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.CreateSnapshot;
  49 import org.opendaylight.controller.cluster.datastore.messages.CreateTransaction;
  50 import org.opendaylight.controller.cluster.datastore.messages.CreateTransactionReply;
  51 import org.opendaylight.controller.cluster.datastore.messages.ForwardedReadyTransaction;
  52 import org.opendaylight.controller.cluster.datastore.messages.PeerAddressResolved;
  53 import org.opendaylight.controller.cluster.datastore.messages.ReadyTransactionReply;
  54 import org.opendaylight.controller.cluster.datastore.messages.RegisterChangeListener;
  55 import org.opendaylight.controller.cluster.datastore.messages.RegisterDataTreeChangeListener;
  56 import org.opendaylight.controller.cluster.datastore.messages.UpdateSchemaContext;
  57 import org.opendaylight.controller.cluster.datastore.modification.Modification;
  58 import org.opendaylight.controller.cluster.datastore.modification.ModificationPayload;
  59 import org.opendaylight.controller.cluster.datastore.modification.MutableCompositeModification;
  60 import org.opendaylight.controller.cluster.datastore.utils.Dispatchers;
  61 import org.opendaylight.controller.cluster.datastore.utils.MessageTracker;
  62 import org.opendaylight.controller.cluster.datastore.utils.SerializationUtils;
  63 import org.opendaylight.controller.cluster.notifications.RegisterRoleChangeListener;
  64 import org.opendaylight.controller.cluster.notifications.RoleChangeNotifier;
  65 import org.opendaylight.controller.cluster.raft.RaftActor;
  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.cluster.raft.protobuff.client.messages.Payload;
  71 import org.opendaylight.controller.md.sal.dom.store.impl.InMemoryDOMDataStore;
  72 import org.opendaylight.controller.md.sal.dom.store.impl.InMemoryDOMDataStoreFactory;
  73 import org.opendaylight.controller.sal.core.spi.data.DOMStoreThreePhaseCommitCohort;
  74 import org.opendaylight.controller.sal.core.spi.data.DOMStoreTransaction;
  75 import org.opendaylight.controller.sal.core.spi.data.DOMStoreWriteTransaction;
  76 import org.opendaylight.yangtools.yang.data.api.YangInstanceIdentifier;
  77 import org.opendaylight.yangtools.yang.data.api.schema.NormalizedNode;
  78 import org.opendaylight.yangtools.yang.model.api.SchemaContext;
  79 import scala.concurrent.duration.Duration;
  80 import scala.concurrent.duration.FiniteDuration;
  81
  82 /**
  83  * A Shard represents a portion of the logical data tree <br/>
  84  * <p>
  85  * Our Shard uses InMemoryDataStore as it's internal representation and delegates all requests it
  86  * </p>
  87  */
  88 public class Shard extends RaftActor {
  89
  90     private static final YangInstanceIdentifier DATASTORE_ROOT = YangInstanceIdentifier.builder().build();
  91
  92     private static final Object TX_COMMIT_TIMEOUT_CHECK_MESSAGE = "txCommitTimeoutCheck";
  93
  94     @VisibleForTesting
  95     static final String DEFAULT_NAME = "default";
  96
  97     // The state of this Shard
  98     private final InMemoryDOMDataStore store;
  99
 100     /// The name of this shard
 101     private final String name;
 102
 103     private final ShardStats shardMBean;
 104
 105     private DatastoreContext datastoreContext;
 106
 107     private SchemaContext schemaContext;
 108
 109     private int createSnapshotTransactionCounter;
 110
 111     private final ShardCommitCoordinator commitCoordinator;
 112
 113     private long transactionCommitTimeout;
 114
 115     private Cancellable txCommitTimeoutCheckSchedule;
 116
 117     private final Optional<ActorRef> roleChangeNotifier;
 118
 119     private final MessageTracker appendEntriesReplyTracker;
 120
 121     private final ReadyTransactionReply READY_TRANSACTION_REPLY = new ReadyTransactionReply(
 122             Serialization.serializedActorPath(getSelf()));
 123
 124
 125     /**
 126      * Coordinates persistence recovery on startup.
 127      */
 128     private ShardRecoveryCoordinator recoveryCoordinator;
 129
 130     private final DOMTransactionFactory transactionFactory;
 131
 132     private final String txnDispatcherPath;
 133
 134     private final DataTreeChangeListenerSupport treeChangeSupport = new DataTreeChangeListenerSupport(this);
 135     private final DataChangeListenerSupport changeSupport = new DataChangeListenerSupport(this);
 136
 137     protected Shard(final ShardIdentifier name, final Map<String, String> peerAddresses,
 138             final DatastoreContext datastoreContext, final SchemaContext schemaContext) {
 139         super(name.toString(), new HashMap<>(peerAddresses), Optional.of(datastoreContext.getShardRaftConfig()));
 140
 141         this.name = name.toString();
 142         this.datastoreContext = datastoreContext;
 143         this.schemaContext = schemaContext;
 144         this.txnDispatcherPath = new Dispatchers(context().system().dispatchers())
 145                 .getDispatcherPath(Dispatchers.DispatcherType.Transaction);
 146
 147         setPersistence(datastoreContext.isPersistent());
 148
 149         LOG.info("Shard created : {}, persistent : {}", name, datastoreContext.isPersistent());
 150
 151         store = InMemoryDOMDataStoreFactory.create(name.toString(), null,
 152                 datastoreContext.getDataStoreProperties());
 153
 154         if (schemaContext != null) {
 155             store.onGlobalContextUpdated(schemaContext);
 156         }
 157
 158         shardMBean = ShardMBeanFactory.getShardStatsMBean(name.toString(),
 159                 datastoreContext.getDataStoreMXBeanType());
 160         shardMBean.setNotificationManager(store.getDataChangeListenerNotificationManager());
 161         shardMBean.setShardActor(getSelf());
 162
 163         if (isMetricsCaptureEnabled()) {
 164             getContext().become(new MeteringBehavior(this));
 165         }
 166
 167         transactionFactory = new DOMTransactionFactory(store, shardMBean, LOG, this.name);
 168
 169         commitCoordinator = new ShardCommitCoordinator(transactionFactory,
 170                 TimeUnit.SECONDS.convert(5, TimeUnit.MINUTES),
 171                 datastoreContext.getShardTransactionCommitQueueCapacity(), self(), LOG, this.name);
 172
 173         setTransactionCommitTimeout();
 174
 175         // create a notifier actor for each cluster member
 176         roleChangeNotifier = createRoleChangeNotifier(name.toString());
 177
 178         appendEntriesReplyTracker = new MessageTracker(AppendEntriesReply.class,
 179                 getRaftActorContext().getConfigParams().getIsolatedCheckIntervalInMillis());
 180
 181         recoveryCoordinator = new ShardRecoveryCoordinator(store, persistenceId(), LOG);
 182     }
 183
 184     private void setTransactionCommitTimeout() {
 185         transactionCommitTimeout = TimeUnit.MILLISECONDS.convert(
 186                 datastoreContext.getShardTransactionCommitTimeoutInSeconds(), TimeUnit.SECONDS);
 187     }
 188
 189     public static Props props(final ShardIdentifier name,
 190         final Map<String, String> peerAddresses,
 191         final DatastoreContext datastoreContext, final SchemaContext schemaContext) {
 192         Preconditions.checkNotNull(name, "name should not be null");
 193         Preconditions.checkNotNull(peerAddresses, "peerAddresses should not be null");
 194         Preconditions.checkNotNull(datastoreContext, "dataStoreContext should not be null");
 195         Preconditions.checkNotNull(schemaContext, "schemaContext should not be null");
 196
 197         return Props.create(new ShardCreator(name, peerAddresses, datastoreContext, schemaContext));
 198     }
 199
 200     private Optional<ActorRef> createRoleChangeNotifier(String shardId) {
 201         ActorRef shardRoleChangeNotifier = this.getContext().actorOf(
 202             RoleChangeNotifier.getProps(shardId), shardId + "-notifier");
 203         return Optional.of(shardRoleChangeNotifier);
 204     }
 205
 206     @Override
 207     public void postStop() {
 208         LOG.info("Stopping Shard {}", persistenceId());
 209
 210         super.postStop();
 211
 212         if(txCommitTimeoutCheckSchedule != null) {
 213             txCommitTimeoutCheckSchedule.cancel();
 214         }
 215
 216         shardMBean.unregisterMBean();
 217     }
 218
 219     @Override
 220     public void onReceiveRecover(final Object message) throws Exception {
 221         if(LOG.isDebugEnabled()) {
 222             LOG.debug("{}: onReceiveRecover: Received message {} from {}", persistenceId(),
 223                 message.getClass().toString(), getSender());
 224         }
 225
 226         if (message instanceof RecoveryFailure){
 227             LOG.error("{}: Recovery failed because of this cause",
 228                     persistenceId(), ((RecoveryFailure) message).cause());
 229
 230             // Even though recovery failed, we still need to finish our recovery, eg send the
 231             // ActorInitialized message and start the txCommitTimeoutCheckSchedule.
 232             onRecoveryComplete();
 233         } else {
 234             super.onReceiveRecover(message);
 235             if(LOG.isTraceEnabled()) {
 236                 appendEntriesReplyTracker.begin();
 237             }
 238         }
 239     }
 240
 241     @Override
 242     public void onReceiveCommand(final Object message) throws Exception {
 243
 244         MessageTracker.Context context = appendEntriesReplyTracker.received(message);
 245
 246         if(context.error().isPresent()){
 247             LOG.trace("{} : AppendEntriesReply failed to arrive at the expected interval {}", persistenceId(),
 248                     context.error());
 249         }
 250
 251         try {
 252             if (CreateTransaction.SERIALIZABLE_CLASS.isInstance(message)) {
 253                 handleCreateTransaction(message);
 254             } else if (BatchedModifications.class.isInstance(message)) {
 255                 handleBatchedModifications((BatchedModifications)message);
 256             } else if (message instanceof ForwardedReadyTransaction) {
 257                 handleForwardedReadyTransaction((ForwardedReadyTransaction) message);
 258             } else if (CanCommitTransaction.SERIALIZABLE_CLASS.isInstance(message)) {
 259                 handleCanCommitTransaction(CanCommitTransaction.fromSerializable(message));
 260             } else if (CommitTransaction.SERIALIZABLE_CLASS.isInstance(message)) {
 261                 handleCommitTransaction(CommitTransaction.fromSerializable(message));
 262             } else if (AbortTransaction.SERIALIZABLE_CLASS.isInstance(message)) {
 263                 handleAbortTransaction(AbortTransaction.fromSerializable(message));
 264             } else if (CloseTransactionChain.SERIALIZABLE_CLASS.isInstance(message)) {
 265                 closeTransactionChain(CloseTransactionChain.fromSerializable(message));
 266             } else if (message instanceof RegisterChangeListener) {
 267                 changeSupport.onMessage((RegisterChangeListener) message, isLeader());
 268             } else if (message instanceof RegisterDataTreeChangeListener) {
 269                 treeChangeSupport.onMessage((RegisterDataTreeChangeListener) message, isLeader());
 270             } else if (message instanceof UpdateSchemaContext) {
 271                 updateSchemaContext((UpdateSchemaContext) message);
 272             } else if (message instanceof PeerAddressResolved) {
 273                 PeerAddressResolved resolved = (PeerAddressResolved) message;
 274                 setPeerAddress(resolved.getPeerId().toString(),
 275                         resolved.getPeerAddress());
 276             } else if (message.equals(TX_COMMIT_TIMEOUT_CHECK_MESSAGE)) {
 277                 handleTransactionCommitTimeoutCheck();
 278             } else if(message instanceof DatastoreContext) {
 279                 onDatastoreContext((DatastoreContext)message);
 280             } else if(message instanceof RegisterRoleChangeListener){
 281                 roleChangeNotifier.get().forward(message, context());
 282             } else if (message instanceof FollowerInitialSyncUpStatus){
 283                 shardMBean.setFollowerInitialSyncStatus(((FollowerInitialSyncUpStatus) message).isInitialSyncDone());
 284                 context().parent().tell(message, self());
 285             } else {
 286                 super.onReceiveCommand(message);
 287             }
 288         } finally {
 289             context.done();
 290         }
 291     }
 292
 293     @Override
 294     protected Optional<ActorRef> getRoleChangeNotifier() {
 295         return roleChangeNotifier;
 296     }
 297
 298     private void onDatastoreContext(DatastoreContext context) {
 299         datastoreContext = context;
 300
 301         commitCoordinator.setQueueCapacity(datastoreContext.getShardTransactionCommitQueueCapacity());
 302
 303         setTransactionCommitTimeout();
 304
 305         if(datastoreContext.isPersistent() && !persistence().isRecoveryApplicable()) {
 306             setPersistence(true);
 307         } else if(!datastoreContext.isPersistent() && persistence().isRecoveryApplicable()) {
 308             setPersistence(false);
 309         }
 310
 311         updateConfigParams(datastoreContext.getShardRaftConfig());
 312     }
 313
 314     private void handleTransactionCommitTimeoutCheck() {
 315         CohortEntry cohortEntry = commitCoordinator.getCurrentCohortEntry();
 316         if(cohortEntry != null) {
 317             long elapsed = System.currentTimeMillis() - cohortEntry.getLastAccessTime();
 318             if(elapsed > transactionCommitTimeout) {
 319                 LOG.warn("{}: Current transaction {} has timed out after {} ms - aborting",
 320                         persistenceId(), cohortEntry.getTransactionID(), transactionCommitTimeout);
 321
 322                 doAbortTransaction(cohortEntry.getTransactionID(), null);
 323             }
 324         }
 325     }
 326
 327     private void handleCommitTransaction(final CommitTransaction commit) {
 328         final String transactionID = commit.getTransactionID();
 329
 330         LOG.debug("{}: Committing transaction {}", persistenceId(), transactionID);
 331
 332         // Get the current in-progress cohort entry in the commitCoordinator if it corresponds to
 333         // this transaction.
 334         final CohortEntry cohortEntry = commitCoordinator.getCohortEntryIfCurrent(transactionID);
 335         if(cohortEntry == null) {
 336             // We're not the current Tx - the Tx was likely expired b/c it took too long in
 337             // between the canCommit and commit messages.
 338             IllegalStateException ex = new IllegalStateException(
 339                     String.format("%s: Cannot commit transaction %s - it is not the current transaction",
 340                             persistenceId(), transactionID));
 341             LOG.error(ex.getMessage());
 342             shardMBean.incrementFailedTransactionsCount();
 343             getSender().tell(new akka.actor.Status.Failure(ex), getSelf());
 344             return;
 345         }
 346
 347         // We perform the preCommit phase here atomically with the commit phase. This is an
 348         // optimization to eliminate the overhead of an extra preCommit message. We lose front-end
 349         // coordination of preCommit across shards in case of failure but preCommit should not
 350         // normally fail since we ensure only one concurrent 3-phase commit.
 351
 352         try {
 353             // We block on the future here so we don't have to worry about possibly accessing our
 354             // state on a different thread outside of our dispatcher. Also, the data store
 355             // currently uses a same thread executor anyway.
 356             cohortEntry.getCohort().preCommit().get();
 357
 358             // If we do not have any followers and we are not using persistence
 359             // or if cohortEntry has no modifications
 360             // we can apply modification to the state immediately
 361             if((!hasFollowers() && !persistence().isRecoveryApplicable()) || (!cohortEntry.hasModifications())){
 362                 applyModificationToState(getSender(), transactionID, cohortEntry.getModification());
 363             } else {
 364                 Shard.this.persistData(getSender(), transactionID,
 365                         new ModificationPayload(cohortEntry.getModification()));
 366             }
 367         } catch (Exception e) {
 368             LOG.error("{} An exception occurred while preCommitting transaction {}",
 369                     persistenceId(), cohortEntry.getTransactionID(), e);
 370             shardMBean.incrementFailedTransactionsCount();
 371             getSender().tell(new akka.actor.Status.Failure(e), getSelf());
 372         }
 373
 374         cohortEntry.updateLastAccessTime();
 375     }
 376
 377     private void finishCommit(@Nonnull final ActorRef sender, final @Nonnull String transactionID) {
 378         // With persistence enabled, this method is called via applyState by the leader strategy
 379         // after the commit has been replicated to a majority of the followers.
 380
 381         CohortEntry cohortEntry = commitCoordinator.getCohortEntryIfCurrent(transactionID);
 382         if(cohortEntry == null) {
 383             // The transaction is no longer the current commit. This can happen if the transaction
 384             // was aborted prior, most likely due to timeout in the front-end. We need to finish
 385             // committing the transaction though since it was successfully persisted and replicated
 386             // however we can't use the original cohort b/c it was already preCommitted and may
 387             // conflict with the current commit or may have been aborted so we commit with a new
 388             // transaction.
 389             cohortEntry = commitCoordinator.getAndRemoveCohortEntry(transactionID);
 390             if(cohortEntry != null) {
 391                 commitWithNewTransaction(cohortEntry.getModification());
 392                 sender.tell(CommitTransactionReply.INSTANCE.toSerializable(), getSelf());
 393             } else {
 394                 // This really shouldn't happen - it likely means that persistence or replication
 395                 // took so long to complete such that the cohort entry was expired from the cache.
 396                 IllegalStateException ex = new IllegalStateException(
 397                         String.format("%s: Could not finish committing transaction %s - no CohortEntry found",
 398                                 persistenceId(), transactionID));
 399                 LOG.error(ex.getMessage());
 400                 sender.tell(new akka.actor.Status.Failure(ex), getSelf());
 401             }
 402
 403             return;
 404         }
 405
 406         LOG.debug("{}: Finishing commit for transaction {}", persistenceId(), cohortEntry.getTransactionID());
 407
 408         try {
 409             // We block on the future here so we don't have to worry about possibly accessing our
 410             // state on a different thread outside of our dispatcher. Also, the data store
 411             // currently uses a same thread executor anyway.
 412             cohortEntry.getCohort().commit().get();
 413
 414             sender.tell(CommitTransactionReply.INSTANCE.toSerializable(), getSelf());
 415
 416             shardMBean.incrementCommittedTransactionCount();
 417             shardMBean.setLastCommittedTransactionTime(System.currentTimeMillis());
 418
 419         } catch (Exception e) {
 420             sender.tell(new akka.actor.Status.Failure(e), getSelf());
 421
 422             LOG.error("{}, An exception occurred while committing transaction {}", persistenceId(),
 423                     transactionID, e);
 424             shardMBean.incrementFailedTransactionsCount();
 425         } finally {
 426             commitCoordinator.currentTransactionComplete(transactionID, true);
 427         }
 428     }
 429
 430     private void handleCanCommitTransaction(final CanCommitTransaction canCommit) {
 431         LOG.debug("{}: Can committing transaction {}", persistenceId(), canCommit.getTransactionID());
 432         commitCoordinator.handleCanCommit(canCommit, getSender(), self());
 433     }
 434
 435     private void handleBatchedModifications(BatchedModifications batched) {
 436         // This message is sent to prepare the modificationsa transaction directly on the Shard as an
 437         // optimization to avoid the extra overhead of a separate ShardTransaction actor. On the last
 438         // BatchedModifications message, the caller sets the ready flag in the message indicating
 439         // modifications are complete. The reply contains the cohort actor path (this actor) for the caller
 440         // to initiate the 3-phase commit. This also avoids the overhead of sending an additional
 441         // ReadyTransaction message.
 442
 443         // If we're not the leader then forward to the leader. This is a safety measure - we shouldn't
 444         // normally get here if we're not the leader as the front-end (TransactionProxy) should determine
 445         // the primary/leader shard. However with timing and caching on the front-end, there's a small
 446         // window where it could have a stale leader during leadership transitions.
 447         //
 448         if(isLeader()) {
 449             try {
 450                 boolean ready = commitCoordinator.handleTransactionModifications(batched);
 451                 if(ready) {
 452                     sender().tell(READY_TRANSACTION_REPLY, self());
 453                 } else {
 454                     sender().tell(new BatchedModificationsReply(batched.getModifications().size()), self());
 455                 }
 456             } catch (Exception e) {
 457                 LOG.error("{}: Error handling BatchedModifications for Tx {}", persistenceId(),
 458                         batched.getTransactionID(), e);
 459                 getSender().tell(new akka.actor.Status.Failure(e), getSelf());
 460             }
 461         } else {
 462             ActorSelection leader = getLeader();
 463             if(leader != null) {
 464                 // TODO: what if this is not the first batch and leadership changed in between batched messages?
 465                 // We could check if the commitCoordinator already has a cached entry and forward all the previous
 466                 // batched modifications.
 467                 LOG.debug("{}: Forwarding BatchedModifications to leader {}", persistenceId(), leader);
 468                 leader.forward(batched, getContext());
 469             } else {
 470                 // TODO: rather than throwing an immediate exception, we could schedule a timer to try again to make
 471                 // it more resilient in case we're in the process of electing a new leader.
 472                 getSender().tell(new akka.actor.Status.Failure(new NoShardLeaderException(String.format(
 473                     "Could not find the leader for shard %s. This typically happens" +
 474                     " when the system is coming up or recovering and a leader is being elected. Try again" +
 475                     " later.", persistenceId()))), getSelf());
 476             }
 477         }
 478     }
 479
 480     private void handleForwardedReadyTransaction(ForwardedReadyTransaction ready) {
 481         LOG.debug("{}: Readying transaction {}, client version {}", persistenceId(),
 482                 ready.getTransactionID(), ready.getTxnClientVersion());
 483
 484         // This message is forwarded by the ShardTransaction on ready. We cache the cohort in the
 485         // commitCoordinator in preparation for the subsequent three phase commit initiated by
 486         // the front-end.
 487         commitCoordinator.transactionReady(ready.getTransactionID(), ready.getCohort(),
 488                 (MutableCompositeModification) ready.getModification());
 489
 490         // Return our actor path as we'll handle the three phase commit, except if the Tx client
 491         // version < 1 (Helium-1 version). This means the Tx was initiated by a base Helium version
 492         // node. In that case, the subsequent 3-phase commit messages won't contain the
 493         // transactionId so to maintain backwards compatibility, we create a separate cohort actor
 494         // to provide the compatible behavior.
 495         if(ready.getTxnClientVersion() < DataStoreVersions.LITHIUM_VERSION) {
 496             ActorRef replyActorPath = getSelf();
 497             if(ready.getTxnClientVersion() < DataStoreVersions.HELIUM_1_VERSION) {
 498                 LOG.debug("{}: Creating BackwardsCompatibleThreePhaseCommitCohort", persistenceId());
 499                 replyActorPath = getContext().actorOf(BackwardsCompatibleThreePhaseCommitCohort.props(
 500                         ready.getTransactionID()));
 501             }
 502
 503             ReadyTransactionReply readyTransactionReply =
 504                     new ReadyTransactionReply(Serialization.serializedActorPath(replyActorPath),
 505                             ready.getTxnClientVersion());
 506             getSender().tell(ready.isReturnSerialized() ? readyTransactionReply.toSerializable() :
 507                 readyTransactionReply, getSelf());
 508         } else {
 509             getSender().tell(READY_TRANSACTION_REPLY, getSelf());
 510         }
 511     }
 512
 513     private void handleAbortTransaction(final AbortTransaction abort) {
 514         doAbortTransaction(abort.getTransactionID(), getSender());
 515     }
 516
 517     void doAbortTransaction(final String transactionID, final ActorRef sender) {
 518         final CohortEntry cohortEntry = commitCoordinator.getCohortEntryIfCurrent(transactionID);
 519         if(cohortEntry != null) {
 520             LOG.debug("{}: Aborting transaction {}", persistenceId(), transactionID);
 521
 522             // We don't remove the cached cohort entry here (ie pass false) in case the Tx was
 523             // aborted during replication in which case we may still commit locally if replication
 524             // succeeds.
 525             commitCoordinator.currentTransactionComplete(transactionID, false);
 526
 527             final ListenableFuture<Void> future = cohortEntry.getCohort().abort();
 528             final ActorRef self = getSelf();
 529
 530             Futures.addCallback(future, new FutureCallback<Void>() {
 531                 @Override
 532                 public void onSuccess(final Void v) {
 533                     shardMBean.incrementAbortTransactionsCount();
 534
 535                     if(sender != null) {
 536                         sender.tell(AbortTransactionReply.INSTANCE.toSerializable(), self);
 537                     }
 538                 }
 539
 540                 @Override
 541                 public void onFailure(final Throwable t) {
 542                     LOG.error("{}: An exception happened during abort", persistenceId(), t);
 543
 544                     if(sender != null) {
 545                         sender.tell(new akka.actor.Status.Failure(t), self);
 546                     }
 547                 }
 548             });
 549         }
 550     }
 551
 552     private void handleCreateTransaction(final Object message) {
 553         if (isLeader()) {
 554             createTransaction(CreateTransaction.fromSerializable(message));
 555         } else if (getLeader() != null) {
 556             getLeader().forward(message, getContext());
 557         } else {
 558             getSender().tell(new akka.actor.Status.Failure(new NoShardLeaderException(String.format(
 559                 "Could not find leader for shard %s so transaction cannot be created. This typically happens" +
 560                 " when the system is coming up or recovering and a leader is being elected. Try again" +
 561                 " later.", persistenceId()))), getSelf());
 562         }
 563     }
 564
 565     private void closeTransactionChain(final CloseTransactionChain closeTransactionChain) {
 566         transactionFactory.closeTransactionChain(closeTransactionChain.getTransactionChainId());
 567     }
 568
 569     private ActorRef createTypedTransactionActor(int transactionType,
 570             ShardTransactionIdentifier transactionId, String transactionChainId,
 571             short clientVersion ) {
 572
 573         DOMStoreTransaction transaction = transactionFactory.newTransaction(
 574                 TransactionProxy.TransactionType.fromInt(transactionType), transactionId.toString(),
 575                 transactionChainId);
 576
 577         return createShardTransaction(transaction, transactionId, clientVersion);
 578     }
 579
 580     private ActorRef createShardTransaction(DOMStoreTransaction transaction, ShardTransactionIdentifier transactionId,
 581                                             short clientVersion){
 582         return getContext().actorOf(
 583                 ShardTransaction.props(transaction, getSelf(),
 584                         schemaContext, datastoreContext, shardMBean,
 585                         transactionId.getRemoteTransactionId(), clientVersion)
 586                         .withDispatcher(txnDispatcherPath),
 587                 transactionId.toString());
 588
 589     }
 590
 591     private void createTransaction(CreateTransaction createTransaction) {
 592         try {
 593             ActorRef transactionActor = createTransaction(createTransaction.getTransactionType(),
 594                 createTransaction.getTransactionId(), createTransaction.getTransactionChainId(),
 595                 createTransaction.getVersion());
 596
 597             getSender().tell(new CreateTransactionReply(Serialization.serializedActorPath(transactionActor),
 598                     createTransaction.getTransactionId()).toSerializable(), getSelf());
 599         } catch (Exception e) {
 600             getSender().tell(new akka.actor.Status.Failure(e), getSelf());
 601         }
 602     }
 603
 604     private ActorRef createTransaction(int transactionType, String remoteTransactionId,
 605             String transactionChainId, short clientVersion) {
 606
 607
 608         ShardTransactionIdentifier transactionId = new ShardTransactionIdentifier(remoteTransactionId);
 609
 610         if(LOG.isDebugEnabled()) {
 611             LOG.debug("{}: Creating transaction : {} ", persistenceId(), transactionId);
 612         }
 613
 614         ActorRef transactionActor = createTypedTransactionActor(transactionType, transactionId,
 615                 transactionChainId, clientVersion);
 616
 617         return transactionActor;
 618     }
 619
 620     private void syncCommitTransaction(final DOMStoreWriteTransaction transaction)
 621         throws ExecutionException, InterruptedException {
 622         DOMStoreThreePhaseCommitCohort commitCohort = transaction.ready();
 623         commitCohort.preCommit().get();
 624         commitCohort.commit().get();
 625     }
 626
 627     private void commitWithNewTransaction(final Modification modification) {
 628         DOMStoreWriteTransaction tx = store.newWriteOnlyTransaction();
 629         modification.apply(tx);
 630         try {
 631             syncCommitTransaction(tx);
 632             shardMBean.incrementCommittedTransactionCount();
 633             shardMBean.setLastCommittedTransactionTime(System.currentTimeMillis());
 634         } catch (InterruptedException | ExecutionException e) {
 635             shardMBean.incrementFailedTransactionsCount();
 636             LOG.error("{}: Failed to commit", persistenceId(), e);
 637         }
 638     }
 639
 640     private void updateSchemaContext(final UpdateSchemaContext message) {
 641         this.schemaContext = message.getSchemaContext();
 642         updateSchemaContext(message.getSchemaContext());
 643         store.onGlobalContextUpdated(message.getSchemaContext());
 644     }
 645
 646     @VisibleForTesting
 647     void updateSchemaContext(final SchemaContext schemaContext) {
 648         store.onGlobalContextUpdated(schemaContext);
 649     }
 650
 651     private boolean isMetricsCaptureEnabled() {
 652         CommonConfig config = new CommonConfig(getContext().system().settings().config());
 653         return config.isMetricCaptureEnabled();
 654     }
 655
 656     @Override
 657     protected
 658     void startLogRecoveryBatch(final int maxBatchSize) {
 659         recoveryCoordinator.startLogRecoveryBatch(maxBatchSize);
 660     }
 661
 662     @Override
 663     protected void appendRecoveredLogEntry(final Payload data) {
 664         recoveryCoordinator.appendRecoveredLogPayload(data);
 665     }
 666
 667     @Override
 668     protected void applyRecoverySnapshot(final byte[] snapshotBytes) {
 669         recoveryCoordinator.applyRecoveredSnapshot(snapshotBytes);
 670     }
 671
 672     @Override
 673     protected void applyCurrentLogRecoveryBatch() {
 674         recoveryCoordinator.applyCurrentLogRecoveryBatch();
 675     }
 676
 677     @Override
 678     protected void onRecoveryComplete() {
 679         recoveryCoordinator = null;
 680
 681         //notify shard manager
 682         getContext().parent().tell(new ActorInitialized(), getSelf());
 683
 684         // Being paranoid here - this method should only be called once but just in case...
 685         if(txCommitTimeoutCheckSchedule == null) {
 686             // Schedule a message to be periodically sent to check if the current in-progress
 687             // transaction should be expired and aborted.
 688             FiniteDuration period = Duration.create(transactionCommitTimeout / 3, TimeUnit.MILLISECONDS);
 689             txCommitTimeoutCheckSchedule = getContext().system().scheduler().schedule(
 690                     period, period, getSelf(),
 691                     TX_COMMIT_TIMEOUT_CHECK_MESSAGE, getContext().dispatcher(), ActorRef.noSender());
 692         }
 693     }
 694
 695     @Override
 696     protected void applyState(final ActorRef clientActor, final String identifier, final Object data) {
 697
 698         if(data instanceof ModificationPayload) {
 699             try {
 700                 applyModificationToState(clientActor, identifier, ((ModificationPayload) data).getModification());
 701             } catch (ClassNotFoundException | IOException e) {
 702                 LOG.error("{}: Error extracting ModificationPayload", persistenceId(), e);
 703             }
 704         }
 705         else if (data instanceof CompositeModificationPayload) {
 706             Object modification = ((CompositeModificationPayload) data).getModification();
 707
 708             applyModificationToState(clientActor, identifier, modification);
 709         } else if(data instanceof CompositeModificationByteStringPayload ){
 710             Object modification = ((CompositeModificationByteStringPayload) data).getModification();
 711
 712             applyModificationToState(clientActor, identifier, modification);
 713         } else {
 714             LOG.error("{}: Unknown state received {} Class loader = {} CompositeNodeMod.ClassLoader = {}",
 715                     persistenceId(), data, data.getClass().getClassLoader(),
 716                     CompositeModificationPayload.class.getClassLoader());
 717         }
 718     }
 719
 720     private void applyModificationToState(ActorRef clientActor, String identifier, Object modification) {
 721         if(modification == null) {
 722             LOG.error(
 723                     "{}: modification is null - this is very unexpected, clientActor = {}, identifier = {}",
 724                     persistenceId(), identifier, clientActor != null ? clientActor.path().toString() : null);
 725         } else if(clientActor == null) {
 726             // There's no clientActor to which to send a commit reply so we must be applying
 727             // replicated state from the leader.
 728             commitWithNewTransaction(MutableCompositeModification.fromSerializable(modification));
 729         } else {
 730             // This must be the OK to commit after replication consensus.
 731             finishCommit(clientActor, identifier);
 732         }
 733     }
 734
 735     @Override
 736     protected void createSnapshot() {
 737         // Create a transaction actor. We are really going to treat the transaction as a worker
 738         // so that this actor does not get block building the snapshot. THe transaction actor will
 739         // after processing the CreateSnapshot message.
 740
 741         ActorRef createSnapshotTransaction = createTransaction(
 742                 TransactionProxy.TransactionType.READ_ONLY.ordinal(),
 743                 "createSnapshot" + ++createSnapshotTransactionCounter, "",
 744                 DataStoreVersions.CURRENT_VERSION);
 745
 746         createSnapshotTransaction.tell(CreateSnapshot.INSTANCE, self());
 747     }
 748
 749     @VisibleForTesting
 750     @Override
 751     protected void applySnapshot(final byte[] snapshotBytes) {
 752         // Since this will be done only on Recovery or when this actor is a Follower
 753         // we can safely commit everything in here. We not need to worry about event notifications
 754         // as they would have already been disabled on the follower
 755
 756         LOG.info("{}: Applying snapshot", persistenceId());
 757         try {
 758             DOMStoreWriteTransaction transaction = store.newWriteOnlyTransaction();
 759
 760             NormalizedNode<?, ?> node = SerializationUtils.deserializeNormalizedNode(snapshotBytes);
 761
 762             // delete everything first
 763             transaction.delete(DATASTORE_ROOT);
 764
 765             // Add everything from the remote node back
 766             transaction.write(DATASTORE_ROOT, node);
 767             syncCommitTransaction(transaction);
 768         } catch (InterruptedException | ExecutionException e) {
 769             LOG.error("{}: An exception occurred when applying snapshot", persistenceId(), e);
 770         } finally {
 771             LOG.info("{}: Done applying snapshot", persistenceId());
 772         }
 773     }
 774
 775     @Override
 776     protected void onStateChanged() {
 777         boolean isLeader = isLeader();
 778         changeSupport.onLeadershipChange(isLeader);
 779         treeChangeSupport.onLeadershipChange(isLeader);
 780
 781         // If this actor is no longer the leader close all the transaction chains
 782         if (!isLeader) {
 783             if(LOG.isDebugEnabled()) {
 784                 LOG.debug(
 785                     "{}: onStateChanged: Closing all transaction chains because shard {} is no longer the leader",
 786                     persistenceId(), getId());
 787             }
 788
 789             transactionFactory.closeAllTransactionChains();
 790         }
 791     }
 792
 793     @Override
 794     public String persistenceId() {
 795         return this.name;
 796     }
 797
 798     @VisibleForTesting
 799     ShardCommitCoordinator getCommitCoordinator() {
 800         return commitCoordinator;
 801     }
 802
 803
 804     private static class ShardCreator implements Creator<Shard> {
 805
 806         private static final long serialVersionUID = 1L;
 807
 808         final ShardIdentifier name;
 809         final Map<String, String> peerAddresses;
 810         final DatastoreContext datastoreContext;
 811         final SchemaContext schemaContext;
 812
 813         ShardCreator(final ShardIdentifier name, final Map<String, String> peerAddresses,
 814                 final DatastoreContext datastoreContext, final SchemaContext schemaContext) {
 815             this.name = name;
 816             this.peerAddresses = peerAddresses;
 817             this.datastoreContext = datastoreContext;
 818             this.schemaContext = schemaContext;
 819         }
 820
 821         @Override
 822         public Shard create() throws Exception {
 823             return new Shard(name, peerAddresses, datastoreContext, schemaContext);
 824         }
 825     }
 826
 827     @VisibleForTesting
 828     public InMemoryDOMDataStore getDataStore() {
 829         return store;
 830     }
 831
 832     @VisibleForTesting
 833     ShardStats getShardMBean() {
 834         return shardMBean;
 835     }
 836 }