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.collect.Lists;
  22 import com.google.common.util.concurrent.FutureCallback;
  23 import com.google.common.util.concurrent.Futures;
  24 import com.google.common.util.concurrent.ListenableFuture;
  25 import java.io.IOException;
  26 import java.util.HashMap;
  27 import java.util.List;
  28 import java.util.Map;
  29 import java.util.concurrent.ExecutionException;
  30 import java.util.concurrent.TimeUnit;
  31 import javax.annotation.Nonnull;
  32 import org.opendaylight.controller.cluster.common.actor.CommonConfig;
  33 import org.opendaylight.controller.cluster.common.actor.MeteringBehavior;
  34 import org.opendaylight.controller.cluster.datastore.ShardCommitCoordinator.CohortEntry;
  35 import org.opendaylight.controller.cluster.datastore.compat.BackwardsCompatibleThreePhaseCommitCohort;
  36 import org.opendaylight.controller.cluster.datastore.exceptions.NoShardLeaderException;
  37 import org.opendaylight.controller.cluster.datastore.identifiers.ShardIdentifier;
  38 import org.opendaylight.controller.cluster.datastore.identifiers.ShardTransactionIdentifier;
  39 import org.opendaylight.controller.cluster.datastore.jmx.mbeans.shard.ShardMBeanFactory;
  40 import org.opendaylight.controller.cluster.datastore.jmx.mbeans.shard.ShardStats;
  41 import org.opendaylight.controller.cluster.datastore.messages.AbortTransaction;
  42 import org.opendaylight.controller.cluster.datastore.messages.AbortTransactionReply;
  43 import org.opendaylight.controller.cluster.datastore.messages.ActorInitialized;
  44 import org.opendaylight.controller.cluster.datastore.messages.BatchedModifications;
  45 import org.opendaylight.controller.cluster.datastore.messages.BatchedModificationsReply;
  46 import org.opendaylight.controller.cluster.datastore.messages.CanCommitTransaction;
  47 import org.opendaylight.controller.cluster.datastore.messages.CloseTransactionChain;
  48 import org.opendaylight.controller.cluster.datastore.messages.CommitTransaction;
  49 import org.opendaylight.controller.cluster.datastore.messages.CommitTransactionReply;
  50 import org.opendaylight.controller.cluster.datastore.messages.CreateSnapshot;
  51 import org.opendaylight.controller.cluster.datastore.messages.CreateTransaction;
  52 import org.opendaylight.controller.cluster.datastore.messages.CreateTransactionReply;
  53 import org.opendaylight.controller.cluster.datastore.messages.EnableNotification;
  54 import org.opendaylight.controller.cluster.datastore.messages.ForwardedReadyTransaction;
  55 import org.opendaylight.controller.cluster.datastore.messages.PeerAddressResolved;
  56 import org.opendaylight.controller.cluster.datastore.messages.ReadyTransactionReply;
  57 import org.opendaylight.controller.cluster.datastore.messages.RegisterChangeListener;
  58 import org.opendaylight.controller.cluster.datastore.messages.RegisterChangeListenerReply;
  59 import org.opendaylight.controller.cluster.datastore.messages.UpdateSchemaContext;
  60 import org.opendaylight.controller.cluster.datastore.modification.Modification;
  61 import org.opendaylight.controller.cluster.datastore.modification.ModificationPayload;
  62 import org.opendaylight.controller.cluster.datastore.modification.MutableCompositeModification;
  63 import org.opendaylight.controller.cluster.datastore.utils.Dispatchers;
  64 import org.opendaylight.controller.cluster.datastore.utils.MessageTracker;
  65 import org.opendaylight.controller.cluster.datastore.utils.SerializationUtils;
  66 import org.opendaylight.controller.cluster.notifications.RegisterRoleChangeListener;
  67 import org.opendaylight.controller.cluster.notifications.RoleChangeNotifier;
  68 import org.opendaylight.controller.cluster.raft.RaftActor;
  69 import org.opendaylight.controller.cluster.raft.base.messages.FollowerInitialSyncUpStatus;
  70 import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
  71 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.CompositeModificationByteStringPayload;
  72 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.CompositeModificationPayload;
  73 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload;
  74 import org.opendaylight.controller.md.sal.common.api.data.AsyncDataChangeListener;
  75 import org.opendaylight.controller.md.sal.dom.store.impl.InMemoryDOMDataStore;
  76 import org.opendaylight.controller.md.sal.dom.store.impl.InMemoryDOMDataStoreFactory;
  77 import org.opendaylight.controller.sal.core.spi.data.DOMStoreThreePhaseCommitCohort;
  78 import org.opendaylight.controller.sal.core.spi.data.DOMStoreTransaction;
  79 import org.opendaylight.controller.sal.core.spi.data.DOMStoreWriteTransaction;
  80 import org.opendaylight.yangtools.concepts.ListenerRegistration;
  81 import org.opendaylight.yangtools.yang.data.api.YangInstanceIdentifier;
  82 import org.opendaylight.yangtools.yang.data.api.schema.NormalizedNode;
  83 import org.opendaylight.yangtools.yang.model.api.SchemaContext;
  84 import scala.concurrent.duration.Duration;
  85 import scala.concurrent.duration.FiniteDuration;
  86
  87 /**
  88  * A Shard represents a portion of the logical data tree <br/>
  89  * <p>
  90  * Our Shard uses InMemoryDataStore as it's internal representation and delegates all requests it
  91  * </p>
  92  */
  93 public class Shard extends RaftActor {
  94
  95     private static final YangInstanceIdentifier DATASTORE_ROOT = YangInstanceIdentifier.builder().build();
  96
  97     private static final Object TX_COMMIT_TIMEOUT_CHECK_MESSAGE = "txCommitTimeoutCheck";
  98
  99     @VisibleForTesting
 100     static final String DEFAULT_NAME = "default";
 101
 102     // The state of this Shard
 103     private final InMemoryDOMDataStore store;
 104
 105     /// The name of this shard
 106     private final String name;
 107
 108     private final ShardStats shardMBean;
 109
 110     private final List<ActorSelection> dataChangeListeners =  Lists.newArrayList();
 111
 112     private final List<DelayedListenerRegistration> delayedListenerRegistrations =
 113                                                                        Lists.newArrayList();
 114
 115     private DatastoreContext datastoreContext;
 116
 117     private SchemaContext schemaContext;
 118
 119     private int createSnapshotTransactionCounter;
 120
 121     private final ShardCommitCoordinator commitCoordinator;
 122
 123     private long transactionCommitTimeout;
 124
 125     private Cancellable txCommitTimeoutCheckSchedule;
 126
 127     private final Optional<ActorRef> roleChangeNotifier;
 128
 129     private final MessageTracker appendEntriesReplyTracker;
 130
 131     private final ReadyTransactionReply READY_TRANSACTION_REPLY = new ReadyTransactionReply(
 132             Serialization.serializedActorPath(getSelf()));
 133
 134
 135     /**
 136      * Coordinates persistence recovery on startup.
 137      */
 138     private ShardRecoveryCoordinator recoveryCoordinator;
 139
 140     private final DOMTransactionFactory transactionFactory;
 141
 142     private final String txnDispatcherPath;
 143
 144     protected Shard(final ShardIdentifier name, final Map<String, String> peerAddresses,
 145             final DatastoreContext datastoreContext, final SchemaContext schemaContext) {
 146         super(name.toString(), new HashMap<>(peerAddresses), Optional.of(datastoreContext.getShardRaftConfig()));
 147
 148         this.name = name.toString();
 149         this.datastoreContext = datastoreContext;
 150         this.schemaContext = schemaContext;
 151         this.txnDispatcherPath = new Dispatchers(context().system().dispatchers())
 152                 .getDispatcherPath(Dispatchers.DispatcherType.Transaction);
 153
 154         setPersistence(datastoreContext.isPersistent());
 155
 156         LOG.info("Shard created : {}, persistent : {}", name, datastoreContext.isPersistent());
 157
 158         store = InMemoryDOMDataStoreFactory.create(name.toString(), null,
 159                 datastoreContext.getDataStoreProperties());
 160
 161         if(schemaContext != null) {
 162             store.onGlobalContextUpdated(schemaContext);
 163         }
 164
 165         shardMBean = ShardMBeanFactory.getShardStatsMBean(name.toString(),
 166                 datastoreContext.getDataStoreMXBeanType());
 167         shardMBean.setNotificationManager(store.getDataChangeListenerNotificationManager());
 168         shardMBean.setShardActor(getSelf());
 169
 170         if (isMetricsCaptureEnabled()) {
 171             getContext().become(new MeteringBehavior(this));
 172         }
 173
 174         transactionFactory = new DOMTransactionFactory(store, shardMBean, LOG, this.name);
 175
 176         commitCoordinator = new ShardCommitCoordinator(transactionFactory,
 177                 TimeUnit.SECONDS.convert(5, TimeUnit.MINUTES),
 178                 datastoreContext.getShardTransactionCommitQueueCapacity(), self(), LOG, this.name);
 179
 180         setTransactionCommitTimeout();
 181
 182         // create a notifier actor for each cluster member
 183         roleChangeNotifier = createRoleChangeNotifier(name.toString());
 184
 185         appendEntriesReplyTracker = new MessageTracker(AppendEntriesReply.class,
 186                 getRaftActorContext().getConfigParams().getIsolatedCheckIntervalInMillis());
 187
 188         recoveryCoordinator = new ShardRecoveryCoordinator(store, persistenceId(), LOG);
 189     }
 190
 191     private void setTransactionCommitTimeout() {
 192         transactionCommitTimeout = TimeUnit.MILLISECONDS.convert(
 193                 datastoreContext.getShardTransactionCommitTimeoutInSeconds(), TimeUnit.SECONDS);
 194     }
 195
 196     public static Props props(final ShardIdentifier name,
 197         final Map<String, String> peerAddresses,
 198         final DatastoreContext datastoreContext, final SchemaContext schemaContext) {
 199         Preconditions.checkNotNull(name, "name should not be null");
 200         Preconditions.checkNotNull(peerAddresses, "peerAddresses should not be null");
 201         Preconditions.checkNotNull(datastoreContext, "dataStoreContext should not be null");
 202         Preconditions.checkNotNull(schemaContext, "schemaContext should not be null");
 203
 204         return Props.create(new ShardCreator(name, peerAddresses, datastoreContext, schemaContext));
 205     }
 206
 207     private Optional<ActorRef> createRoleChangeNotifier(String shardId) {
 208         ActorRef shardRoleChangeNotifier = this.getContext().actorOf(
 209             RoleChangeNotifier.getProps(shardId), shardId + "-notifier");
 210         return Optional.of(shardRoleChangeNotifier);
 211     }
 212
 213     @Override
 214     public void postStop() {
 215         LOG.info("Stopping Shard {}", persistenceId());
 216
 217         super.postStop();
 218
 219         if(txCommitTimeoutCheckSchedule != null) {
 220             txCommitTimeoutCheckSchedule.cancel();
 221         }
 222
 223         shardMBean.unregisterMBean();
 224     }
 225
 226     @Override
 227     public void onReceiveRecover(final Object message) throws Exception {
 228         if(LOG.isDebugEnabled()) {
 229             LOG.debug("{}: onReceiveRecover: Received message {} from {}", persistenceId(),
 230                 message.getClass().toString(), getSender());
 231         }
 232
 233         if (message instanceof RecoveryFailure){
 234             LOG.error("{}: Recovery failed because of this cause",
 235                     persistenceId(), ((RecoveryFailure) message).cause());
 236
 237             // Even though recovery failed, we still need to finish our recovery, eg send the
 238             // ActorInitialized message and start the txCommitTimeoutCheckSchedule.
 239             onRecoveryComplete();
 240         } else {
 241             super.onReceiveRecover(message);
 242             if(LOG.isTraceEnabled()) {
 243                 appendEntriesReplyTracker.begin();
 244             }
 245         }
 246     }
 247
 248     @Override
 249     public void onReceiveCommand(final Object message) throws Exception {
 250
 251         MessageTracker.Context context = appendEntriesReplyTracker.received(message);
 252
 253         if(context.error().isPresent()){
 254             LOG.trace("{} : AppendEntriesReply failed to arrive at the expected interval {}", persistenceId(),
 255                     context.error());
 256         }
 257
 258         try {
 259             if (CreateTransaction.SERIALIZABLE_CLASS.isInstance(message)) {
 260                 handleCreateTransaction(message);
 261             } else if (BatchedModifications.class.isInstance(message)) {
 262                 handleBatchedModifications((BatchedModifications)message);
 263             } else if (message instanceof ForwardedReadyTransaction) {
 264                 handleForwardedReadyTransaction((ForwardedReadyTransaction) message);
 265             } else if (CanCommitTransaction.SERIALIZABLE_CLASS.isInstance(message)) {
 266                 handleCanCommitTransaction(CanCommitTransaction.fromSerializable(message));
 267             } else if (CommitTransaction.SERIALIZABLE_CLASS.isInstance(message)) {
 268                 handleCommitTransaction(CommitTransaction.fromSerializable(message));
 269             } else if (AbortTransaction.SERIALIZABLE_CLASS.isInstance(message)) {
 270                 handleAbortTransaction(AbortTransaction.fromSerializable(message));
 271             } else if (CloseTransactionChain.SERIALIZABLE_CLASS.isInstance(message)) {
 272                 closeTransactionChain(CloseTransactionChain.fromSerializable(message));
 273             } else if (message instanceof RegisterChangeListener) {
 274                 registerChangeListener((RegisterChangeListener) message);
 275             } else if (message instanceof UpdateSchemaContext) {
 276                 updateSchemaContext((UpdateSchemaContext) message);
 277             } else if (message instanceof PeerAddressResolved) {
 278                 PeerAddressResolved resolved = (PeerAddressResolved) message;
 279                 setPeerAddress(resolved.getPeerId().toString(),
 280                         resolved.getPeerAddress());
 281             } else if (message.equals(TX_COMMIT_TIMEOUT_CHECK_MESSAGE)) {
 282                 handleTransactionCommitTimeoutCheck();
 283             } else if(message instanceof DatastoreContext) {
 284                 onDatastoreContext((DatastoreContext)message);
 285             } else if(message instanceof RegisterRoleChangeListener){
 286                 roleChangeNotifier.get().forward(message, context());
 287             } else if (message instanceof FollowerInitialSyncUpStatus){
 288                 shardMBean.setFollowerInitialSyncStatus(((FollowerInitialSyncUpStatus) message).isInitialSyncDone());
 289                 context().parent().tell(message, self());
 290             } else {
 291                 super.onReceiveCommand(message);
 292             }
 293         } finally {
 294             context.done();
 295         }
 296     }
 297
 298     @Override
 299     protected Optional<ActorRef> getRoleChangeNotifier() {
 300         return roleChangeNotifier;
 301     }
 302
 303     private void onDatastoreContext(DatastoreContext context) {
 304         datastoreContext = context;
 305
 306         commitCoordinator.setQueueCapacity(datastoreContext.getShardTransactionCommitQueueCapacity());
 307
 308         setTransactionCommitTimeout();
 309
 310         if(datastoreContext.isPersistent() && !persistence().isRecoveryApplicable()) {
 311             setPersistence(true);
 312         } else if(!datastoreContext.isPersistent() && persistence().isRecoveryApplicable()) {
 313             setPersistence(false);
 314         }
 315
 316         updateConfigParams(datastoreContext.getShardRaftConfig());
 317     }
 318
 319     private void handleTransactionCommitTimeoutCheck() {
 320         CohortEntry cohortEntry = commitCoordinator.getCurrentCohortEntry();
 321         if(cohortEntry != null) {
 322             long elapsed = System.currentTimeMillis() - cohortEntry.getLastAccessTime();
 323             if(elapsed > transactionCommitTimeout) {
 324                 LOG.warn("{}: Current transaction {} has timed out after {} ms - aborting",
 325                         persistenceId(), cohortEntry.getTransactionID(), transactionCommitTimeout);
 326
 327                 doAbortTransaction(cohortEntry.getTransactionID(), null);
 328             }
 329         }
 330     }
 331
 332     private void handleCommitTransaction(final CommitTransaction commit) {
 333         final String transactionID = commit.getTransactionID();
 334
 335         LOG.debug("{}: Committing transaction {}", persistenceId(), transactionID);
 336
 337         // Get the current in-progress cohort entry in the commitCoordinator if it corresponds to
 338         // this transaction.
 339         final CohortEntry cohortEntry = commitCoordinator.getCohortEntryIfCurrent(transactionID);
 340         if(cohortEntry == null) {
 341             // We're not the current Tx - the Tx was likely expired b/c it took too long in
 342             // between the canCommit and commit messages.
 343             IllegalStateException ex = new IllegalStateException(
 344                     String.format("%s: Cannot commit transaction %s - it is not the current transaction",
 345                             persistenceId(), transactionID));
 346             LOG.error(ex.getMessage());
 347             shardMBean.incrementFailedTransactionsCount();
 348             getSender().tell(new akka.actor.Status.Failure(ex), getSelf());
 349             return;
 350         }
 351
 352         // We perform the preCommit phase here atomically with the commit phase. This is an
 353         // optimization to eliminate the overhead of an extra preCommit message. We lose front-end
 354         // coordination of preCommit across shards in case of failure but preCommit should not
 355         // normally fail since we ensure only one concurrent 3-phase commit.
 356
 357         try {
 358             // We block on the future here so we don't have to worry about possibly accessing our
 359             // state on a different thread outside of our dispatcher. Also, the data store
 360             // currently uses a same thread executor anyway.
 361             cohortEntry.getCohort().preCommit().get();
 362
 363             // If we do not have any followers and we are not using persistence
 364             // or if cohortEntry has no modifications
 365             // we can apply modification to the state immediately
 366             if((!hasFollowers() && !persistence().isRecoveryApplicable()) || (!cohortEntry.hasModifications())){
 367                 applyModificationToState(getSender(), transactionID, cohortEntry.getModification());
 368             } else {
 369                 Shard.this.persistData(getSender(), transactionID,
 370                         new ModificationPayload(cohortEntry.getModification()));
 371             }
 372         } catch (Exception e) {
 373             LOG.error("{} An exception occurred while preCommitting transaction {}",
 374                     persistenceId(), cohortEntry.getTransactionID(), e);
 375             shardMBean.incrementFailedTransactionsCount();
 376             getSender().tell(new akka.actor.Status.Failure(e), getSelf());
 377         }
 378
 379         cohortEntry.updateLastAccessTime();
 380     }
 381
 382     private void finishCommit(@Nonnull final ActorRef sender, final @Nonnull String transactionID) {
 383         // With persistence enabled, this method is called via applyState by the leader strategy
 384         // after the commit has been replicated to a majority of the followers.
 385
 386         CohortEntry cohortEntry = commitCoordinator.getCohortEntryIfCurrent(transactionID);
 387         if(cohortEntry == null) {
 388             // The transaction is no longer the current commit. This can happen if the transaction
 389             // was aborted prior, most likely due to timeout in the front-end. We need to finish
 390             // committing the transaction though since it was successfully persisted and replicated
 391             // however we can't use the original cohort b/c it was already preCommitted and may
 392             // conflict with the current commit or may have been aborted so we commit with a new
 393             // transaction.
 394             cohortEntry = commitCoordinator.getAndRemoveCohortEntry(transactionID);
 395             if(cohortEntry != null) {
 396                 commitWithNewTransaction(cohortEntry.getModification());
 397                 sender.tell(CommitTransactionReply.INSTANCE.toSerializable(), getSelf());
 398             } else {
 399                 // This really shouldn't happen - it likely means that persistence or replication
 400                 // took so long to complete such that the cohort entry was expired from the cache.
 401                 IllegalStateException ex = new IllegalStateException(
 402                         String.format("%s: Could not finish committing transaction %s - no CohortEntry found",
 403                                 persistenceId(), transactionID));
 404                 LOG.error(ex.getMessage());
 405                 sender.tell(new akka.actor.Status.Failure(ex), getSelf());
 406             }
 407
 408             return;
 409         }
 410
 411         LOG.debug("{}: Finishing commit for transaction {}", persistenceId(), cohortEntry.getTransactionID());
 412
 413         try {
 414             // We block on the future here so we don't have to worry about possibly accessing our
 415             // state on a different thread outside of our dispatcher. Also, the data store
 416             // currently uses a same thread executor anyway.
 417             cohortEntry.getCohort().commit().get();
 418
 419             sender.tell(CommitTransactionReply.INSTANCE.toSerializable(), getSelf());
 420
 421             shardMBean.incrementCommittedTransactionCount();
 422             shardMBean.setLastCommittedTransactionTime(System.currentTimeMillis());
 423
 424         } catch (Exception e) {
 425             sender.tell(new akka.actor.Status.Failure(e), getSelf());
 426
 427             LOG.error("{}, An exception occurred while committing transaction {}", persistenceId(),
 428                     transactionID, e);
 429             shardMBean.incrementFailedTransactionsCount();
 430         } finally {
 431             commitCoordinator.currentTransactionComplete(transactionID, true);
 432         }
 433     }
 434
 435     private void handleCanCommitTransaction(final CanCommitTransaction canCommit) {
 436         LOG.debug("{}: Can committing transaction {}", persistenceId(), canCommit.getTransactionID());
 437         commitCoordinator.handleCanCommit(canCommit, getSender(), self());
 438     }
 439
 440     private void handleBatchedModifications(BatchedModifications batched) {
 441         // This message is sent to prepare the modificationsa transaction directly on the Shard as an
 442         // optimization to avoid the extra overhead of a separate ShardTransaction actor. On the last
 443         // BatchedModifications message, the caller sets the ready flag in the message indicating
 444         // modifications are complete. The reply contains the cohort actor path (this actor) for the caller
 445         // to initiate the 3-phase commit. This also avoids the overhead of sending an additional
 446         // ReadyTransaction message.
 447
 448         // If we're not the leader then forward to the leader. This is a safety measure - we shouldn't
 449         // normally get here if we're not the leader as the front-end (TransactionProxy) should determine
 450         // the primary/leader shard. However with timing and caching on the front-end, there's a small
 451         // window where it could have a stale leader during leadership transitions.
 452         //
 453         if(isLeader()) {
 454             try {
 455                 BatchedModificationsReply reply = commitCoordinator.handleTransactionModifications(batched);
 456                 sender().tell(reply, self());
 457             } catch (Exception e) {
 458                 LOG.error("{}: Error handling BatchedModifications for Tx {}", persistenceId(),
 459                         batched.getTransactionID(), e);
 460                 getSender().tell(new akka.actor.Status.Failure(e), getSelf());
 461             }
 462         } else {
 463             ActorSelection leader = getLeader();
 464             if(leader != null) {
 465                 // TODO: what if this is not the first batch and leadership changed in between batched messages?
 466                 // We could check if the commitCoordinator already has a cached entry and forward all the previous
 467                 // batched modifications.
 468                 LOG.debug("{}: Forwarding BatchedModifications to leader {}", persistenceId(), leader);
 469                 leader.forward(batched, getContext());
 470             } else {
 471                 // TODO: rather than throwing an immediate exception, we could schedule a timer to try again to make
 472                 // it more resilient in case we're in the process of electing a new leader.
 473                 getSender().tell(new akka.actor.Status.Failure(new NoShardLeaderException(String.format(
 474                     "Could not find the leader for shard %s. This typically happens" +
 475                     " when the system is coming up or recovering and a leader is being elected. Try again" +
 476                     " later.", persistenceId()))), getSelf());
 477             }
 478         }
 479     }
 480
 481     private void handleForwardedReadyTransaction(ForwardedReadyTransaction ready) {
 482         LOG.debug("{}: Readying transaction {}, client version {}", persistenceId(),
 483                 ready.getTransactionID(), ready.getTxnClientVersion());
 484
 485         // This message is forwarded by the ShardTransaction on ready. We cache the cohort in the
 486         // commitCoordinator in preparation for the subsequent three phase commit initiated by
 487         // the front-end.
 488         commitCoordinator.transactionReady(ready.getTransactionID(), ready.getCohort(),
 489                 (MutableCompositeModification) ready.getModification());
 490
 491         // Return our actor path as we'll handle the three phase commit, except if the Tx client
 492         // version < 1 (Helium-1 version). This means the Tx was initiated by a base Helium version
 493         // node. In that case, the subsequent 3-phase commit messages won't contain the
 494         // transactionId so to maintain backwards compatibility, we create a separate cohort actor
 495         // to provide the compatible behavior.
 496         if(ready.getTxnClientVersion() < DataStoreVersions.HELIUM_1_VERSION) {
 497             LOG.debug("{}: Creating BackwardsCompatibleThreePhaseCommitCohort", persistenceId());
 498             ActorRef replyActorPath = getContext().actorOf(BackwardsCompatibleThreePhaseCommitCohort.props(
 499                     ready.getTransactionID()));
 500
 501             ReadyTransactionReply readyTransactionReply =
 502                     new ReadyTransactionReply(Serialization.serializedActorPath(replyActorPath));
 503             getSender().tell(ready.isReturnSerialized() ? readyTransactionReply.toSerializable() :
 504                     readyTransactionReply, getSelf());
 505
 506         } else {
 507
 508             getSender().tell(ready.isReturnSerialized() ? READY_TRANSACTION_REPLY.toSerializable() :
 509                     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 void registerChangeListener(final RegisterChangeListener registerChangeListener) {
 652
 653         LOG.debug("{}: registerDataChangeListener for {}", persistenceId(), registerChangeListener.getPath());
 654
 655         ListenerRegistration<AsyncDataChangeListener<YangInstanceIdentifier,
 656                                                      NormalizedNode<?, ?>>> registration;
 657         if(isLeader()) {
 658             registration = doChangeListenerRegistration(registerChangeListener);
 659         } else {
 660             LOG.debug("{}: Shard is not the leader - delaying registration", persistenceId());
 661
 662             DelayedListenerRegistration delayedReg =
 663                     new DelayedListenerRegistration(registerChangeListener);
 664             delayedListenerRegistrations.add(delayedReg);
 665             registration = delayedReg;
 666         }
 667
 668         ActorRef listenerRegistration = getContext().actorOf(
 669                 DataChangeListenerRegistration.props(registration));
 670
 671         LOG.debug("{}: registerDataChangeListener sending reply, listenerRegistrationPath = {} ",
 672                 persistenceId(), listenerRegistration.path());
 673
 674         getSender().tell(new RegisterChangeListenerReply(listenerRegistration), getSelf());
 675     }
 676
 677     private ListenerRegistration<AsyncDataChangeListener<YangInstanceIdentifier,
 678                                                NormalizedNode<?, ?>>> doChangeListenerRegistration(
 679             final RegisterChangeListener registerChangeListener) {
 680
 681         ActorSelection dataChangeListenerPath = getContext().system().actorSelection(
 682                 registerChangeListener.getDataChangeListenerPath());
 683
 684         // Notify the listener if notifications should be enabled or not
 685         // If this shard is the leader then it will enable notifications else
 686         // it will not
 687         dataChangeListenerPath.tell(new EnableNotification(true), getSelf());
 688
 689         // Now store a reference to the data change listener so it can be notified
 690         // at a later point if notifications should be enabled or disabled
 691         dataChangeListeners.add(dataChangeListenerPath);
 692
 693         AsyncDataChangeListener<YangInstanceIdentifier, NormalizedNode<?, ?>> listener =
 694                 new DataChangeListenerProxy(dataChangeListenerPath);
 695
 696         LOG.debug("{}: Registering for path {}", persistenceId(), registerChangeListener.getPath());
 697
 698         return store.registerChangeListener(registerChangeListener.getPath(), listener,
 699                 registerChangeListener.getScope());
 700     }
 701
 702     private boolean isMetricsCaptureEnabled(){
 703         CommonConfig config = new CommonConfig(getContext().system().settings().config());
 704         return config.isMetricCaptureEnabled();
 705     }
 706
 707     @Override
 708     protected
 709     void startLogRecoveryBatch(final int maxBatchSize) {
 710         recoveryCoordinator.startLogRecoveryBatch(maxBatchSize);
 711     }
 712
 713     @Override
 714     protected void appendRecoveredLogEntry(final Payload data) {
 715         recoveryCoordinator.appendRecoveredLogPayload(data);
 716     }
 717
 718     @Override
 719     protected void applyRecoverySnapshot(final byte[] snapshotBytes) {
 720         recoveryCoordinator.applyRecoveredSnapshot(snapshotBytes);
 721     }
 722
 723     @Override
 724     protected void applyCurrentLogRecoveryBatch() {
 725         recoveryCoordinator.applyCurrentLogRecoveryBatch();
 726     }
 727
 728     @Override
 729     protected void onRecoveryComplete() {
 730         recoveryCoordinator = null;
 731
 732         //notify shard manager
 733         getContext().parent().tell(new ActorInitialized(), getSelf());
 734
 735         // Being paranoid here - this method should only be called once but just in case...
 736         if(txCommitTimeoutCheckSchedule == null) {
 737             // Schedule a message to be periodically sent to check if the current in-progress
 738             // transaction should be expired and aborted.
 739             FiniteDuration period = Duration.create(transactionCommitTimeout / 3, TimeUnit.MILLISECONDS);
 740             txCommitTimeoutCheckSchedule = getContext().system().scheduler().schedule(
 741                     period, period, getSelf(),
 742                     TX_COMMIT_TIMEOUT_CHECK_MESSAGE, getContext().dispatcher(), ActorRef.noSender());
 743         }
 744     }
 745
 746     @Override
 747     protected void applyState(final ActorRef clientActor, final String identifier, final Object data) {
 748
 749         if(data instanceof ModificationPayload) {
 750             try {
 751                 applyModificationToState(clientActor, identifier, ((ModificationPayload) data).getModification());
 752             } catch (ClassNotFoundException | IOException e) {
 753                 LOG.error("{}: Error extracting ModificationPayload", persistenceId(), e);
 754             }
 755         }
 756         else if (data instanceof CompositeModificationPayload) {
 757             Object modification = ((CompositeModificationPayload) data).getModification();
 758
 759             applyModificationToState(clientActor, identifier, modification);
 760         } else if(data instanceof CompositeModificationByteStringPayload ){
 761             Object modification = ((CompositeModificationByteStringPayload) data).getModification();
 762
 763             applyModificationToState(clientActor, identifier, modification);
 764         } else {
 765             LOG.error("{}: Unknown state received {} Class loader = {} CompositeNodeMod.ClassLoader = {}",
 766                     persistenceId(), data, data.getClass().getClassLoader(),
 767                     CompositeModificationPayload.class.getClassLoader());
 768         }
 769     }
 770
 771     private void applyModificationToState(ActorRef clientActor, String identifier, Object modification) {
 772         if(modification == null) {
 773             LOG.error(
 774                     "{}: modification is null - this is very unexpected, clientActor = {}, identifier = {}",
 775                     persistenceId(), identifier, clientActor != null ? clientActor.path().toString() : null);
 776         } else if(clientActor == null) {
 777             // There's no clientActor to which to send a commit reply so we must be applying
 778             // replicated state from the leader.
 779             commitWithNewTransaction(MutableCompositeModification.fromSerializable(modification));
 780         } else {
 781             // This must be the OK to commit after replication consensus.
 782             finishCommit(clientActor, identifier);
 783         }
 784     }
 785
 786     @Override
 787     protected void createSnapshot() {
 788         // Create a transaction actor. We are really going to treat the transaction as a worker
 789         // so that this actor does not get block building the snapshot. THe transaction actor will
 790         // after processing the CreateSnapshot message.
 791
 792         ActorRef createSnapshotTransaction = createTransaction(
 793                 TransactionProxy.TransactionType.READ_ONLY.ordinal(),
 794                 "createSnapshot" + ++createSnapshotTransactionCounter, "",
 795                 DataStoreVersions.CURRENT_VERSION);
 796
 797         createSnapshotTransaction.tell(CreateSnapshot.INSTANCE, self());
 798     }
 799
 800     @VisibleForTesting
 801     @Override
 802     protected void applySnapshot(final byte[] snapshotBytes) {
 803         // Since this will be done only on Recovery or when this actor is a Follower
 804         // we can safely commit everything in here. We not need to worry about event notifications
 805         // as they would have already been disabled on the follower
 806
 807         LOG.info("{}: Applying snapshot", persistenceId());
 808         try {
 809             DOMStoreWriteTransaction transaction = store.newWriteOnlyTransaction();
 810
 811             NormalizedNode<?, ?> node = SerializationUtils.deserializeNormalizedNode(snapshotBytes);
 812
 813             // delete everything first
 814             transaction.delete(DATASTORE_ROOT);
 815
 816             // Add everything from the remote node back
 817             transaction.write(DATASTORE_ROOT, node);
 818             syncCommitTransaction(transaction);
 819         } catch (InterruptedException | ExecutionException e) {
 820             LOG.error("{}: An exception occurred when applying snapshot", persistenceId(), e);
 821         } finally {
 822             LOG.info("{}: Done applying snapshot", persistenceId());
 823         }
 824     }
 825
 826     @Override
 827     protected void onStateChanged() {
 828         boolean isLeader = isLeader();
 829         for (ActorSelection dataChangeListener : dataChangeListeners) {
 830             dataChangeListener.tell(new EnableNotification(isLeader), getSelf());
 831         }
 832
 833         if(isLeader) {
 834             for(DelayedListenerRegistration reg: delayedListenerRegistrations) {
 835                 if(!reg.isClosed()) {
 836                     reg.setDelegate(doChangeListenerRegistration(reg.getRegisterChangeListener()));
 837                 }
 838             }
 839
 840             delayedListenerRegistrations.clear();
 841         }
 842
 843         // If this actor is no longer the leader close all the transaction chains
 844         if(!isLeader) {
 845             if(LOG.isDebugEnabled()) {
 846                 LOG.debug(
 847                     "{}: onStateChanged: Closing all transaction chains because shard {} is no longer the leader",
 848                     persistenceId(), getId());
 849             }
 850
 851             transactionFactory.closeAllTransactionChains();
 852         }
 853     }
 854
 855     @Override
 856     public String persistenceId() {
 857         return this.name;
 858     }
 859
 860     @VisibleForTesting
 861     ShardCommitCoordinator getCommitCoordinator() {
 862         return commitCoordinator;
 863     }
 864
 865
 866     private static class ShardCreator implements Creator<Shard> {
 867
 868         private static final long serialVersionUID = 1L;
 869
 870         final ShardIdentifier name;
 871         final Map<String, String> peerAddresses;
 872         final DatastoreContext datastoreContext;
 873         final SchemaContext schemaContext;
 874
 875         ShardCreator(final ShardIdentifier name, final Map<String, String> peerAddresses,
 876                 final DatastoreContext datastoreContext, final SchemaContext schemaContext) {
 877             this.name = name;
 878             this.peerAddresses = peerAddresses;
 879             this.datastoreContext = datastoreContext;
 880             this.schemaContext = schemaContext;
 881         }
 882
 883         @Override
 884         public Shard create() throws Exception {
 885             return new Shard(name, peerAddresses, datastoreContext, schemaContext);
 886         }
 887     }
 888
 889     @VisibleForTesting
 890     public InMemoryDOMDataStore getDataStore() {
 891         return store;
 892     }
 893
 894     @VisibleForTesting
 895     ShardStats getShardMBean() {
 896         return shardMBean;
 897     }
 898
 899     private static class DelayedListenerRegistration implements
 900         ListenerRegistration<AsyncDataChangeListener<YangInstanceIdentifier, NormalizedNode<?, ?>>> {
 901
 902         private volatile boolean closed;
 903
 904         private final RegisterChangeListener registerChangeListener;
 905
 906         private volatile ListenerRegistration<AsyncDataChangeListener<YangInstanceIdentifier,
 907                                                              NormalizedNode<?, ?>>> delegate;
 908
 909         DelayedListenerRegistration(final RegisterChangeListener registerChangeListener) {
 910             this.registerChangeListener = registerChangeListener;
 911         }
 912
 913         void setDelegate( final ListenerRegistration<AsyncDataChangeListener<YangInstanceIdentifier,
 914                                             NormalizedNode<?, ?>>> registration) {
 915             this.delegate = registration;
 916         }
 917
 918         boolean isClosed() {
 919             return closed;
 920         }
 921
 922         RegisterChangeListener getRegisterChangeListener() {
 923             return registerChangeListener;
 924         }
 925
 926         @Override
 927         public AsyncDataChangeListener<YangInstanceIdentifier, NormalizedNode<?, ?>> getInstance() {
 928             return delegate != null ? delegate.getInstance() : null;
 929         }
 930
 931         @Override
 932         public void close() {
 933             closed = true;
 934             if(delegate != null) {
 935                 delegate.close();
 936             }
 937         }
 938     }
 939 }