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.RegisterDataTreeChangeListener;
  60 import org.opendaylight.controller.cluster.datastore.messages.UpdateSchemaContext;
  61 import org.opendaylight.controller.cluster.datastore.modification.Modification;
  62 import org.opendaylight.controller.cluster.datastore.modification.ModificationPayload;
  63 import org.opendaylight.controller.cluster.datastore.modification.MutableCompositeModification;
  64 import org.opendaylight.controller.cluster.datastore.utils.Dispatchers;
  65 import org.opendaylight.controller.cluster.datastore.utils.MessageTracker;
  66 import org.opendaylight.controller.cluster.datastore.utils.SerializationUtils;
  67 import org.opendaylight.controller.cluster.notifications.RegisterRoleChangeListener;
  68 import org.opendaylight.controller.cluster.notifications.RoleChangeNotifier;
  69 import org.opendaylight.controller.cluster.raft.RaftActor;
  70 import org.opendaylight.controller.cluster.raft.base.messages.FollowerInitialSyncUpStatus;
  71 import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
  72 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.CompositeModificationByteStringPayload;
  73 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.CompositeModificationPayload;
  74 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload;
  75 import org.opendaylight.controller.md.sal.common.api.data.AsyncDataChangeListener;
  76 import org.opendaylight.controller.md.sal.dom.store.impl.InMemoryDOMDataStore;
  77 import org.opendaylight.controller.md.sal.dom.store.impl.InMemoryDOMDataStoreFactory;
  78 import org.opendaylight.controller.sal.core.spi.data.DOMStoreThreePhaseCommitCohort;
  79 import org.opendaylight.controller.sal.core.spi.data.DOMStoreTransaction;
  80 import org.opendaylight.controller.sal.core.spi.data.DOMStoreWriteTransaction;
  81 import org.opendaylight.yangtools.concepts.ListenerRegistration;
  82 import org.opendaylight.yangtools.yang.data.api.YangInstanceIdentifier;
  83 import org.opendaylight.yangtools.yang.data.api.schema.NormalizedNode;
  84 import org.opendaylight.yangtools.yang.model.api.SchemaContext;
  85 import scala.concurrent.duration.Duration;
  86 import scala.concurrent.duration.FiniteDuration;
  87
  88 /**
  89  * A Shard represents a portion of the logical data tree <br/>
  90  * <p>
  91  * Our Shard uses InMemoryDataStore as it's internal representation and delegates all requests it
  92  * </p>
  93  */
  94 public class Shard extends RaftActor {
  95
  96     private static final YangInstanceIdentifier DATASTORE_ROOT = YangInstanceIdentifier.builder().build();
  97
  98     private static final Object TX_COMMIT_TIMEOUT_CHECK_MESSAGE = "txCommitTimeoutCheck";
  99
 100     @VisibleForTesting
 101     static final String DEFAULT_NAME = "default";
 102
 103     // The state of this Shard
 104     private final InMemoryDOMDataStore store;
 105
 106     /// The name of this shard
 107     private final String name;
 108
 109     private final ShardStats shardMBean;
 110
 111     private final List<ActorSelection> dataChangeListeners =  Lists.newArrayList();
 112
 113     private final List<DelayedListenerRegistration> delayedListenerRegistrations =
 114                                                                        Lists.newArrayList();
 115
 116     private DatastoreContext datastoreContext;
 117
 118     private SchemaContext schemaContext;
 119
 120     private int createSnapshotTransactionCounter;
 121
 122     private final ShardCommitCoordinator commitCoordinator;
 123
 124     private long transactionCommitTimeout;
 125
 126     private Cancellable txCommitTimeoutCheckSchedule;
 127
 128     private final Optional<ActorRef> roleChangeNotifier;
 129
 130     private final MessageTracker appendEntriesReplyTracker;
 131
 132     private final ReadyTransactionReply READY_TRANSACTION_REPLY = new ReadyTransactionReply(
 133             Serialization.serializedActorPath(getSelf()));
 134
 135
 136     /**
 137      * Coordinates persistence recovery on startup.
 138      */
 139     private ShardRecoveryCoordinator recoveryCoordinator;
 140
 141     private final DOMTransactionFactory transactionFactory;
 142
 143     private final String txnDispatcherPath;
 144
 145     private final DataTreeChangeListenerSupport treeChangeSupport = new DataTreeChangeListenerSupport(this);
 146
 147     protected Shard(final ShardIdentifier name, final Map<String, String> peerAddresses,
 148             final DatastoreContext datastoreContext, final SchemaContext schemaContext) {
 149         super(name.toString(), new HashMap<>(peerAddresses), Optional.of(datastoreContext.getShardRaftConfig()));
 150
 151         this.name = name.toString();
 152         this.datastoreContext = datastoreContext;
 153         this.schemaContext = schemaContext;
 154         this.txnDispatcherPath = new Dispatchers(context().system().dispatchers())
 155                 .getDispatcherPath(Dispatchers.DispatcherType.Transaction);
 156
 157         setPersistence(datastoreContext.isPersistent());
 158
 159         LOG.info("Shard created : {}, persistent : {}", name, datastoreContext.isPersistent());
 160
 161         store = InMemoryDOMDataStoreFactory.create(name.toString(), null,
 162                 datastoreContext.getDataStoreProperties());
 163
 164         if (schemaContext != null) {
 165             store.onGlobalContextUpdated(schemaContext);
 166         }
 167
 168         shardMBean = ShardMBeanFactory.getShardStatsMBean(name.toString(),
 169                 datastoreContext.getDataStoreMXBeanType());
 170         shardMBean.setNotificationManager(store.getDataChangeListenerNotificationManager());
 171         shardMBean.setShardActor(getSelf());
 172
 173         if (isMetricsCaptureEnabled()) {
 174             getContext().become(new MeteringBehavior(this));
 175         }
 176
 177         transactionFactory = new DOMTransactionFactory(store, shardMBean, LOG, this.name);
 178
 179         commitCoordinator = new ShardCommitCoordinator(transactionFactory,
 180                 TimeUnit.SECONDS.convert(5, TimeUnit.MINUTES),
 181                 datastoreContext.getShardTransactionCommitQueueCapacity(), self(), LOG, this.name);
 182
 183         setTransactionCommitTimeout();
 184
 185         // create a notifier actor for each cluster member
 186         roleChangeNotifier = createRoleChangeNotifier(name.toString());
 187
 188         appendEntriesReplyTracker = new MessageTracker(AppendEntriesReply.class,
 189                 getRaftActorContext().getConfigParams().getIsolatedCheckIntervalInMillis());
 190
 191         recoveryCoordinator = new ShardRecoveryCoordinator(store, persistenceId(), LOG);
 192     }
 193
 194     private void setTransactionCommitTimeout() {
 195         transactionCommitTimeout = TimeUnit.MILLISECONDS.convert(
 196                 datastoreContext.getShardTransactionCommitTimeoutInSeconds(), TimeUnit.SECONDS);
 197     }
 198
 199     public static Props props(final ShardIdentifier name,
 200         final Map<String, String> peerAddresses,
 201         final DatastoreContext datastoreContext, final SchemaContext schemaContext) {
 202         Preconditions.checkNotNull(name, "name should not be null");
 203         Preconditions.checkNotNull(peerAddresses, "peerAddresses should not be null");
 204         Preconditions.checkNotNull(datastoreContext, "dataStoreContext should not be null");
 205         Preconditions.checkNotNull(schemaContext, "schemaContext should not be null");
 206
 207         return Props.create(new ShardCreator(name, peerAddresses, datastoreContext, schemaContext));
 208     }
 209
 210     private Optional<ActorRef> createRoleChangeNotifier(String shardId) {
 211         ActorRef shardRoleChangeNotifier = this.getContext().actorOf(
 212             RoleChangeNotifier.getProps(shardId), shardId + "-notifier");
 213         return Optional.of(shardRoleChangeNotifier);
 214     }
 215
 216     @Override
 217     public void postStop() {
 218         LOG.info("Stopping Shard {}", persistenceId());
 219
 220         super.postStop();
 221
 222         if(txCommitTimeoutCheckSchedule != null) {
 223             txCommitTimeoutCheckSchedule.cancel();
 224         }
 225
 226         shardMBean.unregisterMBean();
 227     }
 228
 229     @Override
 230     public void onReceiveRecover(final Object message) throws Exception {
 231         if(LOG.isDebugEnabled()) {
 232             LOG.debug("{}: onReceiveRecover: Received message {} from {}", persistenceId(),
 233                 message.getClass().toString(), getSender());
 234         }
 235
 236         if (message instanceof RecoveryFailure){
 237             LOG.error("{}: Recovery failed because of this cause",
 238                     persistenceId(), ((RecoveryFailure) message).cause());
 239
 240             // Even though recovery failed, we still need to finish our recovery, eg send the
 241             // ActorInitialized message and start the txCommitTimeoutCheckSchedule.
 242             onRecoveryComplete();
 243         } else {
 244             super.onReceiveRecover(message);
 245             if(LOG.isTraceEnabled()) {
 246                 appendEntriesReplyTracker.begin();
 247             }
 248         }
 249     }
 250
 251     @Override
 252     public void onReceiveCommand(final Object message) throws Exception {
 253
 254         MessageTracker.Context context = appendEntriesReplyTracker.received(message);
 255
 256         if(context.error().isPresent()){
 257             LOG.trace("{} : AppendEntriesReply failed to arrive at the expected interval {}", persistenceId(),
 258                     context.error());
 259         }
 260
 261         try {
 262             if (CreateTransaction.SERIALIZABLE_CLASS.isInstance(message)) {
 263                 handleCreateTransaction(message);
 264             } else if (BatchedModifications.class.isInstance(message)) {
 265                 handleBatchedModifications((BatchedModifications)message);
 266             } else if (message instanceof ForwardedReadyTransaction) {
 267                 handleForwardedReadyTransaction((ForwardedReadyTransaction) message);
 268             } else if (CanCommitTransaction.SERIALIZABLE_CLASS.isInstance(message)) {
 269                 handleCanCommitTransaction(CanCommitTransaction.fromSerializable(message));
 270             } else if (CommitTransaction.SERIALIZABLE_CLASS.isInstance(message)) {
 271                 handleCommitTransaction(CommitTransaction.fromSerializable(message));
 272             } else if (AbortTransaction.SERIALIZABLE_CLASS.isInstance(message)) {
 273                 handleAbortTransaction(AbortTransaction.fromSerializable(message));
 274             } else if (CloseTransactionChain.SERIALIZABLE_CLASS.isInstance(message)) {
 275                 closeTransactionChain(CloseTransactionChain.fromSerializable(message));
 276             } else if (message instanceof RegisterChangeListener) {
 277                 registerChangeListener((RegisterChangeListener) message);
 278             } else if (message instanceof RegisterDataTreeChangeListener) {
 279                 treeChangeSupport.onMessage((RegisterDataTreeChangeListener) message, isLeader());
 280             } else if (message instanceof UpdateSchemaContext) {
 281                 updateSchemaContext((UpdateSchemaContext) message);
 282             } else if (message instanceof PeerAddressResolved) {
 283                 PeerAddressResolved resolved = (PeerAddressResolved) message;
 284                 setPeerAddress(resolved.getPeerId().toString(),
 285                         resolved.getPeerAddress());
 286             } else if (message.equals(TX_COMMIT_TIMEOUT_CHECK_MESSAGE)) {
 287                 handleTransactionCommitTimeoutCheck();
 288             } else if(message instanceof DatastoreContext) {
 289                 onDatastoreContext((DatastoreContext)message);
 290             } else if(message instanceof RegisterRoleChangeListener){
 291                 roleChangeNotifier.get().forward(message, context());
 292             } else if (message instanceof FollowerInitialSyncUpStatus){
 293                 shardMBean.setFollowerInitialSyncStatus(((FollowerInitialSyncUpStatus) message).isInitialSyncDone());
 294                 context().parent().tell(message, self());
 295             } else {
 296                 super.onReceiveCommand(message);
 297             }
 298         } finally {
 299             context.done();
 300         }
 301     }
 302
 303     @Override
 304     protected Optional<ActorRef> getRoleChangeNotifier() {
 305         return roleChangeNotifier;
 306     }
 307
 308     private void onDatastoreContext(DatastoreContext context) {
 309         datastoreContext = context;
 310
 311         commitCoordinator.setQueueCapacity(datastoreContext.getShardTransactionCommitQueueCapacity());
 312
 313         setTransactionCommitTimeout();
 314
 315         if(datastoreContext.isPersistent() && !persistence().isRecoveryApplicable()) {
 316             setPersistence(true);
 317         } else if(!datastoreContext.isPersistent() && persistence().isRecoveryApplicable()) {
 318             setPersistence(false);
 319         }
 320
 321         updateConfigParams(datastoreContext.getShardRaftConfig());
 322     }
 323
 324     private void handleTransactionCommitTimeoutCheck() {
 325         CohortEntry cohortEntry = commitCoordinator.getCurrentCohortEntry();
 326         if(cohortEntry != null) {
 327             long elapsed = System.currentTimeMillis() - cohortEntry.getLastAccessTime();
 328             if(elapsed > transactionCommitTimeout) {
 329                 LOG.warn("{}: Current transaction {} has timed out after {} ms - aborting",
 330                         persistenceId(), cohortEntry.getTransactionID(), transactionCommitTimeout);
 331
 332                 doAbortTransaction(cohortEntry.getTransactionID(), null);
 333             }
 334         }
 335     }
 336
 337     private void handleCommitTransaction(final CommitTransaction commit) {
 338         final String transactionID = commit.getTransactionID();
 339
 340         LOG.debug("{}: Committing transaction {}", persistenceId(), transactionID);
 341
 342         // Get the current in-progress cohort entry in the commitCoordinator if it corresponds to
 343         // this transaction.
 344         final CohortEntry cohortEntry = commitCoordinator.getCohortEntryIfCurrent(transactionID);
 345         if(cohortEntry == null) {
 346             // We're not the current Tx - the Tx was likely expired b/c it took too long in
 347             // between the canCommit and commit messages.
 348             IllegalStateException ex = new IllegalStateException(
 349                     String.format("%s: Cannot commit transaction %s - it is not the current transaction",
 350                             persistenceId(), transactionID));
 351             LOG.error(ex.getMessage());
 352             shardMBean.incrementFailedTransactionsCount();
 353             getSender().tell(new akka.actor.Status.Failure(ex), getSelf());
 354             return;
 355         }
 356
 357         // We perform the preCommit phase here atomically with the commit phase. This is an
 358         // optimization to eliminate the overhead of an extra preCommit message. We lose front-end
 359         // coordination of preCommit across shards in case of failure but preCommit should not
 360         // normally fail since we ensure only one concurrent 3-phase commit.
 361
 362         try {
 363             // We block on the future here so we don't have to worry about possibly accessing our
 364             // state on a different thread outside of our dispatcher. Also, the data store
 365             // currently uses a same thread executor anyway.
 366             cohortEntry.getCohort().preCommit().get();
 367
 368             // If we do not have any followers and we are not using persistence
 369             // or if cohortEntry has no modifications
 370             // we can apply modification to the state immediately
 371             if((!hasFollowers() && !persistence().isRecoveryApplicable()) || (!cohortEntry.hasModifications())){
 372                 applyModificationToState(getSender(), transactionID, cohortEntry.getModification());
 373             } else {
 374                 Shard.this.persistData(getSender(), transactionID,
 375                         new ModificationPayload(cohortEntry.getModification()));
 376             }
 377         } catch (Exception e) {
 378             LOG.error("{} An exception occurred while preCommitting transaction {}",
 379                     persistenceId(), cohortEntry.getTransactionID(), e);
 380             shardMBean.incrementFailedTransactionsCount();
 381             getSender().tell(new akka.actor.Status.Failure(e), getSelf());
 382         }
 383
 384         cohortEntry.updateLastAccessTime();
 385     }
 386
 387     private void finishCommit(@Nonnull final ActorRef sender, final @Nonnull String transactionID) {
 388         // With persistence enabled, this method is called via applyState by the leader strategy
 389         // after the commit has been replicated to a majority of the followers.
 390
 391         CohortEntry cohortEntry = commitCoordinator.getCohortEntryIfCurrent(transactionID);
 392         if(cohortEntry == null) {
 393             // The transaction is no longer the current commit. This can happen if the transaction
 394             // was aborted prior, most likely due to timeout in the front-end. We need to finish
 395             // committing the transaction though since it was successfully persisted and replicated
 396             // however we can't use the original cohort b/c it was already preCommitted and may
 397             // conflict with the current commit or may have been aborted so we commit with a new
 398             // transaction.
 399             cohortEntry = commitCoordinator.getAndRemoveCohortEntry(transactionID);
 400             if(cohortEntry != null) {
 401                 commitWithNewTransaction(cohortEntry.getModification());
 402                 sender.tell(CommitTransactionReply.INSTANCE.toSerializable(), getSelf());
 403             } else {
 404                 // This really shouldn't happen - it likely means that persistence or replication
 405                 // took so long to complete such that the cohort entry was expired from the cache.
 406                 IllegalStateException ex = new IllegalStateException(
 407                         String.format("%s: Could not finish committing transaction %s - no CohortEntry found",
 408                                 persistenceId(), transactionID));
 409                 LOG.error(ex.getMessage());
 410                 sender.tell(new akka.actor.Status.Failure(ex), getSelf());
 411             }
 412
 413             return;
 414         }
 415
 416         LOG.debug("{}: Finishing commit for transaction {}", persistenceId(), cohortEntry.getTransactionID());
 417
 418         try {
 419             // We block on the future here so we don't have to worry about possibly accessing our
 420             // state on a different thread outside of our dispatcher. Also, the data store
 421             // currently uses a same thread executor anyway.
 422             cohortEntry.getCohort().commit().get();
 423
 424             sender.tell(CommitTransactionReply.INSTANCE.toSerializable(), getSelf());
 425
 426             shardMBean.incrementCommittedTransactionCount();
 427             shardMBean.setLastCommittedTransactionTime(System.currentTimeMillis());
 428
 429         } catch (Exception e) {
 430             sender.tell(new akka.actor.Status.Failure(e), getSelf());
 431
 432             LOG.error("{}, An exception occurred while committing transaction {}", persistenceId(),
 433                     transactionID, e);
 434             shardMBean.incrementFailedTransactionsCount();
 435         } finally {
 436             commitCoordinator.currentTransactionComplete(transactionID, true);
 437         }
 438     }
 439
 440     private void handleCanCommitTransaction(final CanCommitTransaction canCommit) {
 441         LOG.debug("{}: Can committing transaction {}", persistenceId(), canCommit.getTransactionID());
 442         commitCoordinator.handleCanCommit(canCommit, getSender(), self());
 443     }
 444
 445     private void handleBatchedModifications(BatchedModifications batched) {
 446         // This message is sent to prepare the modificationsa transaction directly on the Shard as an
 447         // optimization to avoid the extra overhead of a separate ShardTransaction actor. On the last
 448         // BatchedModifications message, the caller sets the ready flag in the message indicating
 449         // modifications are complete. The reply contains the cohort actor path (this actor) for the caller
 450         // to initiate the 3-phase commit. This also avoids the overhead of sending an additional
 451         // ReadyTransaction message.
 452
 453         // If we're not the leader then forward to the leader. This is a safety measure - we shouldn't
 454         // normally get here if we're not the leader as the front-end (TransactionProxy) should determine
 455         // the primary/leader shard. However with timing and caching on the front-end, there's a small
 456         // window where it could have a stale leader during leadership transitions.
 457         //
 458         if(isLeader()) {
 459             try {
 460                 BatchedModificationsReply reply = commitCoordinator.handleTransactionModifications(batched);
 461                 sender().tell(reply, self());
 462             } catch (Exception e) {
 463                 LOG.error("{}: Error handling BatchedModifications for Tx {}", persistenceId(),
 464                         batched.getTransactionID(), e);
 465                 getSender().tell(new akka.actor.Status.Failure(e), getSelf());
 466             }
 467         } else {
 468             ActorSelection leader = getLeader();
 469             if(leader != null) {
 470                 // TODO: what if this is not the first batch and leadership changed in between batched messages?
 471                 // We could check if the commitCoordinator already has a cached entry and forward all the previous
 472                 // batched modifications.
 473                 LOG.debug("{}: Forwarding BatchedModifications to leader {}", persistenceId(), leader);
 474                 leader.forward(batched, getContext());
 475             } else {
 476                 // TODO: rather than throwing an immediate exception, we could schedule a timer to try again to make
 477                 // it more resilient in case we're in the process of electing a new leader.
 478                 getSender().tell(new akka.actor.Status.Failure(new NoShardLeaderException(String.format(
 479                     "Could not find the leader for shard %s. This typically happens" +
 480                     " when the system is coming up or recovering and a leader is being elected. Try again" +
 481                     " later.", persistenceId()))), getSelf());
 482             }
 483         }
 484     }
 485
 486     private void handleForwardedReadyTransaction(ForwardedReadyTransaction ready) {
 487         LOG.debug("{}: Readying transaction {}, client version {}", persistenceId(),
 488                 ready.getTransactionID(), ready.getTxnClientVersion());
 489
 490         // This message is forwarded by the ShardTransaction on ready. We cache the cohort in the
 491         // commitCoordinator in preparation for the subsequent three phase commit initiated by
 492         // the front-end.
 493         commitCoordinator.transactionReady(ready.getTransactionID(), ready.getCohort(),
 494                 (MutableCompositeModification) ready.getModification());
 495
 496         // Return our actor path as we'll handle the three phase commit, except if the Tx client
 497         // version < 1 (Helium-1 version). This means the Tx was initiated by a base Helium version
 498         // node. In that case, the subsequent 3-phase commit messages won't contain the
 499         // transactionId so to maintain backwards compatibility, we create a separate cohort actor
 500         // to provide the compatible behavior.
 501         if(ready.getTxnClientVersion() < DataStoreVersions.HELIUM_1_VERSION) {
 502             LOG.debug("{}: Creating BackwardsCompatibleThreePhaseCommitCohort", persistenceId());
 503             ActorRef replyActorPath = getContext().actorOf(BackwardsCompatibleThreePhaseCommitCohort.props(
 504                     ready.getTransactionID()));
 505
 506             ReadyTransactionReply readyTransactionReply =
 507                     new ReadyTransactionReply(Serialization.serializedActorPath(replyActorPath));
 508             getSender().tell(ready.isReturnSerialized() ? readyTransactionReply.toSerializable() :
 509                     readyTransactionReply, getSelf());
 510
 511         } else {
 512
 513             getSender().tell(ready.isReturnSerialized() ? READY_TRANSACTION_REPLY.toSerializable() :
 514                     READY_TRANSACTION_REPLY, getSelf());
 515         }
 516     }
 517
 518     private void handleAbortTransaction(final AbortTransaction abort) {
 519         doAbortTransaction(abort.getTransactionID(), getSender());
 520     }
 521
 522     void doAbortTransaction(final String transactionID, final ActorRef sender) {
 523         final CohortEntry cohortEntry = commitCoordinator.getCohortEntryIfCurrent(transactionID);
 524         if(cohortEntry != null) {
 525             LOG.debug("{}: Aborting transaction {}", persistenceId(), transactionID);
 526
 527             // We don't remove the cached cohort entry here (ie pass false) in case the Tx was
 528             // aborted during replication in which case we may still commit locally if replication
 529             // succeeds.
 530             commitCoordinator.currentTransactionComplete(transactionID, false);
 531
 532             final ListenableFuture<Void> future = cohortEntry.getCohort().abort();
 533             final ActorRef self = getSelf();
 534
 535             Futures.addCallback(future, new FutureCallback<Void>() {
 536                 @Override
 537                 public void onSuccess(final Void v) {
 538                     shardMBean.incrementAbortTransactionsCount();
 539
 540                     if(sender != null) {
 541                         sender.tell(AbortTransactionReply.INSTANCE.toSerializable(), self);
 542                     }
 543                 }
 544
 545                 @Override
 546                 public void onFailure(final Throwable t) {
 547                     LOG.error("{}: An exception happened during abort", persistenceId(), t);
 548
 549                     if(sender != null) {
 550                         sender.tell(new akka.actor.Status.Failure(t), self);
 551                     }
 552                 }
 553             });
 554         }
 555     }
 556
 557     private void handleCreateTransaction(final Object message) {
 558         if (isLeader()) {
 559             createTransaction(CreateTransaction.fromSerializable(message));
 560         } else if (getLeader() != null) {
 561             getLeader().forward(message, getContext());
 562         } else {
 563             getSender().tell(new akka.actor.Status.Failure(new NoShardLeaderException(String.format(
 564                 "Could not find leader for shard %s so transaction cannot be created. This typically happens" +
 565                 " when the system is coming up or recovering and a leader is being elected. Try again" +
 566                 " later.", persistenceId()))), getSelf());
 567         }
 568     }
 569
 570     private void closeTransactionChain(final CloseTransactionChain closeTransactionChain) {
 571         transactionFactory.closeTransactionChain(closeTransactionChain.getTransactionChainId());
 572     }
 573
 574     private ActorRef createTypedTransactionActor(int transactionType,
 575             ShardTransactionIdentifier transactionId, String transactionChainId,
 576             short clientVersion ) {
 577
 578         DOMStoreTransaction transaction = transactionFactory.newTransaction(
 579                 TransactionProxy.TransactionType.fromInt(transactionType), transactionId.toString(),
 580                 transactionChainId);
 581
 582         return createShardTransaction(transaction, transactionId, clientVersion);
 583     }
 584
 585     private ActorRef createShardTransaction(DOMStoreTransaction transaction, ShardTransactionIdentifier transactionId,
 586                                             short clientVersion){
 587         return getContext().actorOf(
 588                 ShardTransaction.props(transaction, getSelf(),
 589                         schemaContext, datastoreContext, shardMBean,
 590                         transactionId.getRemoteTransactionId(), clientVersion)
 591                         .withDispatcher(txnDispatcherPath),
 592                 transactionId.toString());
 593
 594     }
 595
 596     private void createTransaction(CreateTransaction createTransaction) {
 597         try {
 598             ActorRef transactionActor = createTransaction(createTransaction.getTransactionType(),
 599                 createTransaction.getTransactionId(), createTransaction.getTransactionChainId(),
 600                 createTransaction.getVersion());
 601
 602             getSender().tell(new CreateTransactionReply(Serialization.serializedActorPath(transactionActor),
 603                     createTransaction.getTransactionId()).toSerializable(), getSelf());
 604         } catch (Exception e) {
 605             getSender().tell(new akka.actor.Status.Failure(e), getSelf());
 606         }
 607     }
 608
 609     private ActorRef createTransaction(int transactionType, String remoteTransactionId,
 610             String transactionChainId, short clientVersion) {
 611
 612
 613         ShardTransactionIdentifier transactionId = new ShardTransactionIdentifier(remoteTransactionId);
 614
 615         if(LOG.isDebugEnabled()) {
 616             LOG.debug("{}: Creating transaction : {} ", persistenceId(), transactionId);
 617         }
 618
 619         ActorRef transactionActor = createTypedTransactionActor(transactionType, transactionId,
 620                 transactionChainId, clientVersion);
 621
 622         return transactionActor;
 623     }
 624
 625     private void syncCommitTransaction(final DOMStoreWriteTransaction transaction)
 626         throws ExecutionException, InterruptedException {
 627         DOMStoreThreePhaseCommitCohort commitCohort = transaction.ready();
 628         commitCohort.preCommit().get();
 629         commitCohort.commit().get();
 630     }
 631
 632     private void commitWithNewTransaction(final Modification modification) {
 633         DOMStoreWriteTransaction tx = store.newWriteOnlyTransaction();
 634         modification.apply(tx);
 635         try {
 636             syncCommitTransaction(tx);
 637             shardMBean.incrementCommittedTransactionCount();
 638             shardMBean.setLastCommittedTransactionTime(System.currentTimeMillis());
 639         } catch (InterruptedException | ExecutionException e) {
 640             shardMBean.incrementFailedTransactionsCount();
 641             LOG.error("{}: Failed to commit", persistenceId(), e);
 642         }
 643     }
 644
 645     private void updateSchemaContext(final UpdateSchemaContext message) {
 646         this.schemaContext = message.getSchemaContext();
 647         updateSchemaContext(message.getSchemaContext());
 648         store.onGlobalContextUpdated(message.getSchemaContext());
 649     }
 650
 651     @VisibleForTesting
 652     void updateSchemaContext(final SchemaContext schemaContext) {
 653         store.onGlobalContextUpdated(schemaContext);
 654     }
 655
 656     private void registerChangeListener(final RegisterChangeListener registerChangeListener) {
 657
 658         LOG.debug("{}: registerDataChangeListener for {}", persistenceId(), registerChangeListener.getPath());
 659
 660         ListenerRegistration<AsyncDataChangeListener<YangInstanceIdentifier,
 661                                                      NormalizedNode<?, ?>>> registration;
 662         if(isLeader()) {
 663             registration = doChangeListenerRegistration(registerChangeListener);
 664         } else {
 665             LOG.debug("{}: Shard is not the leader - delaying registration", persistenceId());
 666
 667             DelayedListenerRegistration delayedReg =
 668                     new DelayedListenerRegistration(registerChangeListener);
 669             delayedListenerRegistrations.add(delayedReg);
 670             registration = delayedReg;
 671         }
 672
 673         ActorRef listenerRegistration = getContext().actorOf(
 674                 DataChangeListenerRegistration.props(registration));
 675
 676         LOG.debug("{}: registerDataChangeListener sending reply, listenerRegistrationPath = {} ",
 677                 persistenceId(), listenerRegistration.path());
 678
 679         getSender().tell(new RegisterChangeListenerReply(listenerRegistration), getSelf());
 680     }
 681
 682     private ListenerRegistration<AsyncDataChangeListener<YangInstanceIdentifier,
 683                                                NormalizedNode<?, ?>>> doChangeListenerRegistration(
 684             final RegisterChangeListener registerChangeListener) {
 685
 686         ActorSelection dataChangeListenerPath = getContext().system().actorSelection(
 687                 registerChangeListener.getDataChangeListenerPath());
 688
 689         // Notify the listener if notifications should be enabled or not
 690         // If this shard is the leader then it will enable notifications else
 691         // it will not
 692         dataChangeListenerPath.tell(new EnableNotification(true), getSelf());
 693
 694         // Now store a reference to the data change listener so it can be notified
 695         // at a later point if notifications should be enabled or disabled
 696         dataChangeListeners.add(dataChangeListenerPath);
 697
 698         AsyncDataChangeListener<YangInstanceIdentifier, NormalizedNode<?, ?>> listener =
 699                 new DataChangeListenerProxy(dataChangeListenerPath);
 700
 701         LOG.debug("{}: Registering for path {}", persistenceId(), registerChangeListener.getPath());
 702
 703         return store.registerChangeListener(registerChangeListener.getPath(), listener,
 704                 registerChangeListener.getScope());
 705     }
 706
 707     private boolean isMetricsCaptureEnabled(){
 708         CommonConfig config = new CommonConfig(getContext().system().settings().config());
 709         return config.isMetricCaptureEnabled();
 710     }
 711
 712     @Override
 713     protected
 714     void startLogRecoveryBatch(final int maxBatchSize) {
 715         recoveryCoordinator.startLogRecoveryBatch(maxBatchSize);
 716     }
 717
 718     @Override
 719     protected void appendRecoveredLogEntry(final Payload data) {
 720         recoveryCoordinator.appendRecoveredLogPayload(data);
 721     }
 722
 723     @Override
 724     protected void applyRecoverySnapshot(final byte[] snapshotBytes) {
 725         recoveryCoordinator.applyRecoveredSnapshot(snapshotBytes);
 726     }
 727
 728     @Override
 729     protected void applyCurrentLogRecoveryBatch() {
 730         recoveryCoordinator.applyCurrentLogRecoveryBatch();
 731     }
 732
 733     @Override
 734     protected void onRecoveryComplete() {
 735         recoveryCoordinator = null;
 736
 737         //notify shard manager
 738         getContext().parent().tell(new ActorInitialized(), getSelf());
 739
 740         // Being paranoid here - this method should only be called once but just in case...
 741         if(txCommitTimeoutCheckSchedule == null) {
 742             // Schedule a message to be periodically sent to check if the current in-progress
 743             // transaction should be expired and aborted.
 744             FiniteDuration period = Duration.create(transactionCommitTimeout / 3, TimeUnit.MILLISECONDS);
 745             txCommitTimeoutCheckSchedule = getContext().system().scheduler().schedule(
 746                     period, period, getSelf(),
 747                     TX_COMMIT_TIMEOUT_CHECK_MESSAGE, getContext().dispatcher(), ActorRef.noSender());
 748         }
 749     }
 750
 751     @Override
 752     protected void applyState(final ActorRef clientActor, final String identifier, final Object data) {
 753
 754         if(data instanceof ModificationPayload) {
 755             try {
 756                 applyModificationToState(clientActor, identifier, ((ModificationPayload) data).getModification());
 757             } catch (ClassNotFoundException | IOException e) {
 758                 LOG.error("{}: Error extracting ModificationPayload", persistenceId(), e);
 759             }
 760         }
 761         else if (data instanceof CompositeModificationPayload) {
 762             Object modification = ((CompositeModificationPayload) data).getModification();
 763
 764             applyModificationToState(clientActor, identifier, modification);
 765         } else if(data instanceof CompositeModificationByteStringPayload ){
 766             Object modification = ((CompositeModificationByteStringPayload) data).getModification();
 767
 768             applyModificationToState(clientActor, identifier, modification);
 769         } else {
 770             LOG.error("{}: Unknown state received {} Class loader = {} CompositeNodeMod.ClassLoader = {}",
 771                     persistenceId(), data, data.getClass().getClassLoader(),
 772                     CompositeModificationPayload.class.getClassLoader());
 773         }
 774     }
 775
 776     private void applyModificationToState(ActorRef clientActor, String identifier, Object modification) {
 777         if(modification == null) {
 778             LOG.error(
 779                     "{}: modification is null - this is very unexpected, clientActor = {}, identifier = {}",
 780                     persistenceId(), identifier, clientActor != null ? clientActor.path().toString() : null);
 781         } else if(clientActor == null) {
 782             // There's no clientActor to which to send a commit reply so we must be applying
 783             // replicated state from the leader.
 784             commitWithNewTransaction(MutableCompositeModification.fromSerializable(modification));
 785         } else {
 786             // This must be the OK to commit after replication consensus.
 787             finishCommit(clientActor, identifier);
 788         }
 789     }
 790
 791     @Override
 792     protected void createSnapshot() {
 793         // Create a transaction actor. We are really going to treat the transaction as a worker
 794         // so that this actor does not get block building the snapshot. THe transaction actor will
 795         // after processing the CreateSnapshot message.
 796
 797         ActorRef createSnapshotTransaction = createTransaction(
 798                 TransactionProxy.TransactionType.READ_ONLY.ordinal(),
 799                 "createSnapshot" + ++createSnapshotTransactionCounter, "",
 800                 DataStoreVersions.CURRENT_VERSION);
 801
 802         createSnapshotTransaction.tell(CreateSnapshot.INSTANCE, self());
 803     }
 804
 805     @VisibleForTesting
 806     @Override
 807     protected void applySnapshot(final byte[] snapshotBytes) {
 808         // Since this will be done only on Recovery or when this actor is a Follower
 809         // we can safely commit everything in here. We not need to worry about event notifications
 810         // as they would have already been disabled on the follower
 811
 812         LOG.info("{}: Applying snapshot", persistenceId());
 813         try {
 814             DOMStoreWriteTransaction transaction = store.newWriteOnlyTransaction();
 815
 816             NormalizedNode<?, ?> node = SerializationUtils.deserializeNormalizedNode(snapshotBytes);
 817
 818             // delete everything first
 819             transaction.delete(DATASTORE_ROOT);
 820
 821             // Add everything from the remote node back
 822             transaction.write(DATASTORE_ROOT, node);
 823             syncCommitTransaction(transaction);
 824         } catch (InterruptedException | ExecutionException e) {
 825             LOG.error("{}: An exception occurred when applying snapshot", persistenceId(), e);
 826         } finally {
 827             LOG.info("{}: Done applying snapshot", persistenceId());
 828         }
 829     }
 830
 831     @Override
 832     protected void onStateChanged() {
 833         boolean isLeader = isLeader();
 834         treeChangeSupport.onLeadershipChange(isLeader);
 835
 836         for (ActorSelection dataChangeListener : dataChangeListeners) {
 837             dataChangeListener.tell(new EnableNotification(isLeader), getSelf());
 838         }
 839
 840         if(isLeader) {
 841             for(DelayedListenerRegistration reg: delayedListenerRegistrations) {
 842                 if(!reg.isClosed()) {
 843                     reg.setDelegate(doChangeListenerRegistration(reg.getRegisterChangeListener()));
 844                 }
 845             }
 846
 847             delayedListenerRegistrations.clear();
 848         }
 849
 850         // If this actor is no longer the leader close all the transaction chains
 851         if(!isLeader) {
 852             if(LOG.isDebugEnabled()) {
 853                 LOG.debug(
 854                     "{}: onStateChanged: Closing all transaction chains because shard {} is no longer the leader",
 855                     persistenceId(), getId());
 856             }
 857
 858             transactionFactory.closeAllTransactionChains();
 859         }
 860     }
 861
 862     @Override
 863     public String persistenceId() {
 864         return this.name;
 865     }
 866
 867     @VisibleForTesting
 868     ShardCommitCoordinator getCommitCoordinator() {
 869         return commitCoordinator;
 870     }
 871
 872
 873     private static class ShardCreator implements Creator<Shard> {
 874
 875         private static final long serialVersionUID = 1L;
 876
 877         final ShardIdentifier name;
 878         final Map<String, String> peerAddresses;
 879         final DatastoreContext datastoreContext;
 880         final SchemaContext schemaContext;
 881
 882         ShardCreator(final ShardIdentifier name, final Map<String, String> peerAddresses,
 883                 final DatastoreContext datastoreContext, final SchemaContext schemaContext) {
 884             this.name = name;
 885             this.peerAddresses = peerAddresses;
 886             this.datastoreContext = datastoreContext;
 887             this.schemaContext = schemaContext;
 888         }
 889
 890         @Override
 891         public Shard create() throws Exception {
 892             return new Shard(name, peerAddresses, datastoreContext, schemaContext);
 893         }
 894     }
 895
 896     @VisibleForTesting
 897     public InMemoryDOMDataStore getDataStore() {
 898         return store;
 899     }
 900
 901     @VisibleForTesting
 902     ShardStats getShardMBean() {
 903         return shardMBean;
 904     }
 905
 906     private static class DelayedListenerRegistration implements
 907         ListenerRegistration<AsyncDataChangeListener<YangInstanceIdentifier, NormalizedNode<?, ?>>> {
 908
 909         private volatile boolean closed;
 910
 911         private final RegisterChangeListener registerChangeListener;
 912
 913         private volatile ListenerRegistration<AsyncDataChangeListener<YangInstanceIdentifier,
 914                                                              NormalizedNode<?, ?>>> delegate;
 915
 916         DelayedListenerRegistration(final RegisterChangeListener registerChangeListener) {
 917             this.registerChangeListener = registerChangeListener;
 918         }
 919
 920         void setDelegate( final ListenerRegistration<AsyncDataChangeListener<YangInstanceIdentifier,
 921                                             NormalizedNode<?, ?>>> registration) {
 922             this.delegate = registration;
 923         }
 924
 925         boolean isClosed() {
 926             return closed;
 927         }
 928
 929         RegisterChangeListener getRegisterChangeListener() {
 930             return registerChangeListener;
 931         }
 932
 933         @Override
 934         public AsyncDataChangeListener<YangInstanceIdentifier, NormalizedNode<?, ?>> getInstance() {
 935             return delegate != null ? delegate.getInstance() : null;
 936         }
 937
 938         @Override
 939         public void close() {
 940             closed = true;
 941             if(delegate != null) {
 942                 delegate.close();
 943             }
 944         }
 945     }
 946 }