifier = createRoleChangeNotifier(name.toString());
 158
 159         appendEntriesReplyTracker = new MessageTracker(AppendEntriesReply.class,
 160                 getRaftActorContext().getConfigParams().getIsolatedCheckIntervalInMillis());
 161
 162         transactionActorFactory = new ShardTransactionActorFactory(store, datastoreContext,
 163                 new Dispatchers(context().system().dispatchers()).getDispatcherPath(
 164                         Dispatchers.DispatcherType.Transaction), self(), getContext(), shardMBean);
 165
 166         snapshotCohort = new ShardSnapshotCohort(transactionActorFactory, store, LOG, this.name);
 167
 168         messageRetrySupport = new ShardTransactionMessageRetrySupport(this);
 169     }
 170
 171     private void setTransactionCommitTimeout() {
 172         transactionCommitTimeout = TimeUnit.MILLISECONDS.convert(
 173                 datastoreContext.getShardTransactionCommitTimeoutInSeconds(), TimeUnit.SECONDS) / 2;
 174     }
 175
 176     private Optional<ActorRef> createRoleChangeNotifier(String shardId) {
 177         ActorRef shardRoleChangeNotifier = this.getContext().actorOf(
 178             RoleChangeNotifier.getProps(shardId), shardId + "-notifier");
 179         return Optional.of(shardRoleChangeNotifier);
 180     }
 181
 182     @Override
 183     public void postStop() {
 184         LOG.info("Stopping Shard {}", persistenceId());
 185
 186         super.postStop();
 187
 188         messageRetrySupport.close();
 189
 190         if(txCommitTimeoutCheckSchedule != null) {
 191             txCommitTimeoutCheckSchedule.cancel();
 192         }
 193
 194         commitCoordinator.abortPendingTransactions("Transaction aborted due to shutdown.", this);
 195
 196         shardMBean.unregisterMBean();
 197     }
 198
 199     @Override
 200     protected void handleRecover(final Object message) {
 201         LOG.debug("{}: onReceiveRecover: Received message {} from {}", persistenceId(), message.getClass(),
 202             getSender());
 203
 204         super.handleRecover(message);
 205         if (LOG.isTraceEnabled()) {
 206             appendEntriesReplyTracker.begin();
 207         }
 208     }
 209
 210     @Override
 211     protected void handleCommand(final Object message) {
 212
 213         final MessageTracker.Context context = appendEntriesReplyTracker.received(message);
 214         final Optional<Error> maybeError = context.error();
 215         if (maybeError.isPresent()) {
 216             LOG.trace("{} : AppendEntriesReply failed to arrive at the expected interval {}", persistenceId(),
 217                 maybeError.get());
 218         }
 219
 220         try {
 221             if (CreateTransaction.isSerializedType(message)) {
 222                 handleCreateTransaction(message);
 223             } else if (message instanceof BatchedModifications) {
 224                 handleBatchedModifications((BatchedModifications)message);
 225             } else if (message instanceof ForwardedReadyTransaction) {
 226                 handleForwardedReadyTransaction((ForwardedReadyTransaction) message);
 227             } else if (message instanceof ReadyLocalTransaction) {
 228                 handleReadyLocalTransaction((ReadyLocalTransaction)message);
 229             } else if (CanCommitTransaction.isSerializedType(message)) {
 230                 handleCanCommitTransaction(CanCommitTransaction.fromSerializable(message));
 231             } else if (CommitTransaction.isSerializedType(message)) {
 232                 handleCommitTransaction(CommitTransaction.fromSerializable(message));
 233             } else if (AbortTransaction.isSerializedType(message)) {
 234                 handleAbortTransaction(AbortTransaction.fromSerializable(message));
 235             } else if (CloseTransactionChain.isSerializedType(message)) {
 236                 closeTransactionChain(CloseTransactionChain.fromSerializable(message));
 237             } else if (message instanceof RegisterChangeListener) {
 238                 changeSupport.onMessage((RegisterChangeListener) message, isLeader(), hasLeader());
 239             } else if (message instanceof RegisterDataTreeChangeListener) {
 240                 treeChangeSupport.onMessage((RegisterDataTreeChangeListener) message, isLeader(), hasLeader());
 241             } else if (message instanceof UpdateSchemaContext) {
 242                 updateSchemaContext((UpdateSchemaContext) message);
 243             } else if (message instanceof PeerAddressResolved) {
 244                 PeerAddressResolved resolved = (PeerAddressResolved) message;
 245                 setPeerAddress(resolved.getPeerId().toString(),
 246                         resolved.getPeerAddress());
 247             } else if (TX_COMMIT_TIMEOUT_CHECK_MESSAGE.equals(message)) {
 248                 commitCoordinator.checkForExpiredTransactions(transactionCommitTimeout, this);
 249             } else if(message instanceof DatastoreContext) {
 250                 onDatastoreContext((DatastoreContext)message);
 251             } else if(message instanceof RegisterRoleChangeListener){
 252                 roleChangeNotifier.get().forward(message, context());
 253             } else if (message instanceof FollowerInitialSyncUpStatus) {
 254                 shardMBean.setFollowerInitialSyncStatus(((FollowerInitialSyncUpStatus) message).isInitialSyncDone());
 255                 context().parent().tell(message, self());
 256             } else if(GET_SHARD_MBEAN_MESSAGE.equals(message)){
 257                 sender().tell(getShardMBean(), self());
 258             } else if(message instanceof GetShardDataTree) {
 259                 sender().tell(store.getDataTree(), self());
 260             } else if(message instanceof ServerRemoved){
 261                 context().parent().forward(message, context());
 262             } else if(ShardTransactionMessageRetrySupport.TIMER_MESSAGE_CLASS.isInstance(message)) {
 263                 messageRetrySupport.onTimerMessage(message);
 264             } else {
 265                 super.handleCommand(message);
 266             }
 267         } finally {
 268             context.done();
 269         }
 270     }
 271
 272     private boolean hasLeader() {
 273         return getLeaderId() != null;
 274     }
 275
 276     public int getPendingTxCommitQueueSize() {
 277         return commitCoordinator.getQueueSize();
 278     }
 279
 280     public int getCohortCacheSize() {
 281         return commitCoordinator.getCohortCacheSize();
 282     }
 283
 284     @Override
 285     protected Optional<ActorRef> getRoleChangeNotifier() {
 286         return roleChangeNotifier;
 287     }
 288
 289     @Override
 290     protected LeaderStateChanged newLeaderStateChanged(String memberId, String leaderId, short leaderPayloadVersion) {
 291         return isLeader() ? new ShardLeaderStateChanged(memberId, leaderId, store.getDataTree(), leaderPayloadVersion)
 292                 : new ShardLeaderStateChanged(memberId, leaderId, leaderPayloadVersion);
 293     }
 294
 295     protected void onDatastoreContext(DatastoreContext context) {
 296         datastoreContext = context;
 297
 298         commitCoordinator.setQueueCapacity(datastoreContext.getShardTransactionCommitQueueCapacity());
 299
 300         setTransactionCommitTimeout();
 301
 302         if(datastoreContext.isPersistent() && !persistence().isRecoveryApplicable()) {
 303             setPersistence(true);
 304         } else if(!datastoreContext.isPersistent() && persistence().isRecoveryApplicable()) {
 305             setPersistence(false);
 306         }
 307
 308         updateConfigParams(datastoreContext.getShardRaftConfig());
 309     }
 310
 311     private static boolean isEmptyCommit(final DataTreeCandidate candidate) {
 312         return ModificationType.UNMODIFIED.equals(candidate.getRootNode().getModificationType());
 313     }
 314
 315     void continueCommit(final CohortEntry cohortEntry) {
 316         final DataTreeCandidate candidate = cohortEntry.getCandidate();
 317
 318         // If we do not have any followers and we are not using persistence
 319         // or if cohortEntry has no modifications
 320         // we can apply modification to the state immediately
 321         if ((!hasFollowers() && !persistence().isRecoveryApplicable()) || isEmptyCommit(candidate)) {
 322             applyModificationToState(cohortEntry.getReplySender(), cohortEntry.getTransactionID(), candidate);
 323         } else {
 324             Shard.this.persistData(cohortEntry.getReplySender(), cohortEntry.getTransactionID(),
 325                     DataTreeCandidatePayload.create(candidate));
 326         }
 327     }
 328
 329     private void handleCommitTransaction(final CommitTransaction commit) {
 330         if (isLeader()) {
 331             if(!commitCoordinator.handleCommit(commit.getTransactionID(), getSender(), this)) {
 332                 shardMBean.incrementFailedTransactionsCount();
 333             }
 334         } else {
 335             ActorSelection leader = getLeader();
 336             if (leader == null) {
 337                 messageRetrySupport.addMessageToRetry(commit, getSender(),
 338                         "Could not commit transaction " + commit.getTransactionID());
 339             } else {
 340                 LOG.debug("{}: Forwarding CommitTransaction to leader {}", persistenceId(), leader);
 341                 leader.forward(commit, getContext());
 342             }
 343         }
 344     }
 345
 346     private void finishCommit(@Nonnull final ActorRef sender, @Nonnull final String transactionID, @Nonnull final CohortEntry cohortEntry) {
 347         LOG.debug("{}: Finishing commit for transaction {}", persistenceId(), cohortEntry.getTransactionID());
 348
 349         try {
 350             try {
 351                 cohortEntry.commit();
 352             } catch(ExecutionException e) {
 353                 // We may get a "store tree and candidate base differ" IllegalStateException from commit under
 354                 // certain edge case scenarios so we'll try to re-apply the candidate from scratch as a last
 355                 // resort. Eg, we're a follower and a tx payload is replicated but the leader goes down before
 356                 // applying it to the state. We then become the leader and a second tx is pre-committed and
 357                 // replicated. When consensus occurs, this will cause the first tx to be applied as a foreign
 358                 // candidate via applyState prior to the second tx. Since the second tx has already been
 359                 // pre-committed, when it gets here to commit it will get an IllegalStateException.
 360
 361                 // FIXME - this is not an ideal way to handle this scenario. This is temporary - a cleaner
 362                 // solution will be forthcoming.
 363                 if(e.getCause() instanceof IllegalStateException) {
 364                     LOG.debug("{}: commit failed for transaction {} - retrying as foreign candidate", persistenceId(),
 365                             transactionID, e);
 366                     store.applyForeignCandidate(transactionID, cohortEntry.getCandidate());
 367                 } else {
 368                     throw e;
 369                 }
 370             }
 371
 372             sender.tell(CommitTransactionReply.instance(cohortEntry.getClientVersion()).toSerializable(), getSelf());
 373
 374             shardMBean.incrementCommittedTransactionCount();
 375             shardMBean.setLastCommittedTransactionTime(System.currentTimeMillis());
 376
 377         } catch (Exception e) {
 378             sender.tell(new akka.actor.Status.Failure(e), getSelf());
 379
 380             LOG.error("{}, An exception occurred while committing transaction {}", persistenceId(),
 381                     transactionID, e);
 382             shardMBean.incrementFailedTransactionsCount();
 383         } finally {
 384             commitCoordinator.currentTransactionComplete(transactionID, true);
 385         }
 386     }
 387
 388     private void finishCommit(@Nonnull final ActorRef sender, final @Nonnull String transactionID) {
 389         // With persistence enabled, this method is called via applyState by the leader strategy
 390         // after the commit has been replicated to a majority of the followers.
 391
 392         CohortEntry cohortEntry = commitCoordinator.getCohortEntryIfCurrent(transactionID);
 393         if (cohortEntry == null) {
 394             // The transaction is no longer the current commit. This can happen if the transaction
 395             // was aborted prior, most likely due to timeout in the front-end. We need to finish
 396             // committing the transaction though since it was successfully persisted and replicated
 397             // however we can't use the original cohort b/c it was already preCommitted and may
 398             // conflict with the current commit or may have been aborted so we commit with a new
 399             // transaction.
 400             cohortEntry = commitCoordinator.getAndRemoveCohortEntry(transactionID);
 401             if(cohortEntry != null) {
 402                 try {
 403                     store.applyForeignCandidate(transactionID, cohortEntry.getCandidate());
 404                 } catch (DataValidationFailedException e) {
 405                     shardMBean.incrementFailedTransactionsCount();
 406                     LOG.error("{}: Failed to re-apply transaction {}", persistenceId(), transactionID, e);
 407                 }
 408
 409                 sender.tell(CommitTransactionReply.instance(cohortEntry.getClientVersion()).toSerializable(),
 410                         getSelf());
 411             } else {
 412                 // This really shouldn't happen - it likely means that persistence or replication
 413                 // took so long to complete such that the cohort entry was expired from the cache.
 414                 IllegalStateException ex = new IllegalStateException(
 415                         String.format("%s: Could not finish committing transaction %s - no CohortEntry found",
 416                                 persistenceId(), transactionID));
 417                 LOG.error(ex.getMessage());
 418                 sender.tell(new akka.actor.Status.Failure(ex), getSelf());
 419             }
 420         } else {
 421             finishCommit(sender, transactionID, cohortEntry);
 422         }
 423     }
 424
 425     private void handleCanCommitTransaction(final CanCommitTransaction canCommit) {
 426         LOG.debug("{}: Can committing transaction {}", persistenceId(), canCommit.getTransactionID());
 427
 428         if (isLeader()) {
 429             commitCoordinator.handleCanCommit(canCommit.getTransactionID(), getSender(), this);
 430         } else {
 431             ActorSelection leader = getLeader();
 432             if (leader == null) {
 433                 messageRetrySupport.addMessageToRetry(canCommit, getSender(),
 434                         "Could not canCommit transaction " + canCommit.getTransactionID());
 435             } else {
 436                 LOG.debug("{}: Forwarding CanCommitTransaction to leader {}", persistenceId(), leader);
 437                 leader.forward(canCommit, getContext());
 438             }
 439         }
 440     }
 441
 442     protected void handleBatchedModificationsLocal(BatchedModifications batched, ActorRef sender) {
 443         try {
 444             commitCoordinator.handleBatchedModifications(batched, sender, this);
 445         } catch (Exception e) {
 446             LOG.error("{}: Error handling BatchedModifications for Tx {}", persistenceId(),
 447                     batched.getTransactionID(), e);
 448             sender.tell(new akka.actor.Status.Failure(e), getSelf());
 449         }
 450     }
 451
 452     private void handleBatchedModifications(BatchedModifications batched) {
 453         // This message is sent to prepare the modifications transaction directly on the Shard as an
 454         // optimization to avoid the extra overhead of a separate ShardTransaction actor. On the last
 455         // BatchedModifications message, the caller sets the ready flag in the message indicating
 456         // modifications are complete. The reply contains the cohort actor path (this actor) for the caller
 457         // to initiate the 3-phase commit. This also avoids the overhead of sending an additional
 458         // ReadyTransaction message.
 459
 460         // If we're not the leader then forward to the leader. This is a safety measure - we shouldn't
 461         // normally get here if we're not the leader as the front-end (TransactionProxy) should determine
 462         // the primary/leader shard. However with timing and caching on the front-end, there's a small
 463         // window where it could have a stale leader during leadership transitions.
 464         //
 465         boolean isLeaderActive = isLeaderActive();
 466         if (isLeader() && isLeaderActive) {
 467             handleBatchedModificationsLocal(batched, getSender());
 468         } else {
 469             ActorSelection leader = getLeader();
 470             if (!isLeaderActive || leader == null) {
 471                 messageRetrySupport.addMessageToRetry(batched, getSender(),
 472                         "Could not commit transaction " + batched.getTransactionID());
 473             } else {
 474                 // If this is not the first batch and leadership changed in between batched messages,
 475                 // we need to reconstruct previous BatchedModifications from the transaction
 476                 // DataTreeModification, honoring the max batched modification count, and forward all the
 477                 // previous BatchedModifications to the new leader.
 478                 Collection<BatchedModifications> newModifications = commitCoordinator.createForwardedBatchedModifications(
 479                         batched, datastoreContext.getShardBatchedModificationCount());
 480
 481                 LOG.debug("{}: Forwarding {} BatchedModifications to leader {}", persistenceId(),
 482                         newModifications.size(), leader);
 483
 484                 for(BatchedModifications bm: newModifications) {
 485                     leader.forward(bm, getContext());
 486                 }
 487             }
 488         }
 489     }
 490
 491     private boolean failIfIsolatedLeader(ActorRef sender) {
 492         if(isIsolatedLeader()) {
 493             sender.tell(new akka.actor.Status.Failure(new NoShardLeaderException(String.format(
 494                     "Shard %s was the leader but has lost contact with all of its followers. Either all" +
 495                     " other follower nodes are down or this node is isolated by a network partition.",
 496                     persistenceId()))), getSelf());
 497             return true;
 498         }
 499
 500         return false;
 501     }
 502
 503     protected boolean isIsolatedLeader() {
 504         return getRaftState() == RaftState.IsolatedLeader;
 505     }
 506
 507     private void handleReadyLocalTransaction(final ReadyLocalTransaction message) {
 508         LOG.debug("{}: handleReadyLocalTransaction for {}", persistenceId(), message.getTransactionID());
 509
 510         boolean isLeaderActive = isLeaderActive();
 511         if (isLeader() && isLeaderActive) {
 512             try {
 513                 commitCoordinator.handleReadyLocalTransaction(message, getSender(), this);
 514             } catch (Exception e) {
 515                 LOG.error("{}: Error handling ReadyLocalTransaction for Tx {}", persistenceId(),
 516                         message.getTransactionID(), e);
 517                 getSender().tell(new akka.actor.Status.Failure(e), getSelf());
 518             }
 519         } else {
 520             ActorSelection leader = getLeader();
 521             if (!isLeaderActive || leader == null) {
 522                 messageRetrySupport.addMessageToRetry(message, getSender(),
 523                         "Could not commit transaction " + message.getTransactionID());
 524             } else {
 525                 LOG.debug("{}: Forwarding ReadyLocalTransaction to leader {}", persistenceId(), leader);
 526                 message.setRemoteVersion(getCurrentBehavior().getLeaderPayloadVersion());
 527                 leader.forward(message, getContext());
 528             }
 529         }
 530     }
 531
 532     private void handleForwardedReadyTransaction(ForwardedReadyTransaction forwardedReady) {
 533         LOG.debug("{}: handleForwardedReadyTransaction for {}", persistenceId(), forwardedReady.getTransactionID());
 534
 535         boolean isLeaderActive = isLeaderActive();
 536         if (isLeader() && isLeaderActive) {
 537             commitCoordinator.handleForwardedReadyTransaction(forwardedReady, getSender(), this);
 538         } else {
 539             ActorSelection leader = getLeader();
 540             if (!isLeaderActive || leader == null) {
 541                 messageRetrySupport.addMessageToRetry(forwardedReady, getSender(),
 542                         "Could not commit transaction " + forwardedReady.getTransactionID());
 543             } else {
 544                 LOG.debug("{}: Forwarding ForwardedReadyTransaction to leader {}", persistenceId(), leader);
 545
 546                 ReadyLocalTransaction readyLocal = new ReadyLocalTransaction(forwardedReady.getTransactionID(),
 547                         forwardedReady.getTransaction().getSnapshot(), forwardedReady.isDoImmediateCommit());
 548                 readyLocal.setRemoteVersion(getCurrentBehavior().getLeaderPayloadVersion());
 549                 leader.forward(readyLocal, getContext());
 550             }
 551         }
 552     }
 553
 554     private void handleAbortTransaction(final AbortTransaction abort) {
 555         doAbortTransaction(abort.getTransactionID(), getSender());
 556     }
 557
 558     void doAbortTransaction(final String transactionID, final ActorRef sender) {
 559         commitCoordinator.handleAbort(transactionID, sender, this);
 560     }
 561
 562     private void handleCreateTransaction(final Object message) {
 563         if (isLeader()) {
 564             createTransaction(CreateTransaction.fromSerializable(message));
 565         } else if (getLeader() != null) {
 566             getLeader().forward(message, getContext());
 567         } else {
 568             getSender().tell(new akka.actor.Status.Failure(new NoShardLeaderException(
 569                     "Could not create a shard transaction", persistenceId())), getSelf());
 570         }
 571     }
 572
 573     private void closeTransactionChain(final CloseTransactionChain closeTransactionChain) {
 574         store.closeTransactionChain(closeTransactionChain.getTransactionChainId());
 575     }
 576
 577     private ActorRef createTypedTransactionActor(int transactionType,
 578             ShardTransactionIdentifier transactionId, String transactionChainId) {
 579
 580         return transactionActorFactory.newShardTransaction(TransactionType.fromInt(transactionType),
 581                 transactionId, transactionChainId);
 582     }
 583
 584     private void createTransaction(CreateTransaction createTransaction) {
 585         try {
 586             if(TransactionType.fromInt(createTransaction.getTransactionType()) != TransactionType.READ_ONLY &&
 587                     failIfIsolatedLeader(getSender())) {
 588                 return;
 589             }
 590
 591             ActorRef transactionActor = createTransaction(createTransaction.getTransactionType(),
 592                 createTransaction.getTransactionId(), createTransaction.getTransactionChainId());
 593
 594             getSender().tell(new CreateTransactionReply(Serialization.serializedActorPath(transactionActor),
 595                     createTransaction.getTransactionId(), createTransaction.getVersion()).toSerializable(), getSelf());
 596         } catch (Exception e) {
 597             getSender().tell(new akka.actor.Status.Failure(e), getSelf());
 598         }
 599     }
 600
 601     private ActorRef createTransaction(int transactionType, String remoteTransactionId,
 602             String transactionChainId) {
 603
 604
 605         ShardTransactionIdentifier transactionId = new ShardTransactionIdentifier(remoteTransactionId);
 606
 607         if(LOG.isDebugEnabled()) {
 608             LOG.debug("{}: Creating transaction : {} ", persistenceId(), transactionId);
 609         }
 610
 611         ActorRef transactionActor = createTypedTransactionActor(transactionType, transactionId,
 612                 transactionChainId);
 613
 614         return transactionActor;
 615     }
 616
 617     private void commitWithNewTransaction(final Modification modification) {
 618         ReadWriteShardDataTreeTransaction tx = store.newReadWriteTransaction(modification.toString(), null);
 619         modification.apply(tx.getSnapshot());
 620         try {
 621             snapshotCohort.syncCommitTransaction(tx);
 622             shardMBean.incrementCommittedTransactionCount();
 623             shardMBean.setLastCommittedTransactionTime(System.currentTimeMillis());
 624         } catch (Exception e) {
 625             shardMBean.incrementFailedTransactionsCount();
 626             LOG.error("{}: Failed to commit", persistenceId(), e);
 627         }
 628     }
 629
 630     private void updateSchemaContext(final UpdateSchemaContext message) {
 631         updateSchemaContext(message.getSchemaContext());
 632     }
 633
 634     @VisibleForTesting
 635     void updateSchemaContext(final SchemaContext schemaContext) {
 636         store.updateSchemaContext(schemaContext);
 637     }
 638
 639     private boolean isMetricsCaptureEnabled() {
 640         CommonConfig config = new CommonConfig(getContext().system().settings().config());
 641         return config.isMetricCaptureEnabled();
 642     }
 643
 644     @Override
 645     @VisibleForTesting
 646     public RaftActorSnapshotCohort getRaftActorSnapshotCohort() {
 647         return snapshotCohort;
 648     }
 649
 650     @Override
 651     @Nonnull
 652     protected RaftActorRecoveryCohort getRaftActorRecoveryCohort() {
 653         return new ShardRecoveryCoordinator(store, store.getSchemaContext(),
 654                 restoreFromSnapshot != null ? restoreFromSnapshot.getSnapshot() : null, persistenceId(), LOG);
 655     }
 656
 657     @Override
 658     protected void onRecoveryComplete() {
 659         restoreFromSnapshot = null;
 660
 661         //notify shard manager
 662         getContext().parent().tell(new ActorInitialized(), getSelf());
 663
 664         // Being paranoid here - this method should only be called once but just in case...
 665         if(txCommitTimeoutCheckSchedule == null) {
 666             // Schedule a message to be periodically sent to check if the current in-progress
 667             // transaction should be expired and aborted.
 668             FiniteDuration period = Duration.create(transactionCommitTimeout / 3, TimeUnit.MILLISECONDS);
 669             txCommitTimeoutCheckSchedule = getContext().system().scheduler().schedule(
 670                     period, period, getSelf(),
 671                     TX_COMMIT_TIMEOUT_CHECK_MESSAGE, getContext().dispatcher(), ActorRef.noSender());
 672         }
 673     }
 674
 675     @Override
 676     protected void applyState(final ActorRef clientActor, final String identifier, final Object data) {
 677         if (data instanceof DataTreeCandidatePayload) {
 678             if (clientActor == null) {
 679                 // No clientActor indicates a replica coming from the leader
 680                 try {
 681                     store.applyForeignCandidate(identifier, ((DataTreeCandidatePayload)data).getCandidate());
 682                 } catch (DataValidationFailedException | IOException e) {
 683                     LOG.error("{}: Error applying replica {}", persistenceId(), identifier, e);
 684                 }
 685             } else {
 686                 // Replication consensus reached, proceed to commit
 687                 finishCommit(clientActor, identifier);
 688             }
 689         } else if (data instanceof CompositeModificationPayload) {
 690             Object modification = ((CompositeModificationPayload) data).getModification();
 691
 692             applyModificationToState(clientActor, identifier, modification);
 693         } else if(data instanceof CompositeModificationByteStringPayload ){
 694             Object modification = ((CompositeModificationByteStringPayload) data).getModification();
 695
 696             applyModificationToState(clientActor, identifier, modification);
 697         } else {
 698             LOG.error("{}: Unknown state received {} Class loader = {} CompositeNodeMod.ClassLoader = {}",
 699                     persistenceId(), data, data.getClass().getClassLoader(),
 700                     CompositeModificationPayload.class.getClassLoader());
 701         }
 702     }
 703
 704     private void applyModificationToState(ActorRef clientActor, String identifier, Object modification) {
 705         if(modification == null) {
 706             LOG.error(
 707                     "{}: modification is null - this is very unexpected, clientActor = {}, identifier = {}",
 708                     persistenceId(), identifier, clientActor != null ? clientActor.path().toString() : null);
 709         } else if(clientActor == null) {
 710             // There's no clientActor to which to send a commit reply so we must be applying
 711             // replicated state from the leader.
 712             commitWithNewTransaction(MutableCompositeModification.fromSerializable(modification));
 713         } else {
 714             // This must be the OK to commit after replication consensus.
 715             finishCommit(clientActor, identifier);
 716         }
 717     }
 718
 719     @Override
 720     protected void onStateChanged() {
 721         boolean isLeader = isLeader();
 722         boolean hasLeader = hasLeader();
 723         changeSupport.onLeadershipChange(isLeader, hasLeader);
 724         treeChangeSupport.onLeadershipChange(isLeader, hasLeader);
 725
 726         // If this actor is no longer the leader close all the transaction chains
 727         if (!isLeader) {
 728             if(LOG.isDebugEnabled()) {
 729                 LOG.debug(
 730                     "{}: onStateChanged: Closing all transaction chains because shard {} is no longer the leader",
 731                     persistenceId(), getId());
 732             }
 733
 734             store.closeAllTransactionChains();
 735         }
 736
 737         if(hasLeader && !isIsolatedLeader()) {
 738             messageRetrySupport.retryMessages();
 739         }
 740     }
 741
 742     @Override
 743     protected void onLeaderChanged(String oldLeader, String newLeader) {
 744         shardMBean.incrementLeadershipChangeCount();
 745
 746         boolean hasLeader = hasLeader();
 747         if(hasLeader && !isLeader()) {
 748             // Another leader was elected. If we were the previous leader and had pending transactions, convert
 749             // them to transaction messages and send to the new leader.
 750             ActorSelection leader = getLeader();
 751             if(leader != null) {
 752                 Collection<Object> messagesToForward = commitCoordinator.convertPendingTransactionsToMessages(
 753                         datastoreContext.getShardBatchedModificationCount());
 754
 755                 if(!messagesToForward.isEmpty()) {
 756                     LOG.debug("{}: Forwarding {} pending transaction messages to leader {}", persistenceId(),
 757                             messagesToForward.size(), leader);
 758
 759                     for(Object message: messagesToForward) {
 760                         leader.tell(message, self());
 761                     }
 762                 }
 763             } else {
 764                 commitCoordinator.abortPendingTransactions(
 765                         "The transacton was aborted due to inflight leadership change and the leader address isn't available.",
 766                         this);
 767             }
 768         }
 769
 770         if(hasLeader && !isIsolatedLeader()) {
 771             messageRetrySupport.retryMessages();
 772         }
 773     }
 774
 775     @Override
 776     protected void pauseLeader(Runnable operation) {
 777         LOG.debug("{}: In pauseLeader, operation: {}", persistenceId(), operation);
 778         commitCoordinator.setRunOnPendingTransactionsComplete(operation);
 779     }
 780
 781     @Override
 782     public String persistenceId() {
 783         return this.name;
 784     }
 785
 786     @VisibleForTesting
 787     ShardCommitCoordinator getCommitCoordinator() {
 788         return commitCoordinator;
 789     }
 790
 791     public DatastoreContext getDatastoreContext() {
 792         return datastoreContext;
 793     }
 794
 795     @VisibleForTesting
 796     public ShardDataTree getDataStore() {
 797         return store;
 798     }
 799
 800     @VisibleForTesting
 801     ShardStats getShardMBean() {
 802         return shardMBean;
 803     }
 804
 805     public static Builder builder() {
 806         return new Builder();
 807     }
 808
 809     public static abstract class AbstractBuilder<T extends AbstractBuilder<T, S>, S extends Shard> {
 810         private final Class<S> shardClass;
 811         private ShardIdentifier id;
 812         private Map<String, String> peerAddresses = Collections.emptyMap();
 813         private DatastoreContext datastoreContext;
 814         private SchemaContext schemaContext;
 815         private DatastoreSnapshot.ShardSnapshot restoreFromSnapshot;
 816         private volatile boolean sealed;
 817
 818         protected AbstractBuilder(Class<S> shardClass) {
 819             this.shardClass = shardClass;
 820         }
 821
 822         protected void checkSealed() {
 823             Preconditions.checkState(!sealed, "Builder isalready sealed - further modifications are not allowed");
 824         }
 825
 826         @SuppressWarnings("unchecked")
 827         private T self() {
 828             return (T) this;
 829         }
 830
 831         public T id(ShardIdentifier id) {
 832             checkSealed();
 833             this.id = id;
 834             return self();
 835         }
 836
 837         public T peerAddresses(Map<String, String> peerAddresses) {
 838             checkSealed();
 839             this.peerAddresses = peerAddresses;
 840             return self();
 841         }
 842
 843         public T datastoreContext(DatastoreContext datastoreContext) {
 844             checkSealed();
 845             this.datastoreContext = datastoreContext;
 846             return self();
 847         }
 848
 849         public T schemaContext(SchemaContext schemaContext) {
 850             checkSealed();
 851             this.schemaContext = schemaContext;
 852             return self();
 853         }
 854
 855         public T restoreFromSnapshot(DatastoreSnapshot.ShardSnapshot restoreFromSnapshot) {
 856             checkSealed();
 857             this.restoreFromSnapshot = restoreFromSnapshot;
 858             return self();
 859         }
 860
 861         public ShardIdentifier getId() {
 862             return id;
 863         }
 864
 865         public Map<String, String> getPeerAddresses() {
 866             return peerAddresses;
 867         }
 868
 869         public DatastoreContext getDatastoreContext() {
 870             return datastoreContext;
 871         }
 872
 873         public SchemaContext getSchemaContext() {
 874             return schemaContext;
 875         }
 876
 877         public DatastoreSnapshot.ShardSnapshot getRestoreFromSnapshot() {
 878             return restoreFromSnapshot;
 879         }
 880
 881         public TreeType getTreeType() {
 882             switch (datastoreContext.getLogicalStoreType()) {
 883             case CONFIGURATION:
 884                 return TreeType.CONFIGURATION;
 885             case OPERATIONAL:
 886                 return TreeType.OPERATIONAL;
 887             }
 888
 889             throw new IllegalStateException("Unhandled logical store type " + datastoreContext.getLogicalStoreType());
 890         }
 891
 892         protected void verify() {
 893             Preconditions.checkNotNull(id, "id should not be null");
 894             Preconditions.checkNotNull(peerAddresses, "peerAddresses should not be null");
 895             Preconditions.checkNotNull(datastoreContext, "dataStoreContext should not be null");
 896             Preconditions.checkNotNull(schemaContext, "schemaContext should not be null");
 897         }
 898
 899         public Props props() {
 900             sealed = true;
 901             verify();
 902             return Props.create(shardClass, this);
 903         }
 904     }
 905
 906     public static class Builder extends AbstractBuilder<Builder, Shard> {
 907         private Builder() {
 908             super(Shard.class);
 909         }
 910     }
 911 }