opendaylight/md-sal/sal-distributed-datastore/src/main/java/org/opendaylight/controller/cluster/datastore/Shard.java

   1 /*
   2  * Copyright (c) 2014 Cisco Systems, Inc. and others.  All rights reserved.
   3  *
   4  * This program and the accompanying materials are made available under the
   5  * terms of the Eclipse Public License v1.0 which accompanies this distribution,
   6  * and is available at http://www.eclipse.org/legal/epl-v10.html
   7  */
   8
   9 package org.opendaylight.controller.cluster.datastore;
  10
  11 import akka.actor.ActorRef;
  12 import akka.actor.ActorSelection;
  13 import akka.actor.Cancellable;
  14 import akka.actor.Props;
  15 import akka.japi.Creator;
  16 import akka.persistence.RecoveryFailure;
  17 import akka.serialization.Serialization;
  18 import com.google.common.annotations.VisibleForTesting;
  19 import com.google.common.base.Optional;
  20 import com.google.common.base.Preconditions;
  21 import com.google.common.util.concurrent.FutureCallback;
  22 import com.google.common.util.concurrent.Futures;
  23 import com.google.common.util.concurrent.ListenableFuture;
  24 import java.io.IOException;
  25 import java.util.HashMap;
  26 import java.util.Map;
  27 import java.util.concurrent.ExecutionException;
  28 import java.util.concurrent.TimeUnit;
  29 import javax.annotation.Nonnull;
  30 import org.opendaylight.controller.cluster.common.actor.CommonConfig;
  31 import org.opendaylight.controller.cluster.common.actor.MeteringBehavior;
  32 import org.opendaylight.controller.cluster.datastore.ShardCommitCoordinator.CohortEntry;
  33 import org.opendaylight.controller.cluster.datastore.compat.BackwardsCompatibleThreePhaseCommitCohort;
  34 import org.opendaylight.controller.cluster.datastore.exceptions.NoShardLeaderException;
  35 import org.opendaylight.controller.cluster.datastore.identifiers.ShardIdentifier;
  36 import org.opendaylight.controller.cluster.datastore.identifiers.ShardTransactionIdentifier;
  37 import org.opendaylight.controller.cluster.datastore.jmx.mbeans.shard.ShardMBeanFactory;
  38 import org.opendaylight.controller.cluster.datastore.jmx.mbeans.shard.ShardStats;
  39 import org.opendaylight.controller.cluster.datastore.messages.AbortTransaction;
  40 import org.opendaylight.controller.cluster.datastore.messages.AbortTransactionReply;
  41 import org.opendaylight.controller.cluster.datastore.messages.ActorInitialized;
  42 import org.opendaylight.controller.cluster.datastore.messages.BatchedModifications;
  43 import org.opendaylight.controller.cluster.datastore.messages.BatchedModificationsReply;
  44 import org.opendaylight.controller.cluster.datastore.messages.CanCommitTransaction;
  45 import org.opendaylight.controller.cluster.datastore.messages.CloseTransactionChain;
  46 import org.opendaylight.controller.cluster.datastore.messages.CommitTransaction;
  47 import org.opendaylight.controller.cluster.datastore.messages.CommitTransactionReply;
  48 import org.opendaylight.controller.cluster.datastore.messages.CreateTransaction;
  49 import org.opendaylight.controller.cluster.datastore.messages.CreateTransactionReply;
  50 import org.opendaylight.controller.cluster.datastore.messages.ForwardedReadyTransaction;
  51 import org.opendaylight.controller.cluster.datastore.messages.PeerAddressResolved;
  52 import org.opendaylight.controller.cluster.datastore.messages.ReadyTransactionReply;
  53 import org.opendaylight.controller.cluster.datastore.messages.RegisterChangeListener;
  54 import org.opendaylight.controller.cluster.datastore.messages.RegisterDataTreeChangeListener;
  55 import org.opendaylight.controller.cluster.datastore.messages.UpdateSchemaContext;
  56 import org.opendaylight.controller.cluster.datastore.modification.Modification;
  57 import org.opendaylight.controller.cluster.datastore.modification.ModificationPayload;
  58 import org.opendaylight.controller.cluster.datastore.modification.MutableCompositeModification;
  59 import org.opendaylight.controller.cluster.datastore.utils.Dispatchers;
  60 import org.opendaylight.controller.cluster.datastore.utils.MessageTracker;
  61 import org.opendaylight.controller.cluster.notifications.RegisterRoleChangeListener;
  62 import org.opendaylight.controller.cluster.notifications.RoleChangeNotifier;
  63 import org.opendaylight.controller.cluster.raft.RaftActor;
  64 import org.opendaylight.controller.cluster.raft.RaftActorRecoveryCohort;
  65 import org.opendaylight.controller.cluster.raft.RaftActorSnapshotCohort;
  66 import org.opendaylight.controller.cluster.raft.base.messages.FollowerInitialSyncUpStatus;
  67 import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
  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.controller.md.sal.dom.store.impl.InMemoryDOMDataStore;
  71 import org.opendaylight.controller.md.sal.dom.store.impl.InMemoryDOMDataStoreFactory;
  72 import org.opendaylight.controller.sal.core.spi.data.DOMStoreWriteTransaction;
  73 import org.opendaylight.yangtools.yang.model.api.SchemaContext;
  74 import scala.concurrent.duration.Duration;
  75 import scala.concurrent.duration.FiniteDuration;
  76
  77 /**
  78  * A Shard represents a portion of the logical data tree <br/>
  79  * <p>
  80  * Our Shard uses InMemoryDataStore as it's internal representation and delegates all requests it
  81  * </p>
  82  */
  83 public class Shard extends RaftActor {
  84
  85     private static final Object TX_COMMIT_TIMEOUT_CHECK_MESSAGE = "txCommitTimeoutCheck";
  86
  87     @VisibleForTesting
  88     static final String DEFAULT_NAME = "default";
  89
  90     // The state of this Shard
  91     private final InMemoryDOMDataStore store;
  92
  93     /// The name of this shard
  94     private final String name;
  95
  96     private final ShardStats shardMBean;
  97
  98     private DatastoreContext datastoreContext;
  99
 100     private final ShardCommitCoordinator commitCoordinator;
 101
 102     private long transactionCommitTimeout;
 103
 104     private Cancellable txCommitTimeoutCheckSchedule;
 105
 106     private final Optional<ActorRef> roleChangeNotifier;
 107
 108     private final MessageTracker appendEntriesReplyTracker;
 109
 110     private final ReadyTransactionReply READY_TRANSACTION_REPLY = new ReadyTransactionReply(
 111             Serialization.serializedActorPath(getSelf()));
 112
 113     private final DOMTransactionFactory domTransactionFactory;
 114
 115     private final ShardTransactionActorFactory transactionActorFactory;
 116
 117     private final ShardSnapshotCohort snapshotCohort;
 118
 119     private final DataTreeChangeListenerSupport treeChangeSupport = new DataTreeChangeListenerSupport(this);
 120     private final DataChangeListenerSupport changeSupport = new DataChangeListenerSupport(this);
 121
 122     protected Shard(final ShardIdentifier name, final Map<String, String> peerAddresses,
 123             final DatastoreContext datastoreContext, final SchemaContext schemaContext) {
 124         super(name.toString(), new HashMap<>(peerAddresses), Optional.of(datastoreContext.getShardRaftConfig()));
 125
 126         this.name = name.toString();
 127         this.datastoreContext = datastoreContext;
 128
 129         setPersistence(datastoreContext.isPersistent());
 130
 131         LOG.info("Shard created : {}, persistent : {}", name, datastoreContext.isPersistent());
 132
 133         store = InMemoryDOMDataStoreFactory.create(name.toString(), null,
 134                 datastoreContext.getDataStoreProperties());
 135
 136         if (schemaContext != null) {
 137             store.onGlobalContextUpdated(schemaContext);
 138         }
 139
 140         shardMBean = ShardMBeanFactory.getShardStatsMBean(name.toString(),
 141                 datastoreContext.getDataStoreMXBeanType());
 142         shardMBean.setNotificationManager(store.getDataChangeListenerNotificationManager());
 143         shardMBean.setShardActor(getSelf());
 144
 145         if (isMetricsCaptureEnabled()) {
 146             getContext().become(new MeteringBehavior(this));
 147         }
 148
 149         domTransactionFactory = new DOMTransactionFactory(store, shardMBean, LOG, this.name);
 150
 151         commitCoordinator = new ShardCommitCoordinator(domTransactionFactory,
 152                 TimeUnit.SECONDS.convert(5, TimeUnit.MINUTES),
 153                 datastoreContext.getShardTransactionCommitQueueCapacity(), self(), LOG, this.name);
 154
 155         setTransactionCommitTimeout();
 156
 157         // create a notifier actor for each cluster member
 158         roleChangeNotifier = createRoleChangeNotifier(name.toString());
 159
 160         appendEntriesReplyTracker = new MessageTracker(AppendEntriesReply.class,
 161                 getRaftActorContext().getConfigParams().getIsolatedCheckIntervalInMillis());
 162
 163         transactionActorFactory = new ShardTransactionActorFactory(domTransactionFactory, datastoreContext,
 164                 new Dispatchers(context().system().dispatchers()).getDispatcherPath(
 165                         Dispatchers.DispatcherType.Transaction), self(), getContext(), shardMBean);
 166
 167         snapshotCohort = new ShardSnapshotCohort(transactionActorFactory, store, LOG, this.name);
 168     }
 169
 170     private void setTransactionCommitTimeout() {
 171         transactionCommitTimeout = TimeUnit.MILLISECONDS.convert(
 172                 datastoreContext.getShardTransactionCommitTimeoutInSeconds(), TimeUnit.SECONDS);
 173     }
 174
 175     public static Props props(final ShardIdentifier name,
 176         final Map<String, String> peerAddresses,
 177         final DatastoreContext datastoreContext, final SchemaContext schemaContext) {
 178         Preconditions.checkNotNull(name, "name should not be null");
 179         Preconditions.checkNotNull(peerAddresses, "peerAddresses should not be null");
 180         Preconditions.checkNotNull(datastoreContext, "dataStoreContext should not be null");
 181         Preconditions.checkNotNull(schemaContext, "schemaContext should not be null");
 182
 183         return Props.create(new ShardCreator(name, peerAddresses, datastoreContext, schemaContext));
 184     }
 185
 186     private Optional<ActorRef> createRoleChangeNotifier(String shardId) {
 187         ActorRef shardRoleChangeNotifier = this.getContext().actorOf(
 188             RoleChangeNotifier.getProps(shardId), shardId + "-notifier");
 189         return Optional.of(shardRoleChangeNotifier);
 190     }
 191
 192     @Override
 193     public void postStop() {
 194         LOG.info("Stopping Shard {}", persistenceId());
 195
 196         super.postStop();
 197
 198         if(txCommitTimeoutCheckSchedule != null) {
 199             txCommitTimeoutCheckSchedule.cancel();
 200         }
 201
 202         shardMBean.unregisterMBean();
 203     }
 204
 205     @Override
 206     public void onReceiveRecover(final Object message) throws Exception {
 207         if(LOG.isDebugEnabled()) {
 208             LOG.debug("{}: onReceiveRecover: Received message {} from {}", persistenceId(),
 209                 message.getClass().toString(), getSender());
 210         }
 211
 212         if (message instanceof RecoveryFailure){
 213             LOG.error("{}: Recovery failed because of this cause",
 214                     persistenceId(), ((RecoveryFailure) message).cause());
 215
 216             // Even though recovery failed, we still need to finish our recovery, eg send the
 217             // ActorInitialized message and start the txCommitTimeoutCheckSchedule.
 218             onRecoveryComplete();
 219         } else {
 220             super.onReceiveRecover(message);
 221             if(LOG.isTraceEnabled()) {
 222                 appendEntriesReplyTracker.begin();
 223             }
 224         }
 225     }
 226
 227     @Override
 228     public void onReceiveCommand(final Object message) throws Exception {
 229
 230         MessageTracker.Context context = appendEntriesReplyTracker.received(message);
 231
 232         if(context.error().isPresent()){
 233             LOG.trace("{} : AppendEntriesReply failed to arrive at the expected interval {}", persistenceId(),
 234                     context.error());
 235         }
 236
 237         try {
 238             if (CreateTransaction.SERIALIZABLE_CLASS.isInstance(message)) {
 239                 handleCreateTransaction(message);
 240             } else if (BatchedModifications.class.isInstance(message)) {
 241                 handleBatchedModifications((BatchedModifications)message);
 242             } else if (message instanceof ForwardedReadyTransaction) {
 243                 handleForwardedReadyTransaction((ForwardedReadyTransaction) message);
 244             } else if (CanCommitTransaction.SERIALIZABLE_CLASS.isInstance(message)) {
 245                 handleCanCommitTransaction(CanCommitTransaction.fromSerializable(message));
 246             } else if (CommitTransaction.SERIALIZABLE_CLASS.isInstance(message)) {
 247                 handleCommitTransaction(CommitTransaction.fromSerializable(message));
 248             } else if (AbortTransaction.SERIALIZABLE_CLASS.isInstance(message)) {
 249                 handleAbortTransaction(AbortTransaction.fromSerializable(message));
 250             } else if (CloseTransactionChain.SERIALIZABLE_CLASS.isInstance(message)) {
 251                 closeTransactionChain(CloseTransactionChain.fromSerializable(message));
 252             } else if (message instanceof RegisterChangeListener) {
 253                 changeSupport.onMessage((RegisterChangeListener) message, isLeader());
 254             } else if (message instanceof RegisterDataTreeChangeListener) {
 255                 treeChangeSupport.onMessage((RegisterDataTreeChangeListener) message, isLeader());
 256             } else if (message instanceof UpdateSchemaContext) {
 257                 updateSchemaContext((UpdateSchemaContext) message);
 258             } else if (message instanceof PeerAddressResolved) {
 259                 PeerAddressResolved resolved = (PeerAddressResolved) message;
 260                 setPeerAddress(resolved.getPeerId().toString(),
 261                         resolved.getPeerAddress());
 262             } else if (message.equals(TX_COMMIT_TIMEOUT_CHECK_MESSAGE)) {
 263                 handleTransactionCommitTimeoutCheck();
 264             } else if(message instanceof DatastoreContext) {
 265                 onDatastoreContext((DatastoreContext)message);
 266             } else if(message instanceof RegisterRoleChangeListener){
 267                 roleChangeNotifier.get().forward(message, context());
 268             } else if (message instanceof FollowerInitialSyncUpStatus){
 269                 shardMBean.setFollowerInitialSyncStatus(((FollowerInitialSyncUpStatus) message).isInitialSyncDone());
 270                 context().parent().tell(message, self());
 271             } else {
 272                 super.onReceiveCommand(message);
 273             }
 274         } finally {
 275             context.done();
 276         }
 277     }
 278
 279     @Override
 280     protected Optional<ActorRef> getRoleChangeNotifier() {
 281         return roleChangeNotifier;
 282     }
 283
 284     private void onDatastoreContext(DatastoreContext context) {
 285         datastoreContext = context;
 286
 287         commitCoordinator.setQueueCapacity(datastoreContext.getShardTransactionCommitQueueCapacity());
 288
 289         setTransactionCommitTimeout();
 290
 291         if(datastoreContext.isPersistent() && !persistence().isRecoveryApplicable()) {
 292             setPersistence(true);
 293         } else if(!datastoreContext.isPersistent() && persistence().isRecoveryApplicable()) {
 294             setPersistence(false);
 295         }
 296
 297         updateConfigParams(datastoreContext.getShardRaftConfig());
 298     }
 299
 300     private void handleTransactionCommitTimeoutCheck() {
 301         CohortEntry cohortEntry = commitCoordinator.getCurrentCohortEntry();
 302         if(cohortEntry != null) {
 303             long elapsed = System.currentTimeMillis() - cohortEntry.getLastAccessTime();
 304             if(elapsed > transactionCommitTimeout) {
 305                 LOG.warn("{}: Current transaction {} has timed out after {} ms - aborting",
 306                         persistenceId(), cohortEntry.getTransactionID(), transactionCommitTimeout);
 307
 308                 doAbortTransaction(cohortEntry.getTransactionID(), null);
 309             }
 310         }
 311     }
 312
 313     private void handleCommitTransaction(final CommitTransaction commit) {
 314         final String transactionID = commit.getTransactionID();
 315
 316         LOG.debug("{}: Committing transaction {}", persistenceId(), transactionID);
 317
 318         // Get the current in-progress cohort entry in the commitCoordinator if it corresponds to
 319         // this transaction.
 320         final CohortEntry cohortEntry = commitCoordinator.getCohortEntryIfCurrent(transactionID);
 321         if(cohortEntry == null) {
 322             // We're not the current Tx - the Tx was likely expired b/c it took too long in
 323             // between the canCommit and commit messages.
 324             IllegalStateException ex = new IllegalStateException(
 325                     String.format("%s: Cannot commit transaction %s - it is not the current transaction",
 326                             persistenceId(), transactionID));
 327             LOG.error(ex.getMessage());
 328             shardMBean.incrementFailedTransactionsCount();
 329             getSender().tell(new akka.actor.Status.Failure(ex), getSelf());
 330             return;
 331         }
 332
 333         // We perform the preCommit phase here atomically with the commit phase. This is an
 334         // optimization to eliminate the overhead of an extra preCommit message. We lose front-end
 335         // coordination of preCommit across shards in case of failure but preCommit should not
 336         // normally fail since we ensure only one concurrent 3-phase commit.
 337
 338         try {
 339             // We block on the future here so we don't have to worry about possibly accessing our
 340             // state on a different thread outside of our dispatcher. Also, the data store
 341             // currently uses a same thread executor anyway.
 342             cohortEntry.getCohort().preCommit().get();
 343
 344             // If we do not have any followers and we are not using persistence
 345             // or if cohortEntry has no modifications
 346             // we can apply modification to the state immediately
 347             if((!hasFollowers() && !persistence().isRecoveryApplicable()) || (!cohortEntry.hasModifications())){
 348                 applyModificationToState(getSender(), transactionID, cohortEntry.getModification());
 349             } else {
 350                 Shard.this.persistData(getSender(), transactionID,
 351                         new ModificationPayload(cohortEntry.getModification()));
 352             }
 353         } catch (Exception e) {
 354             LOG.error("{} An exception occurred while preCommitting transaction {}",
 355                     persistenceId(), cohortEntry.getTransactionID(), e);
 356             shardMBean.incrementFailedTransactionsCount();
 357             getSender().tell(new akka.actor.Status.Failure(e), getSelf());
 358         }
 359
 360         cohortEntry.updateLastAccessTime();
 361     }
 362
 363     private void finishCommit(@Nonnull final ActorRef sender, final @Nonnull String transactionID) {
 364         // With persistence enabled, this method is called via applyState by the leader strategy
 365         // after the commit has been replicated to a majority of the followers.
 366
 367         CohortEntry cohortEntry = commitCoordinator.getCohortEntryIfCurrent(transactionID);
 368         if(cohortEntry == null) {
 369             // The transaction is no longer the current commit. This can happen if the transaction
 370             // was aborted prior, most likely due to timeout in the front-end. We need to finish
 371             // committing the transaction though since it was successfully persisted and replicated
 372             // however we can't use the original cohort b/c it was already preCommitted and may
 373             // conflict with the current commit or may have been aborted so we commit with a new
 374             // transaction.
 375             cohortEntry = commitCoordinator.getAndRemoveCohortEntry(transactionID);
 376             if(cohortEntry != null) {
 377                 commitWithNewTransaction(cohortEntry.getModification());
 378                 sender.tell(CommitTransactionReply.INSTANCE.toSerializable(), getSelf());
 379             } else {
 380                 // This really shouldn't happen - it likely means that persistence or replication
 381                 // took so long to complete such that the cohort entry was expired from the cache.
 382                 IllegalStateException ex = new IllegalStateException(
 383                         String.format("%s: Could not finish committing transaction %s - no CohortEntry found",
 384                                 persistenceId(), transactionID));
 385                 LOG.error(ex.getMessage());
 386                 sender.tell(new akka.actor.Status.Failure(ex), getSelf());
 387             }
 388
 389             return;
 390         }
 391
 392         LOG.debug("{}: Finishing commit for transaction {}", persistenceId(), cohortEntry.getTransactionID());
 393
 394         try {
 395             // We block on the future here so we don't have to worry about possibly accessing our
 396             // state on a different thread outside of our dispatcher. Also, the data store
 397             // currently uses a same thread executor anyway.
 398             cohortEntry.getCohort().commit().get();
 399
 400             sender.tell(CommitTransactionReply.INSTANCE.toSerializable(), getSelf());
 401
 402             shardMBean.incrementCommittedTransactionCount();
 403             shardMBean.setLastCommittedTransactionTime(System.currentTimeMillis());
 404
 405         } catch (Exception e) {
 406             sender.tell(new akka.actor.Status.Failure(e), getSelf());
 407
 408             LOG.error("{}, An exception occurred while committing transaction {}", persistenceId(),
 409                     transactionID, e);
 410             shardMBean.incrementFailedTransactionsCount();
 411         } finally {
 412             commitCoordinator.currentTransactionComplete(transactionID, true);
 413         }
 414     }
 415
 416     private void handleCanCommitTransaction(final CanCommitTransaction canCommit) {
 417         LOG.debug("{}: Can committing transaction {}", persistenceId(), canCommit.getTransactionID());
 418         commitCoordinator.handleCanCommit(canCommit, getSender(), self());
 419     }
 420
 421     private void handleBatchedModifications(BatchedModifications batched) {
 422         // This message is sent to prepare the modificationsa transaction directly on the Shard as an
 423         // optimization to avoid the extra overhead of a separate ShardTransaction actor. On the last
 424         // BatchedModifications message, the caller sets the ready flag in the message indicating
 425         // modifications are complete. The reply contains the cohort actor path (this actor) for the caller
 426         // to initiate the 3-phase commit. This also avoids the overhead of sending an additional
 427         // ReadyTransaction message.
 428
 429         // If we're not the leader then forward to the leader. This is a safety measure - we shouldn't
 430         // normally get here if we're not the leader as the front-end (TransactionProxy) should determine
 431         // the primary/leader shard. However with timing and caching on the front-end, there's a small
 432         // window where it could have a stale leader during leadership transitions.
 433         //
 434         if(isLeader()) {
 435             try {
 436                 boolean ready = commitCoordinator.handleTransactionModifications(batched);
 437                 if(ready) {
 438                     sender().tell(READY_TRANSACTION_REPLY, self());
 439                 } else {
 440                     sender().tell(new BatchedModificationsReply(batched.getModifications().size()), self());
 441                 }
 442             } catch (Exception e) {
 443                 LOG.error("{}: Error handling BatchedModifications for Tx {}", persistenceId(),
 444                         batched.getTransactionID(), e);
 445                 getSender().tell(new akka.actor.Status.Failure(e), getSelf());
 446             }
 447         } else {
 448             ActorSelection leader = getLeader();
 449             if(leader != null) {
 450                 // TODO: what if this is not the first batch and leadership changed in between batched messages?
 451                 // We could check if the commitCoordinator already has a cached entry and forward all the previous
 452                 // batched modifications.
 453                 LOG.debug("{}: Forwarding BatchedModifications to leader {}", persistenceId(), leader);
 454                 leader.forward(batched, getContext());
 455             } else {
 456                 // TODO: rather than throwing an immediate exception, we could schedule a timer to try again to make
 457                 // it more resilient in case we're in the process of electing a new leader.
 458                 getSender().tell(new akka.actor.Status.Failure(new NoShardLeaderException(String.format(
 459                     "Could not find the leader for shard %s. This typically happens" +
 460                     " when the system is coming up or recovering and a leader is being elected. Try again" +
 461                     " later.", persistenceId()))), getSelf());
 462             }
 463         }
 464     }
 465
 466     private void handleForwardedReadyTransaction(ForwardedReadyTransaction ready) {
 467         LOG.debug("{}: Readying transaction {}, client version {}", persistenceId(),
 468                 ready.getTransactionID(), ready.getTxnClientVersion());
 469
 470         // This message is forwarded by the ShardTransaction on ready. We cache the cohort in the
 471         // commitCoordinator in preparation for the subsequent three phase commit initiated by
 472         // the front-end.
 473         commitCoordinator.transactionReady(ready.getTransactionID(), ready.getCohort(),
 474                 (MutableCompositeModification) ready.getModification());
 475
 476         // Return our actor path as we'll handle the three phase commit, except if the Tx client
 477         // version < 1 (Helium-1 version). This means the Tx was initiated by a base Helium version
 478         // node. In that case, the subsequent 3-phase commit messages won't contain the
 479         // transactionId so to maintain backwards compatibility, we create a separate cohort actor
 480         // to provide the compatible behavior.
 481         if(ready.getTxnClientVersion() < DataStoreVersions.LITHIUM_VERSION) {
 482             ActorRef replyActorPath = getSelf();
 483             if(ready.getTxnClientVersion() < DataStoreVersions.HELIUM_1_VERSION) {
 484                 LOG.debug("{}: Creating BackwardsCompatibleThreePhaseCommitCohort", persistenceId());
 485                 replyActorPath = getContext().actorOf(BackwardsCompatibleThreePhaseCommitCohort.props(
 486                         ready.getTransactionID()));
 487             }
 488
 489             ReadyTransactionReply readyTransactionReply =
 490                     new ReadyTransactionReply(Serialization.serializedActorPath(replyActorPath),
 491                             ready.getTxnClientVersion());
 492             getSender().tell(ready.isReturnSerialized() ? readyTransactionReply.toSerializable() :
 493                 readyTransactionReply, getSelf());
 494         } else {
 495             getSender().tell(READY_TRANSACTION_REPLY, getSelf());
 496         }
 497     }
 498
 499     private void handleAbortTransaction(final AbortTransaction abort) {
 500         doAbortTransaction(abort.getTransactionID(), getSender());
 501     }
 502
 503     void doAbortTransaction(final String transactionID, final ActorRef sender) {
 504         final CohortEntry cohortEntry = commitCoordinator.getCohortEntryIfCurrent(transactionID);
 505         if(cohortEntry != null) {
 506             LOG.debug("{}: Aborting transaction {}", persistenceId(), transactionID);
 507
 508             // We don't remove the cached cohort entry here (ie pass false) in case the Tx was
 509             // aborted during replication in which case we may still commit locally if replication
 510             // succeeds.
 511             commitCoordinator.currentTransactionComplete(transactionID, false);
 512
 513             final ListenableFuture<Void> future = cohortEntry.getCohort().abort();
 514             final ActorRef self = getSelf();
 515
 516             Futures.addCallback(future, new FutureCallback<Void>() {
 517                 @Override
 518                 public void onSuccess(final Void v) {
 519                     shardMBean.incrementAbortTransactionsCount();
 520
 521                     if(sender != null) {
 522                         sender.tell(AbortTransactionReply.INSTANCE.toSerializable(), self);
 523                     }
 524                 }
 525
 526                 @Override
 527                 public void onFailure(final Throwable t) {
 528                     LOG.error("{}: An exception happened during abort", persistenceId(), t);
 529
 530                     if(sender != null) {
 531                         sender.tell(new akka.actor.Status.Failure(t), self);
 532                     }
 533                 }
 534             });
 535         }
 536     }
 537
 538     private void handleCreateTransaction(final Object message) {
 539         if (isLeader()) {
 540             createTransaction(CreateTransaction.fromSerializable(message));
 541         } else if (getLeader() != null) {
 542             getLeader().forward(message, getContext());
 543         } else {
 544             getSender().tell(new akka.actor.Status.Failure(new NoShardLeaderException(String.format(
 545                 "Could not find leader for shard %s so transaction cannot be created. This typically happens" +
 546                 " when the system is coming up or recovering and a leader is being elected. Try again" +
 547                 " later.", persistenceId()))), getSelf());
 548         }
 549     }
 550
 551     private void closeTransactionChain(final CloseTransactionChain closeTransactionChain) {
 552         domTransactionFactory.closeTransactionChain(closeTransactionChain.getTransactionChainId());
 553     }
 554
 555     private ActorRef createTypedTransactionActor(int transactionType,
 556             ShardTransactionIdentifier transactionId, String transactionChainId,
 557             short clientVersion ) {
 558
 559         return transactionActorFactory.newShardTransaction(TransactionProxy.TransactionType.fromInt(transactionType),
 560                 transactionId, transactionChainId, clientVersion);
 561     }
 562
 563     private void createTransaction(CreateTransaction createTransaction) {
 564         try {
 565             ActorRef transactionActor = createTransaction(createTransaction.getTransactionType(),
 566                 createTransaction.getTransactionId(), createTransaction.getTransactionChainId(),
 567                 createTransaction.getVersion());
 568
 569             getSender().tell(new CreateTransactionReply(Serialization.serializedActorPath(transactionActor),
 570                     createTransaction.getTransactionId()).toSerializable(), getSelf());
 571         } catch (Exception e) {
 572             getSender().tell(new akka.actor.Status.Failure(e), getSelf());
 573         }
 574     }
 575
 576     private ActorRef createTransaction(int transactionType, String remoteTransactionId,
 577             String transactionChainId, short clientVersion) {
 578
 579
 580         ShardTransactionIdentifier transactionId = new ShardTransactionIdentifier(remoteTransactionId);
 581
 582         if(LOG.isDebugEnabled()) {
 583             LOG.debug("{}: Creating transaction : {} ", persistenceId(), transactionId);
 584         }
 585
 586         ActorRef transactionActor = createTypedTransactionActor(transactionType, transactionId,
 587                 transactionChainId, clientVersion);
 588
 589         return transactionActor;
 590     }
 591
 592     private void commitWithNewTransaction(final Modification modification) {
 593         DOMStoreWriteTransaction tx = store.newWriteOnlyTransaction();
 594         modification.apply(tx);
 595         try {
 596             snapshotCohort.syncCommitTransaction(tx);
 597             shardMBean.incrementCommittedTransactionCount();
 598             shardMBean.setLastCommittedTransactionTime(System.currentTimeMillis());
 599         } catch (InterruptedException | ExecutionException e) {
 600             shardMBean.incrementFailedTransactionsCount();
 601             LOG.error("{}: Failed to commit", persistenceId(), e);
 602         }
 603     }
 604
 605     private void updateSchemaContext(final UpdateSchemaContext message) {
 606         updateSchemaContext(message.getSchemaContext());
 607     }
 608
 609     @VisibleForTesting
 610     void updateSchemaContext(final SchemaContext schemaContext) {
 611         store.onGlobalContextUpdated(schemaContext);
 612     }
 613
 614     private boolean isMetricsCaptureEnabled() {
 615         CommonConfig config = new CommonConfig(getContext().system().settings().config());
 616         return config.isMetricCaptureEnabled();
 617     }
 618
 619     @Override
 620     protected RaftActorSnapshotCohort getRaftActorSnapshotCohort() {
 621         return snapshotCohort;
 622     }
 623
 624     @Override
 625     @Nonnull
 626     protected RaftActorRecoveryCohort getRaftActorRecoveryCohort() {
 627         return new ShardRecoveryCoordinator(store, persistenceId(), LOG);
 628     }
 629
 630     @Override
 631     protected void onRecoveryComplete() {
 632         //notify shard manager
 633         getContext().parent().tell(new ActorInitialized(), getSelf());
 634
 635         // Being paranoid here - this method should only be called once but just in case...
 636         if(txCommitTimeoutCheckSchedule == null) {
 637             // Schedule a message to be periodically sent to check if the current in-progress
 638             // transaction should be expired and aborted.
 639             FiniteDuration period = Duration.create(transactionCommitTimeout / 3, TimeUnit.MILLISECONDS);
 640             txCommitTimeoutCheckSchedule = getContext().system().scheduler().schedule(
 641                     period, period, getSelf(),
 642                     TX_COMMIT_TIMEOUT_CHECK_MESSAGE, getContext().dispatcher(), ActorRef.noSender());
 643         }
 644     }
 645
 646     @Override
 647     protected void applyState(final ActorRef clientActor, final String identifier, final Object data) {
 648
 649         if(data instanceof ModificationPayload) {
 650             try {
 651                 applyModificationToState(clientActor, identifier, ((ModificationPayload) data).getModification());
 652             } catch (ClassNotFoundException | IOException e) {
 653                 LOG.error("{}: Error extracting ModificationPayload", persistenceId(), e);
 654             }
 655         }
 656         else if (data instanceof CompositeModificationPayload) {
 657             Object modification = ((CompositeModificationPayload) data).getModification();
 658
 659             applyModificationToState(clientActor, identifier, modification);
 660         } else if(data instanceof CompositeModificationByteStringPayload ){
 661             Object modification = ((CompositeModificationByteStringPayload) data).getModification();
 662
 663             applyModificationToState(clientActor, identifier, modification);
 664         } else {
 665             LOG.error("{}: Unknown state received {} Class loader = {} CompositeNodeMod.ClassLoader = {}",
 666                     persistenceId(), data, data.getClass().getClassLoader(),
 667                     CompositeModificationPayload.class.getClassLoader());
 668         }
 669     }
 670
 671     private void applyModificationToState(ActorRef clientActor, String identifier, Object modification) {
 672         if(modification == null) {
 673             LOG.error(
 674                     "{}: modification is null - this is very unexpected, clientActor = {}, identifier = {}",
 675                     persistenceId(), identifier, clientActor != null ? clientActor.path().toString() : null);
 676         } else if(clientActor == null) {
 677             // There's no clientActor to which to send a commit reply so we must be applying
 678             // replicated state from the leader.
 679             commitWithNewTransaction(MutableCompositeModification.fromSerializable(modification));
 680         } else {
 681             // This must be the OK to commit after replication consensus.
 682             finishCommit(clientActor, identifier);
 683         }
 684     }
 685
 686     @Override
 687     protected void onStateChanged() {
 688         boolean isLeader = isLeader();
 689         changeSupport.onLeadershipChange(isLeader);
 690         treeChangeSupport.onLeadershipChange(isLeader);
 691
 692         // If this actor is no longer the leader close all the transaction chains
 693         if (!isLeader) {
 694             if(LOG.isDebugEnabled()) {
 695                 LOG.debug(
 696                     "{}: onStateChanged: Closing all transaction chains because shard {} is no longer the leader",
 697                     persistenceId(), getId());
 698             }
 699
 700             domTransactionFactory.closeAllTransactionChains();
 701         }
 702     }
 703
 704     @Override
 705     protected void onLeaderChanged(String oldLeader, String newLeader) {
 706         shardMBean.incrementLeadershipChangeCount();
 707     }
 708
 709     @Override
 710     public String persistenceId() {
 711         return this.name;
 712     }
 713
 714     @VisibleForTesting
 715     ShardCommitCoordinator getCommitCoordinator() {
 716         return commitCoordinator;
 717     }
 718
 719
 720     private static class ShardCreator implements Creator<Shard> {
 721
 722         private static final long serialVersionUID = 1L;
 723
 724         final ShardIdentifier name;
 725         final Map<String, String> peerAddresses;
 726         final DatastoreContext datastoreContext;
 727         final SchemaContext schemaContext;
 728
 729         ShardCreator(final ShardIdentifier name, final Map<String, String> peerAddresses,
 730                 final DatastoreContext datastoreContext, final SchemaContext schemaContext) {
 731             this.name = name;
 732             this.peerAddresses = peerAddresses;
 733             this.datastoreContext = datastoreContext;
 734             this.schemaContext = schemaContext;
 735         }
 736
 737         @Override
 738         public Shard create() throws Exception {
 739             return new Shard(name, peerAddresses, datastoreContext, schemaContext);
 740         }
 741     }
 742
 743     @VisibleForTesting
 744     public InMemoryDOMDataStore getDataStore() {
 745         return store;
 746     }
 747
 748     @VisibleForTesting
 749     ShardStats getShardMBean() {
 750         return shardMBean;
 751     }
 752 }