+public class Shard extends RaftActor {
+
+ private static final YangInstanceIdentifier DATASTORE_ROOT = YangInstanceIdentifier.builder().build();
+
+ private static final Object TX_COMMIT_TIMEOUT_CHECK_MESSAGE = "txCommitTimeoutCheck";
+
+ @VisibleForTesting
+ static final String DEFAULT_NAME = "default";
+
+ // The state of this Shard
+ private final InMemoryDOMDataStore store;
+
+ /// The name of this shard
+ private final ShardIdentifier name;
+
+ private final ShardStats shardMBean;
+
+ private final List<ActorSelection> dataChangeListeners = Lists.newArrayList();
+
+ private final List<DelayedListenerRegistration> delayedListenerRegistrations =
+ Lists.newArrayList();
+
+ private final DatastoreContext datastoreContext;
+
+ private final DataPersistenceProvider dataPersistenceProvider;
+
+ private SchemaContext schemaContext;
+
+ private int createSnapshotTransactionCounter;
+
+ private final ShardCommitCoordinator commitCoordinator;
+
+ private final long transactionCommitTimeout;
+
+ private Cancellable txCommitTimeoutCheckSchedule;
+
+ private final Optional<ActorRef> roleChangeNotifier;
+
+ private final MessageTracker appendEntriesReplyTracker;
+
+ /**
+ * Coordinates persistence recovery on startup.
+ */
+ private ShardRecoveryCoordinator recoveryCoordinator;
+ private List<Object> currentLogRecoveryBatch;
+
+ private final Map<String, DOMStoreTransactionChain> transactionChains = new HashMap<>();
+
+ private final String txnDispatcherPath;
+
+ protected Shard(final ShardIdentifier name, final Map<ShardIdentifier, String> peerAddresses,
+ final DatastoreContext datastoreContext, final SchemaContext schemaContext) {
+ super(name.toString(), mapPeerAddresses(peerAddresses),
+ Optional.of(datastoreContext.getShardRaftConfig()));
+
+ this.name = name;
+ this.datastoreContext = datastoreContext;
+ this.schemaContext = schemaContext;
+ this.dataPersistenceProvider = (datastoreContext.isPersistent())
+ ? new PersistentDataProvider() : new NonPersistentRaftDataProvider();
+ this.txnDispatcherPath = new Dispatchers(context().system().dispatchers())
+ .getDispatcherPath(Dispatchers.DispatcherType.Transaction);
+
+
+ LOG.info("Shard created : {}, persistent : {}", name, datastoreContext.isPersistent());
+
+ store = InMemoryDOMDataStoreFactory.create(name.toString(), null,
+ datastoreContext.getDataStoreProperties());
+
+ if(schemaContext != null) {
+ store.onGlobalContextUpdated(schemaContext);
+ }
+
+ shardMBean = ShardMBeanFactory.getShardStatsMBean(name.toString(),
+ datastoreContext.getDataStoreMXBeanType());
+ shardMBean.setNotificationManager(store.getDataChangeListenerNotificationManager());
+
+ if (isMetricsCaptureEnabled()) {
+ getContext().become(new MeteringBehavior(this));
+ }
+
+ commitCoordinator = new ShardCommitCoordinator(TimeUnit.SECONDS.convert(1, TimeUnit.MINUTES),
+ datastoreContext.getShardTransactionCommitQueueCapacity(), LOG, name.toString());
+
+ transactionCommitTimeout = TimeUnit.MILLISECONDS.convert(
+ datastoreContext.getShardTransactionCommitTimeoutInSeconds(), TimeUnit.SECONDS);
+
+ // create a notifier actor for each cluster member
+ roleChangeNotifier = createRoleChangeNotifier(name.toString());
+
+ appendEntriesReplyTracker = new MessageTracker(AppendEntriesReply.class,
+ getRaftActorContext().getConfigParams().getIsolatedCheckIntervalInMillis());
+ }
+
+ private static Map<String, String> mapPeerAddresses(
+ final Map<ShardIdentifier, String> peerAddresses) {
+ Map<String, String> map = new HashMap<>();
+
+ for (Map.Entry<ShardIdentifier, String> entry : peerAddresses
+ .entrySet()) {
+ map.put(entry.getKey().toString(), entry.getValue());
+ }
+
+ return map;
+ }
+
+ public static Props props(final ShardIdentifier name,
+ final Map<ShardIdentifier, String> peerAddresses,
+ final DatastoreContext datastoreContext, final SchemaContext schemaContext) {
+ Preconditions.checkNotNull(name, "name should not be null");
+ Preconditions.checkNotNull(peerAddresses, "peerAddresses should not be null");
+ Preconditions.checkNotNull(datastoreContext, "dataStoreContext should not be null");
+ Preconditions.checkNotNull(schemaContext, "schemaContext should not be null");
+
+ return Props.create(new ShardCreator(name, peerAddresses, datastoreContext, schemaContext));
+ }
+
+ private Optional<ActorRef> createRoleChangeNotifier(String shardId) {
+ ActorRef shardRoleChangeNotifier = this.getContext().actorOf(
+ RoleChangeNotifier.getProps(shardId), shardId + "-notifier");
+ return Optional.<ActorRef>of(shardRoleChangeNotifier);
+ }
+
+ @Override
+ public void postStop() {
+ super.postStop();
+
+ if(txCommitTimeoutCheckSchedule != null) {
+ txCommitTimeoutCheckSchedule.cancel();
+ }
+ }
+
+ @Override
+ public void onReceiveRecover(final Object message) throws Exception {
+ if(LOG.isDebugEnabled()) {
+ LOG.debug("{}: onReceiveRecover: Received message {} from {}", persistenceId(),
+ message.getClass().toString(), getSender());
+ }
+
+ if (message instanceof RecoveryFailure){
+ LOG.error("{}: Recovery failed because of this cause",
+ persistenceId(), ((RecoveryFailure) message).cause());
+
+ // Even though recovery failed, we still need to finish our recovery, eg send the
+ // ActorInitialized message and start the txCommitTimeoutCheckSchedule.
+ onRecoveryComplete();
+ } else {
+ super.onReceiveRecover(message);
+ if(LOG.isTraceEnabled()) {
+ appendEntriesReplyTracker.begin();
+ }
+ }
+ }
+
+ @Override
+ public void onReceiveCommand(final Object message) throws Exception {
+
+ MessageTracker.Context context = appendEntriesReplyTracker.received(message);
+
+ if(context.error().isPresent()){
+ LOG.trace("{} : AppendEntriesReply failed to arrive at the expected interval {}", persistenceId(),
+ context.error());
+ }
+
+ try {
+ if (message.getClass().equals(CreateTransaction.SERIALIZABLE_CLASS)) {
+ handleCreateTransaction(message);
+ } else if (message instanceof ForwardedReadyTransaction) {
+ handleForwardedReadyTransaction((ForwardedReadyTransaction) message);
+ } else if (message.getClass().equals(CanCommitTransaction.SERIALIZABLE_CLASS)) {
+ handleCanCommitTransaction(CanCommitTransaction.fromSerializable(message));
+ } else if (message.getClass().equals(CommitTransaction.SERIALIZABLE_CLASS)) {
+ handleCommitTransaction(CommitTransaction.fromSerializable(message));
+ } else if (message.getClass().equals(AbortTransaction.SERIALIZABLE_CLASS)) {
+ handleAbortTransaction(AbortTransaction.fromSerializable(message));
+ } else if (message.getClass().equals(CloseTransactionChain.SERIALIZABLE_CLASS)) {
+ closeTransactionChain(CloseTransactionChain.fromSerializable(message));
+ } else if (message instanceof RegisterChangeListener) {
+ registerChangeListener((RegisterChangeListener) message);
+ } else if (message instanceof UpdateSchemaContext) {
+ updateSchemaContext((UpdateSchemaContext) message);
+ } else if (message instanceof PeerAddressResolved) {
+ PeerAddressResolved resolved = (PeerAddressResolved) message;
+ setPeerAddress(resolved.getPeerId().toString(),
+ resolved.getPeerAddress());
+ } else if (message.equals(TX_COMMIT_TIMEOUT_CHECK_MESSAGE)) {
+ handleTransactionCommitTimeoutCheck();
+ } else {
+ super.onReceiveCommand(message);
+ }
+ } finally {
+ context.done();
+ }
+ }
+
+ @Override
+ protected Optional<ActorRef> getRoleChangeNotifier() {
+ return roleChangeNotifier;
+ }
+
+ private void handleTransactionCommitTimeoutCheck() {
+ CohortEntry cohortEntry = commitCoordinator.getCurrentCohortEntry();
+ if(cohortEntry != null) {
+ long elapsed = System.currentTimeMillis() - cohortEntry.getLastAccessTime();
+ if(elapsed > transactionCommitTimeout) {
+ LOG.warn("{}: Current transaction {} has timed out after {} ms - aborting",
+ persistenceId(), cohortEntry.getTransactionID(), transactionCommitTimeout);
+
+ doAbortTransaction(cohortEntry.getTransactionID(), null);
+ }
+ }
+ }
+
+ private void handleCommitTransaction(final CommitTransaction commit) {
+ final String transactionID = commit.getTransactionID();
+
+ LOG.debug("{}: Committing transaction {}", persistenceId(), transactionID);
+
+ // Get the current in-progress cohort entry in the commitCoordinator if it corresponds to
+ // this transaction.
+ final CohortEntry cohortEntry = commitCoordinator.getCohortEntryIfCurrent(transactionID);
+ if(cohortEntry == null) {
+ // We're not the current Tx - the Tx was likely expired b/c it took too long in
+ // between the canCommit and commit messages.
+ IllegalStateException ex = new IllegalStateException(
+ String.format("%s: Cannot commit transaction %s - it is not the current transaction",
+ persistenceId(), transactionID));
+ LOG.error(ex.getMessage());
+ shardMBean.incrementFailedTransactionsCount();
+ getSender().tell(new akka.actor.Status.Failure(ex), getSelf());
+ return;
+ }
+
+ // We perform the preCommit phase here atomically with the commit phase. This is an
+ // optimization to eliminate the overhead of an extra preCommit message. We lose front-end
+ // coordination of preCommit across shards in case of failure but preCommit should not
+ // normally fail since we ensure only one concurrent 3-phase commit.
+
+ try {
+ // We block on the future here so we don't have to worry about possibly accessing our
+ // state on a different thread outside of our dispatcher. Also, the data store
+ // currently uses a same thread executor anyway.
+ cohortEntry.getCohort().preCommit().get();
+
+ // If we do not have any followers and we are not using persistence we can
+ // apply modification to the state immediately
+ if(!hasFollowers() && !persistence().isRecoveryApplicable()){
+ applyModificationToState(getSender(), transactionID, cohortEntry.getModification());
+ } else {
+ Shard.this.persistData(getSender(), transactionID,
+ new ModificationPayload(cohortEntry.getModification()));
+ }
+ } catch (Exception e) {
+ LOG.error("{} An exception occurred while preCommitting transaction {}",
+ persistenceId(), cohortEntry.getTransactionID(), e);
+ shardMBean.incrementFailedTransactionsCount();
+ getSender().tell(new akka.actor.Status.Failure(e), getSelf());
+ }
+
+ cohortEntry.updateLastAccessTime();
+ }
+
+ private void finishCommit(@Nonnull final ActorRef sender, final @Nonnull String transactionID) {
+ // With persistence enabled, this method is called via applyState by the leader strategy
+ // after the commit has been replicated to a majority of the followers.
+
+ CohortEntry cohortEntry = commitCoordinator.getCohortEntryIfCurrent(transactionID);
+ if(cohortEntry == null) {
+ // The transaction is no longer the current commit. This can happen if the transaction
+ // was aborted prior, most likely due to timeout in the front-end. We need to finish
+ // committing the transaction though since it was successfully persisted and replicated
+ // however we can't use the original cohort b/c it was already preCommitted and may
+ // conflict with the current commit or may have been aborted so we commit with a new
+ // transaction.
+ cohortEntry = commitCoordinator.getAndRemoveCohortEntry(transactionID);
+ if(cohortEntry != null) {
+ commitWithNewTransaction(cohortEntry.getModification());
+ sender.tell(CommitTransactionReply.INSTANCE.toSerializable(), getSelf());
+ } else {
+ // This really shouldn't happen - it likely means that persistence or replication
+ // took so long to complete such that the cohort entry was expired from the cache.
+ IllegalStateException ex = new IllegalStateException(
+ String.format("%s: Could not finish committing transaction %s - no CohortEntry found",
+ persistenceId(), transactionID));
+ LOG.error(ex.getMessage());
+ sender.tell(new akka.actor.Status.Failure(ex), getSelf());
+ }
+
+ return;
+ }
+
+ LOG.debug("{}: Finishing commit for transaction {}", persistenceId(), cohortEntry.getTransactionID());
+
+ try {
+ // We block on the future here so we don't have to worry about possibly accessing our
+ // state on a different thread outside of our dispatcher. Also, the data store
+ // currently uses a same thread executor anyway.
+ cohortEntry.getCohort().commit().get();
+
+ sender.tell(CommitTransactionReply.INSTANCE.toSerializable(), getSelf());
+
+ shardMBean.incrementCommittedTransactionCount();
+ shardMBean.setLastCommittedTransactionTime(System.currentTimeMillis());
+
+ } catch (Exception e) {
+ sender.tell(new akka.actor.Status.Failure(e), getSelf());
+
+ LOG.error("{}, An exception occurred while committing transaction {}", persistenceId(),
+ transactionID, e);
+ shardMBean.incrementFailedTransactionsCount();
+ } finally {
+ commitCoordinator.currentTransactionComplete(transactionID, true);
+ }
+ }
+
+ private void handleCanCommitTransaction(final CanCommitTransaction canCommit) {
+ LOG.debug("{}: Can committing transaction {}", persistenceId(), canCommit.getTransactionID());
+ commitCoordinator.handleCanCommit(canCommit, getSender(), self());
+ }
+
+ private void handleForwardedReadyTransaction(ForwardedReadyTransaction ready) {
+ LOG.debug("{}: Readying transaction {}, client version {}", persistenceId(),
+ ready.getTransactionID(), ready.getTxnClientVersion());
+
+ // This message is forwarded by the ShardTransaction on ready. We cache the cohort in the
+ // commitCoordinator in preparation for the subsequent three phase commit initiated by
+ // the front-end.
+ commitCoordinator.transactionReady(ready.getTransactionID(), ready.getCohort(),
+ ready.getModification());
+
+ // Return our actor path as we'll handle the three phase commit, except if the Tx client
+ // version < 1 (Helium-1 version). This means the Tx was initiated by a base Helium version
+ // node. In that case, the subsequent 3-phase commit messages won't contain the
+ // transactionId so to maintain backwards compatibility, we create a separate cohort actor
+ // to provide the compatible behavior.
+ ActorRef replyActorPath = self();
+ if(ready.getTxnClientVersion() < DataStoreVersions.HELIUM_1_VERSION) {
+ LOG.debug("{}: Creating BackwardsCompatibleThreePhaseCommitCohort", persistenceId());
+ replyActorPath = getContext().actorOf(BackwardsCompatibleThreePhaseCommitCohort.props(
+ ready.getTransactionID()));
+ }
+
+ ReadyTransactionReply readyTransactionReply = new ReadyTransactionReply(
+ Serialization.serializedActorPath(replyActorPath));
+ getSender().tell(ready.isReturnSerialized() ? readyTransactionReply.toSerializable() :
+ readyTransactionReply, getSelf());
+ }
+
+ private void handleAbortTransaction(final AbortTransaction abort) {
+ doAbortTransaction(abort.getTransactionID(), getSender());
+ }
+
+ void doAbortTransaction(final String transactionID, final ActorRef sender) {
+ final CohortEntry cohortEntry = commitCoordinator.getCohortEntryIfCurrent(transactionID);
+ if(cohortEntry != null) {
+ LOG.debug("{}: Aborting transaction {}", persistenceId(), transactionID);
+
+ // We don't remove the cached cohort entry here (ie pass false) in case the Tx was
+ // aborted during replication in which case we may still commit locally if replication
+ // succeeds.
+ commitCoordinator.currentTransactionComplete(transactionID, false);
+
+ final ListenableFuture<Void> future = cohortEntry.getCohort().abort();
+ final ActorRef self = getSelf();
+
+ Futures.addCallback(future, new FutureCallback<Void>() {
+ @Override
+ public void onSuccess(final Void v) {
+ shardMBean.incrementAbortTransactionsCount();
+
+ if(sender != null) {
+ sender.tell(AbortTransactionReply.INSTANCE.toSerializable(), self);
+ }
+ }
+
+ @Override
+ public void onFailure(final Throwable t) {
+ LOG.error("{}: An exception happened during abort", persistenceId(), t);
+
+ if(sender != null) {
+ sender.tell(new akka.actor.Status.Failure(t), self);
+ }
+ }
+ });
+ }
+ }
+
+ private void handleCreateTransaction(final Object message) {
+ if (isLeader()) {
+ createTransaction(CreateTransaction.fromSerializable(message));
+ } else if (getLeader() != null) {
+ getLeader().forward(message, getContext());
+ } else {
+ getSender().tell(new akka.actor.Status.Failure(new NoShardLeaderException(String.format(
+ "Could not find leader for shard %s so transaction cannot be created. This typically happens" +
+ " when the system is coming up or recovering and a leader is being elected. Try again" +
+ " later.", persistenceId()))), getSelf());
+ }
+ }
+
+ private void closeTransactionChain(final CloseTransactionChain closeTransactionChain) {
+ DOMStoreTransactionChain chain =
+ transactionChains.remove(closeTransactionChain.getTransactionChainId());
+
+ if(chain != null) {
+ chain.close();
+ }
+ }
+
+ private ActorRef createTypedTransactionActor(int transactionType,
+ ShardTransactionIdentifier transactionId, String transactionChainId,
+ short clientVersion ) {
+
+ DOMStoreTransactionFactory factory = store;
+
+ if(!transactionChainId.isEmpty()) {
+ factory = transactionChains.get(transactionChainId);
+ if(factory == null){
+ DOMStoreTransactionChain transactionChain = store.createTransactionChain();
+ transactionChains.put(transactionChainId, transactionChain);
+ factory = transactionChain;
+ }
+ }
+
+ if(this.schemaContext == null) {
+ throw new IllegalStateException("SchemaContext is not set");
+ }
+
+ if (transactionType == TransactionProxy.TransactionType.READ_ONLY.ordinal()) {
+
+ shardMBean.incrementReadOnlyTransactionCount();
+
+ return createShardTransaction(factory.newReadOnlyTransaction(), transactionId, clientVersion);
+
+ } else if (transactionType == TransactionProxy.TransactionType.READ_WRITE.ordinal()) {
+
+ shardMBean.incrementReadWriteTransactionCount();
+
+ return createShardTransaction(factory.newReadWriteTransaction(), transactionId, clientVersion);
+
+ } else if (transactionType == TransactionProxy.TransactionType.WRITE_ONLY.ordinal()) {