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.Collection;
  27 import java.util.HashMap;
  28 import java.util.List;
  29 import java.util.Map;
  30 import java.util.concurrent.ExecutionException;
  31 import java.util.concurrent.TimeUnit;
  32 import javax.annotation.Nonnull;
  33 import org.opendaylight.controller.cluster.DataPersistenceProvider;
  34 import org.opendaylight.controller.cluster.common.actor.CommonConfig;
  35 import org.opendaylight.controller.cluster.common.actor.MeteringBehavior;
  36 import org.opendaylight.controller.cluster.datastore.ShardCommitCoordinator.CohortEntry;
  37 import org.opendaylight.controller.cluster.datastore.compat.BackwardsCompatibleThreePhaseCommitCohort;
  38 import org.opendaylight.controller.cluster.datastore.exceptions.NoShardLeaderException;
  39 import org.opendaylight.controller.cluster.datastore.identifiers.ShardIdentifier;
  40 import org.opendaylight.controller.cluster.datastore.identifiers.ShardTransactionIdentifier;
  41 import org.opendaylight.controller.cluster.datastore.jmx.mbeans.shard.ShardMBeanFactory;
  42 import org.opendaylight.controller.cluster.datastore.jmx.mbeans.shard.ShardStats;
  43 import org.opendaylight.controller.cluster.datastore.messages.AbortTransaction;
  44 import org.opendaylight.controller.cluster.datastore.messages.AbortTransactionReply;
  45 import org.opendaylight.controller.cluster.datastore.messages.ActorInitialized;
  46 import org.opendaylight.controller.cluster.datastore.messages.BatchedModifications;
  47 import org.opendaylight.controller.cluster.datastore.messages.BatchedModificationsReply;
  48 import org.opendaylight.controller.cluster.datastore.messages.CanCommitTransaction;
  49 import org.opendaylight.controller.cluster.datastore.messages.CloseTransactionChain;
  50 import org.opendaylight.controller.cluster.datastore.messages.CommitTransaction;
  51 import org.opendaylight.controller.cluster.datastore.messages.CommitTransactionReply;
  52 import org.opendaylight.controller.cluster.datastore.messages.CreateSnapshot;
  53 import org.opendaylight.controller.cluster.datastore.messages.CreateTransaction;
  54 import org.opendaylight.controller.cluster.datastore.messages.CreateTransactionReply;
  55 import org.opendaylight.controller.cluster.datastore.messages.EnableNotification;
  56 import org.opendaylight.controller.cluster.datastore.messages.ForwardedReadyTransaction;
  57 import org.opendaylight.controller.cluster.datastore.messages.PeerAddressResolved;
  58 import org.opendaylight.controller.cluster.datastore.messages.ReadyTransactionReply;
  59 import org.opendaylight.controller.cluster.datastore.messages.RegisterChangeListener;
  60 import org.opendaylight.controller.cluster.datastore.messages.RegisterChangeListenerReply;
  61 import org.opendaylight.controller.cluster.datastore.messages.UpdateSchemaContext;
  62 import org.opendaylight.controller.cluster.datastore.modification.Modification;
  63 import org.opendaylight.controller.cluster.datastore.modification.ModificationPayload;
  64 import org.opendaylight.controller.cluster.datastore.modification.MutableCompositeModification;
  65 import org.opendaylight.controller.cluster.datastore.utils.Dispatchers;
  66 import org.opendaylight.controller.cluster.datastore.utils.MessageTracker;
  67 import org.opendaylight.controller.cluster.datastore.utils.SerializationUtils;
  68 import org.opendaylight.controller.cluster.notifications.RegisterRoleChangeListener;
  69 import org.opendaylight.controller.cluster.notifications.RoleChangeNotifier;
  70 import org.opendaylight.controller.cluster.raft.RaftActor;
  71 import org.opendaylight.controller.cluster.raft.base.messages.FollowerInitialSyncUpStatus;
  72 import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
  73 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.CompositeModificationByteStringPayload;
  74 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.CompositeModificationPayload;
  75 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload;
  76 import org.opendaylight.controller.md.sal.common.api.data.AsyncDataChangeListener;
  77 import org.opendaylight.controller.md.sal.dom.store.impl.InMemoryDOMDataStore;
  78 import org.opendaylight.controller.md.sal.dom.store.impl.InMemoryDOMDataStoreFactory;
  79 import org.opendaylight.controller.sal.core.spi.data.DOMStoreThreePhaseCommitCohort;
  80 import org.opendaylight.controller.sal.core.spi.data.DOMStoreTransaction;
  81 import org.opendaylight.controller.sal.core.spi.data.DOMStoreWriteTransaction;
  82 import org.opendaylight.yangtools.concepts.ListenerRegistration;
  83 import org.opendaylight.yangtools.yang.data.api.YangInstanceIdentifier;
  84 import org.opendaylight.yangtools.yang.data.api.schema.NormalizedNode;
  85 import org.opendaylight.yangtools.yang.model.api.SchemaContext;
  86 import scala.concurrent.duration.Duration;
  87 import scala.concurrent.duration.FiniteDuration;
  88
  89 /**
  90  * A Shard represents a portion of the logical data tree <br/>
  91  * <p>
  92  * Our Shard uses InMemoryDataStore as it's internal representation and delegates all requests it
  93  * </p>
  94  */
  95 public class Shard extends RaftActor {
  96
  97     private static final YangInstanceIdentifier DATASTORE_ROOT = YangInstanceIdentifier.builder().build();
  98
  99     private static final Object TX_COMMIT_TIMEOUT_CHECK_MESSAGE = "txCommitTimeoutCheck";
 100
 101     @VisibleForTesting
 102     static final String DEFAULT_NAME = "default";
 103
 104     // The state of this Shard
 105     private final InMemoryDOMDataStore store;
 106
 107     /// The name of this shard
 108     private final String name;
 109
 110     private final ShardStats shardMBean;
 111
 112     private final List<ActorSelection> dataChangeListeners =  Lists.newArrayList();
 113
 114     private final List<DelayedListenerRegistration> delayedListenerRegistrations =
 115                                                                        Lists.newArrayList();
 116
 117     private DatastoreContext datastoreContext;
 118
 119     private DataPersistenceProvider dataPersistenceProvider;
 120
 121     private SchemaContext schemaContext;
 122
 123     private int createSnapshotTransactionCounter;
 124
 125     private final ShardCommitCoordinator commitCoordinator;
 126
 127     private long transactionCommitTimeout;
 128
 129     private Cancellable txCommitTimeoutCheckSchedule;
 130
 131     private final Optional<ActorRef> roleChangeNotifier;
 132
 133     private final MessageTracker appendEntriesReplyTracker;
 134
 135     private final ReadyTransactionReply READY_TRANSACTION_REPLY = new ReadyTransactionReply(
 136             Serialization.serializedActorPath(getSelf()));
 137
 138
 139     /**
 140      * Coordinates persistence recovery on startup.
 141      */
 142     private ShardRecoveryCoordinator recoveryCoordinator;
 143     private List<Object> currentLogRecoveryBatch;
 144
 145     private final DOMTransactionFactory transactionFactory;
 146
 147     private final String txnDispatcherPath;
 148
 149     protected Shard(final ShardIdentifier name, final Map<ShardIdentifier, String> peerAddresses,
 150             final DatastoreContext datastoreContext, final SchemaContext schemaContext) {
 151         super(name.toString(), mapPeerAddresses(peerAddresses),
 152                 Optional.of(datastoreContext.getShardRaftConfig()));
 153
 154         this.name = name.toString();
 155         this.datastoreContext = datastoreContext;
 156         this.schemaContext = schemaContext;
 157         this.dataPersistenceProvider = (datastoreContext.isPersistent())
 158                 ? new PersistentDataProvider() : new NonPersistentRaftDataProvider();
 159         this.txnDispatcherPath = new Dispatchers(context().system().dispatchers())
 160                 .getDispatcherPath(Dispatchers.DispatcherType.Transaction);
 161
 162
 163         LOG.info("Shard created : {}, persistent : {}", name, datastoreContext.isPersistent());
 164
 165         store = InMemoryDOMDataStoreFactory.create(name.toString(), null,
 166                 datastoreContext.getDataStoreProperties());
 167
 168         if(schemaContext != null) {
 169             store.onGlobalContextUpdated(schemaContext);
 170         }
 171
 172         shardMBean = ShardMBeanFactory.getShardStatsMBean(name.toString(),
 173                 datastoreContext.getDataStoreMXBeanType());
 174         shardMBean.setNotificationManager(store.getDataChangeListenerNotificationManager());
 175         shardMBean.setShardActor(getSelf());
 176
 177         if (isMetricsCaptureEnabled()) {
 178             getContext().become(new MeteringBehavior(this));
 179         }
 180
 181         transactionFactory = new DOMTransactionFactory(store, shardMBean, LOG, this.name);
 182
 183         commitCoordinator = new ShardCommitCoordinator(transactionFactory,
 184                 TimeUnit.SECONDS.convert(5, TimeUnit.MINUTES),
 185                 datastoreContext.getShardTransactionCommitQueueCapacity(), self(), LOG, this.name);
 186
 187         setTransactionCommitTimeout();
 188
 189         // create a notifier actor for each cluster member
 190         roleChangeNotifier = createRoleChangeNotifier(name.toString());
 191
 192         appendEntriesReplyTracker = new MessageTracker(AppendEntriesReply.class,
 193                 getRaftActorContext().getConfigParams().getIsolatedCheckIntervalInMillis());
 194     }
 195
 196     private void setTransactionCommitTimeout() {
 197         transactionCommitTimeout = TimeUnit.MILLISECONDS.convert(
 198                 datastoreContext.getShardTransactionCommitTimeoutInSeconds(), TimeUnit.SECONDS);
 199     }
 200
 201     private static Map<String, String> mapPeerAddresses(
 202         final Map<ShardIdentifier, String> peerAddresses) {
 203         Map<String, String> map = new HashMap<>();
 204
 205         for (Map.Entry<ShardIdentifier, String> entry : peerAddresses
 206             .entrySet()) {
 207             map.put(entry.getKey().toString(), entry.getValue());
 208         }
 209
 210         return map;
 211     }
 212
 213     public static Props props(final ShardIdentifier name,
 214         final Map<ShardIdentifier, String> peerAddresses,
 215         final DatastoreContext datastoreContext, final SchemaContext schemaContext) {
 216         Preconditions.checkNotNull(name, "name should not be null");
 217         Preconditions.checkNotNull(peerAddresses, "peerAddresses should not be null");
 218         Preconditions.checkNotNull(datastoreContext, "dataStoreContext should not be null");
 219         Preconditions.checkNotNull(schemaContext, "schemaContext should not be null");
 220
 221         return Props.create(new ShardCreator(name, peerAddresses, datastoreContext, schemaContext));
 222     }
 223
 224     private Optional<ActorRef> createRoleChangeNotifier(String shardId) {
 225         ActorRef shardRoleChangeNotifier = this.getContext().actorOf(
 226             RoleChangeNotifier.getProps(shardId), shardId + "-notifier");
 227         return Optional.of(shardRoleChangeNotifier);
 228     }
 229
 230     @Override
 231     public void postStop() {
 232         LOG.info("Stopping Shard {}", persistenceId());
 233
 234         super.postStop();
 235
 236         if(txCommitTimeoutCheckSchedule != null) {
 237             txCommitTimeoutCheckSchedule.cancel();
 238         }
 239
 240         shardMBean.unregisterMBean();
 241     }
 242
 243     @Override
 244     public void onReceiveRecover(final Object message) throws Exception {
 245         if(LOG.isDebugEnabled()) {
 246             LOG.debug("{}: onReceiveRecover: Received message {} from {}", persistenceId(),
 247                 message.getClass().toString(), getSender());
 248         }
 249
 250         if (message instanceof RecoveryFailure){
 251             LOG.error("{}: Recovery failed because of this cause",
 252                     persistenceId(), ((RecoveryFailure) message).cause());
 253
 254             // Even though recovery failed, we still need to finish our recovery, eg send the
 255             // ActorInitialized message and start the txCommitTimeoutCheckSchedule.
 256             onRecoveryComplete();
 257         } else {
 258             super.onReceiveRecover(message);
 259             if(LOG.isTraceEnabled()) {
 260                 appendEntriesReplyTracker.begin();
 261             }
 262         }
 263     }
 264
 265     @Override
 266     public void onReceiveCommand(final Object message) throws Exception {
 267
 268         MessageTracker.Context context = appendEntriesReplyTracker.received(message);
 269
 270         if(context.error().isPresent()){
 271             LOG.trace("{} : AppendEntriesReply failed to arrive at the expected interval {}", persistenceId(),
 272                     context.error());
 273         }
 274
 275         try {
 276             if (CreateTransaction.SERIALIZABLE_CLASS.isInstance(message)) {
 277                 handleCreateTransaction(message);
 278             } else if (BatchedModifications.class.isInstance(message)) {
 279                 handleBatchedModifications((BatchedModifications)message);
 280             } else if (message instanceof ForwardedReadyTransaction) {
 281                 handleForwardedReadyTransaction((ForwardedReadyTransaction) message);
 282             } else if (CanCommitTransaction.SERIALIZABLE_CLASS.isInstance(message)) {
 283                 handleCanCommitTransaction(CanCommitTransaction.fromSerializable(message));
 284             } else if (CommitTransaction.SERIALIZABLE_CLASS.isInstance(message)) {
 285                 handleCommitTransaction(CommitTransaction.fromSerializable(message));
 286             } else if (AbortTransaction.SERIALIZABLE_CLASS.isInstance(message)) {
 287                 handleAbortTransaction(AbortTransaction.fromSerializable(message));
 288             } else if (CloseTransactionChain.SERIALIZABLE_CLASS.isInstance(message)) {
 289                 closeTransactionChain(CloseTransactionChain.fromSerializable(message));
 290             } else if (message instanceof RegisterChangeListener) {
 291                 registerChangeListener((RegisterChangeListener) message);
 292             } else if (message instanceof UpdateSchemaContext) {
 293                 updateSchemaContext((UpdateSchemaContext) message);
 294             } else if (message instanceof PeerAddressResolved) {
 295                 PeerAddressResolved resolved = (PeerAddressResolved) message;
 296                 setPeerAddress(resolved.getPeerId().toString(),
 297                         resolved.getPeerAddress());
 298             } else if (message.equals(TX_COMMIT_TIMEOUT_CHECK_MESSAGE)) {
 299                 handleTransactionCommitTimeoutCheck();
 300             } else if(message instanceof DatastoreContext) {
 301                 onDatastoreContext((DatastoreContext)message);
 302             } else if(message instanceof RegisterRoleChangeListener){
 303                 roleChangeNotifier.get().forward(message, context());
 304             } else if (message instanceof FollowerInitialSyncUpStatus){
 305                 shardMBean.setFollowerInitialSyncStatus(((FollowerInitialSyncUpStatus) message).isInitialSyncDone());
 306                 context().parent().tell(message, self());
 307             } else {
 308                 super.onReceiveCommand(message);
 309             }
 310         } finally {
 311             context.done();
 312         }
 313     }
 314
 315     @Override
 316     protected Optional<ActorRef> getRoleChangeNotifier() {
 317         return roleChangeNotifier;
 318     }
 319
 320     private void onDatastoreContext(DatastoreContext context) {
 321         datastoreContext = context;
 322
 323         commitCoordinator.setQueueCapacity(datastoreContext.getShardTransactionCommitQueueCapacity());
 324
 325         setTransactionCommitTimeout();
 326
 327         if(datastoreContext.isPersistent() &&
 328                 dataPersistenceProvider instanceof NonPersistentRaftDataProvider) {
 329             dataPersistenceProvider = new PersistentDataProvider();
 330         } else if(!datastoreContext.isPersistent() &&
 331                 dataPersistenceProvider instanceof PersistentDataProvider) {
 332             dataPersistenceProvider = new NonPersistentRaftDataProvider();
 333         }
 334
 335         updateConfigParams(datastoreContext.getShardRaftConfig());
 336     }
 337
 338     private void handleTransactionCommitTimeoutCheck() {
 339         CohortEntry cohortEntry = commitCoordinator.getCurrentCohortEntry();
 340         if(cohortEntry != null) {
 341             long elapsed = System.currentTimeMillis() - cohortEntry.getLastAccessTime();
 342             if(elapsed > transactionCommitTimeout) {
 343                 LOG.warn("{}: Current transaction {} has timed out after {} ms - aborting",
 344                         persistenceId(), cohortEntry.getTransactionID(), transactionCommitTimeout);
 345
 346                 doAbortTransaction(cohortEntry.getTransactionID(), null);
 347             }
 348         }
 349     }
 350
 351     private void handleCommitTransaction(final CommitTransaction commit) {
 352         final String transactionID = commit.getTransactionID();
 353
 354         LOG.debug("{}: Committing transaction {}", persistenceId(), transactionID);
 355
 356         // Get the current in-progress cohort entry in the commitCoordinator if it corresponds to
 357         // this transaction.
 358         final CohortEntry cohortEntry = commitCoordinator.getCohortEntryIfCurrent(transactionID);
 359         if(cohortEntry == null) {
 360             // We're not the current Tx - the Tx was likely expired b/c it took too long in
 361             // between the canCommit and commit messages.
 362             IllegalStateException ex = new IllegalStateException(
 363                     String.format("%s: Cannot commit transaction %s - it is not the current transaction",
 364                             persistenceId(), transactionID));
 365             LOG.error(ex.getMessage());
 366             shardMBean.incrementFailedTransactionsCount();
 367             getSender().tell(new akka.actor.Status.Failure(ex), getSelf());
 368             return;
 369         }
 370
 371         // We perform the preCommit phase here atomically with the commit phase. This is an
 372         // optimization to eliminate the overhead of an extra preCommit message. We lose front-end
 373         // coordination of preCommit across shards in case of failure but preCommit should not
 374         // normally fail since we ensure only one concurrent 3-phase commit.
 375
 376         try {
 377             // We block on the future here so we don't have to worry about possibly accessing our
 378             // state on a different thread outside of our dispatcher. Also, the data store
 379             // currently uses a same thread executor anyway.
 380             cohortEntry.getCohort().preCommit().get();
 381
 382             // If we do not have any followers and we are not using persistence
 383             // or if cohortEntry has no modifications
 384             // we can apply modification to the state immediately
 385             if((!hasFollowers() && !persistence().isRecoveryApplicable()) || (!cohortEntry.hasModifications())){
 386                 applyModificationToState(getSender(), transactionID, cohortEntry.getModification());
 387             } else {
 388                 Shard.this.persistData(getSender(), transactionID,
 389                         new ModificationPayload(cohortEntry.getModification()));
 390             }
 391         } catch (Exception e) {
 392             LOG.error("{} An exception occurred while preCommitting transaction {}",
 393                     persistenceId(), cohortEntry.getTransactionID(), e);
 394             shardMBean.incrementFailedTransactionsCount();
 395             getSender().tell(new akka.actor.Status.Failure(e), getSelf());
 396         }
 397
 398         cohortEntry.updateLastAccessTime();
 399     }
 400
 401     private void finishCommit(@Nonnull final ActorRef sender, final @Nonnull String transactionID) {
 402         // With persistence enabled, this method is called via applyState by the leader strategy
 403         // after the commit has been replicated to a majority of the followers.
 404
 405         CohortEntry cohortEntry = commitCoordinator.getCohortEntryIfCurrent(transactionID);
 406         if(cohortEntry == null) {
 407             // The transaction is no longer the current commit. This can happen if the transaction
 408             // was aborted prior, most likely due to timeout in the front-end. We need to finish
 409             // committing the transaction though since it was successfully persisted and replicated
 410             // however we can't use the original cohort b/c it was already preCommitted and may
 411             // conflict with the current commit or may have been aborted so we commit with a new
 412             // transaction.
 413             cohortEntry = commitCoordinator.getAndRemoveCohortEntry(transactionID);
 414             if(cohortEntry != null) {
 415                 commitWithNewTransaction(cohortEntry.getModification());
 416                 sender.tell(CommitTransactionReply.INSTANCE.toSerializable(), getSelf());
 417             } else {
 418                 // This really shouldn't happen - it likely means that persistence or replication
 419                 // took so long to complete such that the cohort entry was expired from the cache.
 420                 IllegalStateException ex = new IllegalStateException(
 421                         String.format("%s: Could not finish committing transaction %s - no CohortEntry found",
 422                                 persistenceId(), transactionID));
 423                 LOG.error(ex.getMessage());
 424                 sender.tell(new akka.actor.Status.Failure(ex), getSelf());
 425             }
 426
 427             return;
 428         }
 429
 430         LOG.debug("{}: Finishing commit for transaction {}", persistenceId(), cohortEntry.getTransactionID());
 431
 432         try {
 433             // We block on the future here so we don't have to worry about possibly accessing our
 434             // state on a different thread outside of our dispatcher. Also, the data store
 435             // currently uses a same thread executor anyway.
 436             cohortEntry.getCohort().commit().get();
 437
 438             sender.tell(CommitTransactionReply.INSTANCE.toSerializable(), getSelf());
 439
 440             shardMBean.incrementCommittedTransactionCount();
 441             shardMBean.setLastCommittedTransactionTime(System.currentTimeMillis());
 442
 443         } catch (Exception e) {
 444             sender.tell(new akka.actor.Status.Failure(e), getSelf());
 445
 446             LOG.error("{}, An exception occurred while committing transaction {}", persistenceId(),
 447                     transactionID, e);
 448             shardMBean.incrementFailedTransactionsCount();
 449         } finally {
 450             commitCoordinator.currentTransactionComplete(transactionID, true);
 451         }
 452     }
 453
 454     private void handleCanCommitTransaction(final CanCommitTransaction canCommit) {
 455         LOG.debug("{}: Can committing transaction {}", persistenceId(), canCommit.getTransactionID());
 456         commitCoordinator.handleCanCommit(canCommit, getSender(), self());
 457     }
 458
 459     private void handleBatchedModifications(BatchedModifications batched) {
 460         // This message is sent to prepare the modificationsa transaction directly on the Shard as an
 461         // optimization to avoid the extra overhead of a separate ShardTransaction actor. On the last
 462         // BatchedModifications message, the caller sets the ready flag in the message indicating
 463         // modifications are complete. The reply contains the cohort actor path (this actor) for the caller
 464         // to initiate the 3-phase commit. This also avoids the overhead of sending an additional
 465         // ReadyTransaction message.
 466
 467         // If we're not the leader then forward to the leader. This is a safety measure - we shouldn't
 468         // normally get here if we're not the leader as the front-end (TransactionProxy) should determine
 469         // the primary/leader shard. However with timing and caching on the front-end, there's a small
 470         // window where it could have a stale leader during leadership transitions.
 471         //
 472         if(isLeader()) {
 473             try {
 474                 BatchedModificationsReply reply = commitCoordinator.handleTransactionModifications(batched);
 475                 sender().tell(reply, self());
 476             } catch (Exception e) {
 477                 LOG.error("{}: Error handling BatchedModifications for Tx {}", persistenceId(),
 478                         batched.getTransactionID(), e);
 479                 getSender().tell(new akka.actor.Status.Failure(e), getSelf());
 480             }
 481         } else {
 482             ActorSelection leader = getLeader();
 483             if(leader != null) {
 484                 // TODO: what if this is not the first batch and leadership changed in between batched messages?
 485                 // We could check if the commitCoordinator already has a cached entry and forward all the previous
 486                 // batched modifications.
 487                 LOG.debug("{}: Forwarding BatchedModifications to leader {}", persistenceId(), leader);
 488                 leader.forward(batched, getContext());
 489             } else {
 490                 // TODO: rather than throwing an immediate exception, we could schedule a timer to try again to make
 491                 // it more resilient in case we're in the process of electing a new leader.
 492                 getSender().tell(new akka.actor.Status.Failure(new NoShardLeaderException(String.format(
 493                     "Could not find the leader for shard %s. This typically happens" +
 494                     " when the system is coming up or recovering and a leader is being elected. Try again" +
 495                     " later.", persistenceId()))), getSelf());
 496             }
 497         }
 498     }
 499
 500     private void handleForwardedReadyTransaction(ForwardedReadyTransaction ready) {
 501         LOG.debug("{}: Readying transaction {}, client version {}", persistenceId(),
 502                 ready.getTransactionID(), ready.getTxnClientVersion());
 503
 504         // This message is forwarded by the ShardTransaction on ready. We cache the cohort in the
 505         // commitCoordinator in preparation for the subsequent three phase commit initiated by
 506         // the front-end.
 507         commitCoordinator.transactionReady(ready.getTransactionID(), ready.getCohort(),
 508                 (MutableCompositeModification) ready.getModification());
 509
 510         // Return our actor path as we'll handle the three phase commit, except if the Tx client
 511         // version < 1 (Helium-1 version). This means the Tx was initiated by a base Helium version
 512         // node. In that case, the subsequent 3-phase commit messages won't contain the
 513         // transactionId so to maintain backwards compatibility, we create a separate cohort actor
 514         // to provide the compatible behavior.
 515         if(ready.getTxnClientVersion() < DataStoreVersions.HELIUM_1_VERSION) {
 516             LOG.debug("{}: Creating BackwardsCompatibleThreePhaseCommitCohort", persistenceId());
 517             ActorRef replyActorPath = getContext().actorOf(BackwardsCompatibleThreePhaseCommitCohort.props(
 518                     ready.getTransactionID()));
 519
 520             ReadyTransactionReply readyTransactionReply =
 521                     new ReadyTransactionReply(Serialization.serializedActorPath(replyActorPath));
 522             getSender().tell(ready.isReturnSerialized() ? readyTransactionReply.toSerializable() :
 523                     readyTransactionReply, getSelf());
 524
 525         } else {
 526
 527             getSender().tell(ready.isReturnSerialized() ? READY_TRANSACTION_REPLY.toSerializable() :
 528                     READY_TRANSACTION_REPLY, getSelf());
 529         }
 530     }
 531
 532     private void handleAbortTransaction(final AbortTransaction abort) {
 533         doAbortTransaction(abort.getTransactionID(), getSender());
 534     }
 535
 536     void doAbortTransaction(final String transactionID, final ActorRef sender) {
 537         final CohortEntry cohortEntry = commitCoordinator.getCohortEntryIfCurrent(transactionID);
 538         if(cohortEntry != null) {
 539             LOG.debug("{}: Aborting transaction {}", persistenceId(), transactionID);
 540
 541             // We don't remove the cached cohort entry here (ie pass false) in case the Tx was
 542             // aborted during replication in which case we may still commit locally if replication
 543             // succeeds.
 544             commitCoordinator.currentTransactionComplete(transactionID, false);
 545
 546             final ListenableFuture<Void> future = cohortEntry.getCohort().abort();
 547             final ActorRef self = getSelf();
 548
 549             Futures.addCallback(future, new FutureCallback<Void>() {
 550                 @Override
 551                 public void onSuccess(final Void v) {
 552                     shardMBean.incrementAbortTransactionsCount();
 553
 554                     if(sender != null) {
 555                         sender.tell(AbortTransactionReply.INSTANCE.toSerializable(), self);
 556                     }
 557                 }
 558
 559                 @Override
 560                 public void onFailure(final Throwable t) {
 561                     LOG.error("{}: An exception happened during abort", persistenceId(), t);
 562
 563                     if(sender != null) {
 564                         sender.tell(new akka.actor.Status.Failure(t), self);
 565                     }
 566                 }
 567             });
 568         }
 569     }
 570
 571     private void handleCreateTransaction(final Object message) {
 572         if (isLeader()) {
 573             createTransaction(CreateTransaction.fromSerializable(message));
 574         } else if (getLeader() != null) {
 575             getLeader().forward(message, getContext());
 576         } else {
 577             getSender().tell(new akka.actor.Status.Failure(new NoShardLeaderException(String.format(
 578                 "Could not find leader for shard %s so transaction cannot be created. This typically happens" +
 579                 " when the system is coming up or recovering and a leader is being elected. Try again" +
 580                 " later.", persistenceId()))), getSelf());
 581         }
 582     }
 583
 584     private void closeTransactionChain(final CloseTransactionChain closeTransactionChain) {
 585         transactionFactory.closeTransactionChain(closeTransactionChain.getTransactionChainId());
 586     }
 587
 588     private ActorRef createTypedTransactionActor(int transactionType,
 589             ShardTransactionIdentifier transactionId, String transactionChainId,
 590             short clientVersion ) {
 591
 592         DOMStoreTransaction transaction = transactionFactory.newTransaction(
 593                 TransactionProxy.TransactionType.fromInt(transactionType), transactionId.toString(),
 594                 transactionChainId);
 595
 596         return createShardTransaction(transaction, transactionId, clientVersion);
 597     }
 598
 599     private ActorRef createShardTransaction(DOMStoreTransaction transaction, ShardTransactionIdentifier transactionId,
 600                                             short clientVersion){
 601         return getContext().actorOf(
 602                 ShardTransaction.props(transaction, getSelf(),
 603                         schemaContext, datastoreContext, shardMBean,
 604                         transactionId.getRemoteTransactionId(), clientVersion)
 605                         .withDispatcher(txnDispatcherPath),
 606                 transactionId.toString());
 607
 608     }
 609
 610     private void createTransaction(CreateTransaction createTransaction) {
 611         try {
 612             ActorRef transactionActor = createTransaction(createTransaction.getTransactionType(),
 613                 createTransaction.getTransactionId(), createTransaction.getTransactionChainId(),
 614                 createTransaction.getVersion());
 615
 616             getSender().tell(new CreateTransactionReply(Serialization.serializedActorPath(transactionActor),
 617                     createTransaction.getTransactionId()).toSerializable(), getSelf());
 618         } catch (Exception e) {
 619             getSender().tell(new akka.actor.Status.Failure(e), getSelf());
 620         }
 621     }
 622
 623     private ActorRef createTransaction(int transactionType, String remoteTransactionId,
 624             String transactionChainId, short clientVersion) {
 625
 626
 627         ShardTransactionIdentifier transactionId = new ShardTransactionIdentifier(remoteTransactionId);
 628
 629         if(LOG.isDebugEnabled()) {
 630             LOG.debug("{}: Creating transaction : {} ", persistenceId(), transactionId);
 631         }
 632
 633         ActorRef transactionActor = createTypedTransactionActor(transactionType, transactionId,
 634                 transactionChainId, clientVersion);
 635
 636         return transactionActor;
 637     }
 638
 639     private void syncCommitTransaction(final DOMStoreWriteTransaction transaction)
 640         throws ExecutionException, InterruptedException {
 641         DOMStoreThreePhaseCommitCohort commitCohort = transaction.ready();
 642         commitCohort.preCommit().get();
 643         commitCohort.commit().get();
 644     }
 645
 646     private void commitWithNewTransaction(final Modification modification) {
 647         DOMStoreWriteTransaction tx = store.newWriteOnlyTransaction();
 648         modification.apply(tx);
 649         try {
 650             syncCommitTransaction(tx);
 651             shardMBean.incrementCommittedTransactionCount();
 652             shardMBean.setLastCommittedTransactionTime(System.currentTimeMillis());
 653         } catch (InterruptedException | ExecutionException e) {
 654             shardMBean.incrementFailedTransactionsCount();
 655             LOG.error("{}: Failed to commit", persistenceId(), e);
 656         }
 657     }
 658
 659     private void updateSchemaContext(final UpdateSchemaContext message) {
 660         this.schemaContext = message.getSchemaContext();
 661         updateSchemaContext(message.getSchemaContext());
 662         store.onGlobalContextUpdated(message.getSchemaContext());
 663     }
 664
 665     @VisibleForTesting
 666     void updateSchemaContext(final SchemaContext schemaContext) {
 667         store.onGlobalContextUpdated(schemaContext);
 668     }
 669
 670     private void registerChangeListener(final RegisterChangeListener registerChangeListener) {
 671
 672         LOG.debug("{}: registerDataChangeListener for {}", persistenceId(), registerChangeListener.getPath());
 673
 674         ListenerRegistration<AsyncDataChangeListener<YangInstanceIdentifier,
 675                                                      NormalizedNode<?, ?>>> registration;
 676         if(isLeader()) {
 677             registration = doChangeListenerRegistration(registerChangeListener);
 678         } else {
 679             LOG.debug("{}: Shard is not the leader - delaying registration", persistenceId());
 680
 681             DelayedListenerRegistration delayedReg =
 682                     new DelayedListenerRegistration(registerChangeListener);
 683             delayedListenerRegistrations.add(delayedReg);
 684             registration = delayedReg;
 685         }
 686
 687         ActorRef listenerRegistration = getContext().actorOf(
 688                 DataChangeListenerRegistration.props(registration));
 689
 690         LOG.debug("{}: registerDataChangeListener sending reply, listenerRegistrationPath = {} ",
 691                 persistenceId(), listenerRegistration.path());
 692
 693         getSender().tell(new RegisterChangeListenerReply(listenerRegistration.path()), getSelf());
 694     }
 695
 696     private ListenerRegistration<AsyncDataChangeListener<YangInstanceIdentifier,
 697                                                NormalizedNode<?, ?>>> doChangeListenerRegistration(
 698             final RegisterChangeListener registerChangeListener) {
 699
 700         ActorSelection dataChangeListenerPath = getContext().system().actorSelection(
 701                 registerChangeListener.getDataChangeListenerPath());
 702
 703         // Notify the listener if notifications should be enabled or not
 704         // If this shard is the leader then it will enable notifications else
 705         // it will not
 706         dataChangeListenerPath.tell(new EnableNotification(true), getSelf());
 707
 708         // Now store a reference to the data change listener so it can be notified
 709         // at a later point if notifications should be enabled or disabled
 710         dataChangeListeners.add(dataChangeListenerPath);
 711
 712         AsyncDataChangeListener<YangInstanceIdentifier, NormalizedNode<?, ?>> listener =
 713                 new DataChangeListenerProxy(dataChangeListenerPath);
 714
 715         LOG.debug("{}: Registering for path {}", persistenceId(), registerChangeListener.getPath());
 716
 717         return store.registerChangeListener(registerChangeListener.getPath(), listener,
 718                 registerChangeListener.getScope());
 719     }
 720
 721     private boolean isMetricsCaptureEnabled(){
 722         CommonConfig config = new CommonConfig(getContext().system().settings().config());
 723         return config.isMetricCaptureEnabled();
 724     }
 725
 726     @Override
 727     protected
 728     void startLogRecoveryBatch(final int maxBatchSize) {
 729         currentLogRecoveryBatch = Lists.newArrayListWithCapacity(maxBatchSize);
 730
 731         if(LOG.isDebugEnabled()) {
 732             LOG.debug("{}: starting log recovery batch with max size {}", persistenceId(), maxBatchSize);
 733         }
 734     }
 735
 736     @Override
 737     protected void appendRecoveredLogEntry(final Payload data) {
 738         if(data instanceof ModificationPayload) {
 739             try {
 740                 currentLogRecoveryBatch.add(((ModificationPayload) data).getModification());
 741             } catch (ClassNotFoundException | IOException e) {
 742                 LOG.error("{}: Error extracting ModificationPayload", persistenceId(), e);
 743             }
 744         } else if (data instanceof CompositeModificationPayload) {
 745             currentLogRecoveryBatch.add(((CompositeModificationPayload) data).getModification());
 746         } else if (data instanceof CompositeModificationByteStringPayload) {
 747             currentLogRecoveryBatch.add(((CompositeModificationByteStringPayload) data).getModification());
 748         } else {
 749             LOG.error("{}: Unknown state received {} during recovery", persistenceId(), data);
 750         }
 751     }
 752
 753     @Override
 754     protected void applyRecoverySnapshot(final byte[] snapshotBytes) {
 755         if(recoveryCoordinator == null) {
 756             recoveryCoordinator = new ShardRecoveryCoordinator(persistenceId(), schemaContext,
 757                     LOG, name.toString());
 758         }
 759
 760         recoveryCoordinator.submit(snapshotBytes, store.newWriteOnlyTransaction());
 761
 762         if(LOG.isDebugEnabled()) {
 763             LOG.debug("{}: submitted recovery sbapshot", persistenceId());
 764         }
 765     }
 766
 767     @Override
 768     protected void applyCurrentLogRecoveryBatch() {
 769         if(recoveryCoordinator == null) {
 770             recoveryCoordinator = new ShardRecoveryCoordinator(persistenceId(), schemaContext,
 771                     LOG, name.toString());
 772         }
 773
 774         recoveryCoordinator.submit(currentLogRecoveryBatch, store.newWriteOnlyTransaction());
 775
 776         if(LOG.isDebugEnabled()) {
 777             LOG.debug("{}: submitted log recovery batch with size {}", persistenceId(),
 778                     currentLogRecoveryBatch.size());
 779         }
 780     }
 781
 782     @Override
 783     protected void onRecoveryComplete() {
 784         if(recoveryCoordinator != null) {
 785             Collection<DOMStoreWriteTransaction> txList = recoveryCoordinator.getTransactions();
 786
 787             if(LOG.isDebugEnabled()) {
 788                 LOG.debug("{}: recovery complete - committing {} Tx's", persistenceId(), txList.size());
 789             }
 790
 791             for(DOMStoreWriteTransaction tx: txList) {
 792                 try {
 793                     syncCommitTransaction(tx);
 794                     shardMBean.incrementCommittedTransactionCount();
 795                 } catch (InterruptedException | ExecutionException e) {
 796                     shardMBean.incrementFailedTransactionsCount();
 797                     LOG.error("{}: Failed to commit", persistenceId(), e);
 798                 }
 799             }
 800         }
 801
 802         recoveryCoordinator = null;
 803         currentLogRecoveryBatch = null;
 804
 805         //notify shard manager
 806         getContext().parent().tell(new ActorInitialized(), getSelf());
 807
 808         // Being paranoid here - this method should only be called once but just in case...
 809         if(txCommitTimeoutCheckSchedule == null) {
 810             // Schedule a message to be periodically sent to check if the current in-progress
 811             // transaction should be expired and aborted.
 812             FiniteDuration period = Duration.create(transactionCommitTimeout / 3, TimeUnit.MILLISECONDS);
 813             txCommitTimeoutCheckSchedule = getContext().system().scheduler().schedule(
 814                     period, period, getSelf(),
 815                     TX_COMMIT_TIMEOUT_CHECK_MESSAGE, getContext().dispatcher(), ActorRef.noSender());
 816         }
 817     }
 818
 819     @Override
 820     protected void applyState(final ActorRef clientActor, final String identifier, final Object data) {
 821
 822         if(data instanceof ModificationPayload) {
 823             try {
 824                 applyModificationToState(clientActor, identifier, ((ModificationPayload) data).getModification());
 825             } catch (ClassNotFoundException | IOException e) {
 826                 LOG.error("{}: Error extracting ModificationPayload", persistenceId(), e);
 827             }
 828         }
 829         else if (data instanceof CompositeModificationPayload) {
 830             Object modification = ((CompositeModificationPayload) data).getModification();
 831
 832             applyModificationToState(clientActor, identifier, modification);
 833         } else if(data instanceof CompositeModificationByteStringPayload ){
 834             Object modification = ((CompositeModificationByteStringPayload) data).getModification();
 835
 836             applyModificationToState(clientActor, identifier, modification);
 837         } else {
 838             LOG.error("{}: Unknown state received {} Class loader = {} CompositeNodeMod.ClassLoader = {}",
 839                     persistenceId(), data, data.getClass().getClassLoader(),
 840                     CompositeModificationPayload.class.getClassLoader());
 841         }
 842     }
 843
 844     private void applyModificationToState(ActorRef clientActor, String identifier, Object modification) {
 845         if(modification == null) {
 846             LOG.error(
 847                     "{}: modification is null - this is very unexpected, clientActor = {}, identifier = {}",
 848                     persistenceId(), identifier, clientActor != null ? clientActor.path().toString() : null);
 849         } else if(clientActor == null) {
 850             // There's no clientActor to which to send a commit reply so we must be applying
 851             // replicated state from the leader.
 852             commitWithNewTransaction(MutableCompositeModification.fromSerializable(modification));
 853         } else {
 854             // This must be the OK to commit after replication consensus.
 855             finishCommit(clientActor, identifier);
 856         }
 857     }
 858
 859     @Override
 860     protected void createSnapshot() {
 861         // Create a transaction actor. We are really going to treat the transaction as a worker
 862         // so that this actor does not get block building the snapshot. THe transaction actor will
 863         // after processing the CreateSnapshot message.
 864
 865         ActorRef createSnapshotTransaction = createTransaction(
 866                 TransactionProxy.TransactionType.READ_ONLY.ordinal(),
 867                 "createSnapshot" + ++createSnapshotTransactionCounter, "",
 868                 DataStoreVersions.CURRENT_VERSION);
 869
 870         createSnapshotTransaction.tell(CreateSnapshot.INSTANCE, self());
 871     }
 872
 873     @VisibleForTesting
 874     @Override
 875     protected void applySnapshot(final byte[] snapshotBytes) {
 876         // Since this will be done only on Recovery or when this actor is a Follower
 877         // we can safely commit everything in here. We not need to worry about event notifications
 878         // as they would have already been disabled on the follower
 879
 880         LOG.info("{}: Applying snapshot", persistenceId());
 881         try {
 882             DOMStoreWriteTransaction transaction = store.newWriteOnlyTransaction();
 883
 884             NormalizedNode<?, ?> node = SerializationUtils.deserializeNormalizedNode(snapshotBytes);
 885
 886             // delete everything first
 887             transaction.delete(DATASTORE_ROOT);
 888
 889             // Add everything from the remote node back
 890             transaction.write(DATASTORE_ROOT, node);
 891             syncCommitTransaction(transaction);
 892         } catch (InterruptedException | ExecutionException e) {
 893             LOG.error("{}: An exception occurred when applying snapshot", persistenceId(), e);
 894         } finally {
 895             LOG.info("{}: Done applying snapshot", persistenceId());
 896         }
 897     }
 898
 899     @Override
 900     protected void onStateChanged() {
 901         boolean isLeader = isLeader();
 902         for (ActorSelection dataChangeListener : dataChangeListeners) {
 903             dataChangeListener.tell(new EnableNotification(isLeader), getSelf());
 904         }
 905
 906         if(isLeader) {
 907             for(DelayedListenerRegistration reg: delayedListenerRegistrations) {
 908                 if(!reg.isClosed()) {
 909                     reg.setDelegate(doChangeListenerRegistration(reg.getRegisterChangeListener()));
 910                 }
 911             }
 912
 913             delayedListenerRegistrations.clear();
 914         }
 915
 916         // If this actor is no longer the leader close all the transaction chains
 917         if(!isLeader) {
 918             if(LOG.isDebugEnabled()) {
 919                 LOG.debug(
 920                     "{}: onStateChanged: Closing all transaction chains because shard {} is no longer the leader",
 921                     persistenceId(), getId());
 922             }
 923
 924             transactionFactory.closeAllTransactionChains();
 925         }
 926     }
 927
 928     @Override
 929     protected DataPersistenceProvider persistence() {
 930         return dataPersistenceProvider;
 931     }
 932
 933     @Override public String persistenceId() {
 934         return this.name;
 935     }
 936
 937     @VisibleForTesting
 938     DataPersistenceProvider getDataPersistenceProvider() {
 939         return dataPersistenceProvider;
 940     }
 941
 942     @VisibleForTesting
 943     ShardCommitCoordinator getCommitCoordinator() {
 944         return commitCoordinator;
 945     }
 946
 947
 948     private static class ShardCreator implements Creator<Shard> {
 949
 950         private static final long serialVersionUID = 1L;
 951
 952         final ShardIdentifier name;
 953         final Map<ShardIdentifier, String> peerAddresses;
 954         final DatastoreContext datastoreContext;
 955         final SchemaContext schemaContext;
 956
 957         ShardCreator(final ShardIdentifier name, final Map<ShardIdentifier, String> peerAddresses,
 958                 final DatastoreContext datastoreContext, final SchemaContext schemaContext) {
 959             this.name = name;
 960             this.peerAddresses = peerAddresses;
 961             this.datastoreContext = datastoreContext;
 962             this.schemaContext = schemaContext;
 963         }
 964
 965         @Override
 966         public Shard create() throws Exception {
 967             return new Shard(name, peerAddresses, datastoreContext, schemaContext);
 968         }
 969     }
 970
 971     @VisibleForTesting
 972     public InMemoryDOMDataStore getDataStore() {
 973         return store;
 974     }
 975
 976     @VisibleForTesting
 977     ShardStats getShardMBean() {
 978         return shardMBean;
 979     }
 980
 981     private static class DelayedListenerRegistration implements
 982         ListenerRegistration<AsyncDataChangeListener<YangInstanceIdentifier, NormalizedNode<?, ?>>> {
 983
 984         private volatile boolean closed;
 985
 986         private final RegisterChangeListener registerChangeListener;
 987
 988         private volatile ListenerRegistration<AsyncDataChangeListener<YangInstanceIdentifier,
 989                                                              NormalizedNode<?, ?>>> delegate;
 990
 991         DelayedListenerRegistration(final RegisterChangeListener registerChangeListener) {
 992             this.registerChangeListener = registerChangeListener;
 993         }
 994
 995         void setDelegate( final ListenerRegistration<AsyncDataChangeListener<YangInstanceIdentifier,
 996                                             NormalizedNode<?, ?>>> registration) {
 997             this.delegate = registration;
 998         }
 999
1000         boolean isClosed() {
1001             return closed;
1002         }
1003
1004         RegisterChangeListener getRegisterChangeListener() {
1005             return registerChangeListener;
1006         }
1007
1008         @Override
1009         public AsyncDataChangeListener<YangInstanceIdentifier, NormalizedNode<?, ?>> getInstance() {
1010             return delegate != null ? delegate.getInstance() : null;
1011         }
1012
1013         @Override
1014         public void close() {
1015             closed = true;
1016             if(delegate != null) {
1017                 delegate.close();
1018             }
1019         }
1020     }
1021 }