X-Git-Url: https://git.opendaylight.org/gerrit/gitweb?p=controller.git;a=blobdiff_plain;f=opendaylight%2Fmd-sal%2Fsal-distributed-datastore%2Fsrc%2Fmain%2Fjava%2Forg%2Fopendaylight%2Fcontroller%2Fcluster%2Fdatastore%2FShard.java;h=b2a86376941d4e8e609d8e3f976041b9f7e665b3;hp=e704e42465b99e1183ca20c19742ccd2424e410f;hb=4062f5241a2a6f58ffb83dd1e9939ee66122d217;hpb=95589305be51630beab2f6b80c098ebf72bca4b9 diff --git a/opendaylight/md-sal/sal-distributed-datastore/src/main/java/org/opendaylight/controller/cluster/datastore/Shard.java b/opendaylight/md-sal/sal-distributed-datastore/src/main/java/org/opendaylight/controller/cluster/datastore/Shard.java index e704e42465..b76c0fac34 100644 --- a/opendaylight/md-sal/sal-distributed-datastore/src/main/java/org/opendaylight/controller/cluster/datastore/Shard.java +++ b/opendaylight/md-sal/sal-distributed-datastore/src/main/java/org/opendaylight/controller/cluster/datastore/Shard.java @@ -12,77 +12,63 @@ import akka.actor.ActorRef; import akka.actor.ActorSelection; import akka.actor.Cancellable; import akka.actor.Props; -import akka.japi.Creator; -import akka.persistence.RecoveryFailure; import akka.serialization.Serialization; import com.google.common.annotations.VisibleForTesting; import com.google.common.base.Optional; import com.google.common.base.Preconditions; -import com.google.common.collect.Lists; -import com.google.common.util.concurrent.FutureCallback; -import com.google.common.util.concurrent.Futures; -import com.google.common.util.concurrent.ListenableFuture; import java.io.IOException; import java.util.Collection; -import java.util.HashMap; -import java.util.List; +import java.util.Collections; import java.util.Map; import java.util.concurrent.ExecutionException; import java.util.concurrent.TimeUnit; import javax.annotation.Nonnull; -import org.opendaylight.controller.cluster.DataPersistenceProvider; import org.opendaylight.controller.cluster.common.actor.CommonConfig; +import org.opendaylight.controller.cluster.common.actor.MessageTracker; +import org.opendaylight.controller.cluster.common.actor.MessageTracker.Error; import org.opendaylight.controller.cluster.common.actor.MeteringBehavior; -import org.opendaylight.controller.cluster.datastore.ShardCommitCoordinator.CohortEntry; -import org.opendaylight.controller.cluster.datastore.compat.BackwardsCompatibleThreePhaseCommitCohort; import org.opendaylight.controller.cluster.datastore.exceptions.NoShardLeaderException; import org.opendaylight.controller.cluster.datastore.identifiers.ShardIdentifier; -import org.opendaylight.controller.cluster.datastore.identifiers.ShardTransactionIdentifier; import org.opendaylight.controller.cluster.datastore.jmx.mbeans.shard.ShardMBeanFactory; import org.opendaylight.controller.cluster.datastore.jmx.mbeans.shard.ShardStats; import org.opendaylight.controller.cluster.datastore.messages.AbortTransaction; -import org.opendaylight.controller.cluster.datastore.messages.AbortTransactionReply; import org.opendaylight.controller.cluster.datastore.messages.ActorInitialized; +import org.opendaylight.controller.cluster.datastore.messages.BatchedModifications; import org.opendaylight.controller.cluster.datastore.messages.CanCommitTransaction; import org.opendaylight.controller.cluster.datastore.messages.CloseTransactionChain; import org.opendaylight.controller.cluster.datastore.messages.CommitTransaction; import org.opendaylight.controller.cluster.datastore.messages.CommitTransactionReply; -import org.opendaylight.controller.cluster.datastore.messages.CreateSnapshot; import org.opendaylight.controller.cluster.datastore.messages.CreateTransaction; import org.opendaylight.controller.cluster.datastore.messages.CreateTransactionReply; -import org.opendaylight.controller.cluster.datastore.messages.EnableNotification; +import org.opendaylight.controller.cluster.datastore.messages.DatastoreSnapshot; +import org.opendaylight.controller.cluster.datastore.messages.DatastoreSnapshot.ShardSnapshot; import org.opendaylight.controller.cluster.datastore.messages.ForwardedReadyTransaction; +import org.opendaylight.controller.cluster.datastore.messages.GetShardDataTree; import org.opendaylight.controller.cluster.datastore.messages.PeerAddressResolved; -import org.opendaylight.controller.cluster.datastore.messages.ReadyTransactionReply; +import org.opendaylight.controller.cluster.datastore.messages.ReadyLocalTransaction; import org.opendaylight.controller.cluster.datastore.messages.RegisterChangeListener; -import org.opendaylight.controller.cluster.datastore.messages.RegisterChangeListenerReply; +import org.opendaylight.controller.cluster.datastore.messages.RegisterDataTreeChangeListener; +import org.opendaylight.controller.cluster.datastore.messages.ShardLeaderStateChanged; import org.opendaylight.controller.cluster.datastore.messages.UpdateSchemaContext; import org.opendaylight.controller.cluster.datastore.modification.Modification; -import org.opendaylight.controller.cluster.datastore.modification.ModificationPayload; import org.opendaylight.controller.cluster.datastore.modification.MutableCompositeModification; import org.opendaylight.controller.cluster.datastore.utils.Dispatchers; -import org.opendaylight.controller.cluster.datastore.utils.MessageTracker; -import org.opendaylight.controller.cluster.datastore.utils.SerializationUtils; +import org.opendaylight.controller.cluster.notifications.LeaderStateChanged; import org.opendaylight.controller.cluster.notifications.RegisterRoleChangeListener; import org.opendaylight.controller.cluster.notifications.RoleChangeNotifier; import org.opendaylight.controller.cluster.raft.RaftActor; -import org.opendaylight.controller.cluster.raft.ReplicatedLogEntry; +import org.opendaylight.controller.cluster.raft.RaftActorRecoveryCohort; +import org.opendaylight.controller.cluster.raft.RaftActorSnapshotCohort; +import org.opendaylight.controller.cluster.raft.RaftState; import org.opendaylight.controller.cluster.raft.base.messages.FollowerInitialSyncUpStatus; import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply; -import org.opendaylight.controller.cluster.raft.protobuff.client.messages.CompositeModificationByteStringPayload; -import org.opendaylight.controller.cluster.raft.protobuff.client.messages.CompositeModificationPayload; -import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload; -import org.opendaylight.controller.md.sal.common.api.data.AsyncDataChangeListener; -import org.opendaylight.controller.md.sal.dom.store.impl.InMemoryDOMDataStore; -import org.opendaylight.controller.md.sal.dom.store.impl.InMemoryDOMDataStoreFactory; -import org.opendaylight.controller.sal.core.spi.data.DOMStoreThreePhaseCommitCohort; -import org.opendaylight.controller.sal.core.spi.data.DOMStoreTransaction; -import org.opendaylight.controller.sal.core.spi.data.DOMStoreTransactionChain; -import org.opendaylight.controller.sal.core.spi.data.DOMStoreTransactionFactory; -import org.opendaylight.controller.sal.core.spi.data.DOMStoreWriteTransaction; -import org.opendaylight.yangtools.concepts.ListenerRegistration; -import org.opendaylight.yangtools.yang.data.api.YangInstanceIdentifier; -import org.opendaylight.yangtools.yang.data.api.schema.NormalizedNode; +import org.opendaylight.controller.cluster.raft.messages.ServerRemoved; +import org.opendaylight.yangtools.concepts.Identifier; +import org.opendaylight.yangtools.util.StringIdentifier; +import org.opendaylight.yangtools.yang.data.api.schema.tree.DataTreeCandidate; +import org.opendaylight.yangtools.yang.data.api.schema.tree.DataValidationFailedException; +import org.opendaylight.yangtools.yang.data.api.schema.tree.ModificationType; +import org.opendaylight.yangtools.yang.data.api.schema.tree.TreeType; import org.opendaylight.yangtools.yang.model.api.SchemaContext; import scala.concurrent.duration.Duration; import scala.concurrent.duration.FiniteDuration; @@ -90,39 +76,40 @@ import scala.concurrent.duration.FiniteDuration; /** * A Shard represents a portion of the logical data tree
*

- * Our Shard uses InMemoryDataStore as it's internal representation and delegates all requests it + * Our Shard uses InMemoryDataTree as it's internal representation and delegates all requests it *

*/ 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 Object TX_COMMIT_TIMEOUT_CHECK_MESSAGE = new Object() { + @Override + public String toString() { + return "txCommitTimeoutCheck"; + } + }; @VisibleForTesting - static final String DEFAULT_NAME = "default"; + static final Object GET_SHARD_MBEAN_MESSAGE = new Object() { + @Override + public String toString() { + return "getShardMBeanMessage"; + } + }; + + // FIXME: shard names should be encapsulated in their own class and this should be exposed as a constant. + public static final String DEFAULT_NAME = "default"; // The state of this Shard - private final InMemoryDOMDataStore store; + private final ShardDataTree store; /// The name of this shard - private final ShardIdentifier name; + private final String name; private final ShardStats shardMBean; - private final List dataChangeListeners = Lists.newArrayList(); - - private final List delayedListenerRegistrations = - Lists.newArrayList(); - private DatastoreContext datastoreContext; - private DataPersistenceProvider dataPersistenceProvider; - - private SchemaContext schemaContext; - - private int createSnapshotTransactionCounter; - private final ShardCommitCoordinator commitCoordinator; private long transactionCommitTimeout; @@ -133,53 +120,45 @@ public class Shard extends RaftActor { private final MessageTracker appendEntriesReplyTracker; - private final ReadyTransactionReply READY_TRANSACTION_REPLY = new ReadyTransactionReply( - Serialization.serializedActorPath(getSelf())); + private final ShardTransactionActorFactory transactionActorFactory; + private final ShardSnapshotCohort snapshotCohort; - /** - * Coordinates persistence recovery on startup. - */ - private ShardRecoveryCoordinator recoveryCoordinator; - private List currentLogRecoveryBatch; + private final DataTreeChangeListenerSupport treeChangeSupport = new DataTreeChangeListenerSupport(this); + private final DataChangeListenerSupport changeSupport = new DataChangeListenerSupport(this); - private final Map transactionChains = new HashMap<>(); - private final String txnDispatcherPath; + private ShardSnapshot restoreFromSnapshot; - protected Shard(final ShardIdentifier name, final Map peerAddresses, - final DatastoreContext datastoreContext, final SchemaContext schemaContext) { - super(name.toString(), mapPeerAddresses(peerAddresses), - Optional.of(datastoreContext.getShardRaftConfig())); + private final ShardTransactionMessageRetrySupport messageRetrySupport; - 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); + protected Shard(AbstractBuilder builder) { + super(builder.getId().toString(), builder.getPeerAddresses(), + Optional.of(builder.getDatastoreContext().getShardRaftConfig()), DataStoreVersions.CURRENT_VERSION); + this.name = builder.getId().toString(); + this.datastoreContext = builder.getDatastoreContext(); + this.restoreFromSnapshot = builder.getRestoreFromSnapshot(); - LOG.info("Shard created : {}, persistent : {}", name, datastoreContext.isPersistent()); + setPersistence(datastoreContext.isPersistent()); - store = InMemoryDOMDataStoreFactory.create(name.toString(), null, - datastoreContext.getDataStoreProperties()); + LOG.info("Shard created : {}, persistent : {}", name, datastoreContext.isPersistent()); - if(schemaContext != null) { - store.onGlobalContextUpdated(schemaContext); - } + store = new ShardDataTree(builder.getSchemaContext(), builder.getTreeType(), + new ShardDataTreeChangeListenerPublisherActorProxy(getContext(), name + "-DTCL-publisher"), + new ShardDataChangeListenerPublisherActorProxy(getContext(), name + "-DCL-publisher"), name); shardMBean = ShardMBeanFactory.getShardStatsMBean(name.toString(), datastoreContext.getDataStoreMXBeanType()); - shardMBean.setNotificationManager(store.getDataChangeListenerNotificationManager()); + shardMBean.setShard(this); if (isMetricsCaptureEnabled()) { getContext().become(new MeteringBehavior(this)); } - commitCoordinator = new ShardCommitCoordinator(TimeUnit.SECONDS.convert(1, TimeUnit.MINUTES), - datastoreContext.getShardTransactionCommitQueueCapacity(), LOG, name.toString()); + commitCoordinator = new ShardCommitCoordinator(store, + datastoreContext.getShardCommitQueueExpiryTimeoutInMillis(), + datastoreContext.getShardTransactionCommitQueueCapacity(), LOG, this.name); setTransactionCommitTimeout(); @@ -188,34 +167,19 @@ public class Shard extends RaftActor { appendEntriesReplyTracker = new MessageTracker(AppendEntriesReply.class, getRaftActorContext().getConfigParams().getIsolatedCheckIntervalInMillis()); - } - - private void setTransactionCommitTimeout() { - transactionCommitTimeout = TimeUnit.MILLISECONDS.convert( - datastoreContext.getShardTransactionCommitTimeoutInSeconds(), TimeUnit.SECONDS); - } - private static Map mapPeerAddresses( - final Map peerAddresses) { - Map map = new HashMap<>(); + transactionActorFactory = new ShardTransactionActorFactory(store, datastoreContext, + new Dispatchers(context().system().dispatchers()).getDispatcherPath( + Dispatchers.DispatcherType.Transaction), self(), getContext(), shardMBean); - for (Map.Entry entry : peerAddresses - .entrySet()) { - map.put(entry.getKey().toString(), entry.getValue()); - } + snapshotCohort = new ShardSnapshotCohort(transactionActorFactory, store, LOG, this.name); - return map; + messageRetrySupport = new ShardTransactionMessageRetrySupport(this); } - public static Props props(final ShardIdentifier name, - final Map 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 void setTransactionCommitTimeout() { + transactionCommitTimeout = TimeUnit.MILLISECONDS.convert( + datastoreContext.getShardTransactionCommitTimeoutInSeconds(), TimeUnit.SECONDS) / 2; } private Optional createRoleChangeNotifier(String shardId) { @@ -230,174 +194,212 @@ public class Shard extends RaftActor { super.postStop(); + messageRetrySupport.close(); + if(txCommitTimeoutCheckSchedule != null) { txCommitTimeoutCheckSchedule.cancel(); } + commitCoordinator.abortPendingTransactions("Transaction aborted due to shutdown.", this); + shardMBean.unregisterMBean(); } @Override - public void onReceiveRecover(final Object message) throws Exception { - if(LOG.isDebugEnabled()) { - LOG.debug("{}: onReceiveRecover: Received message {} from {}", persistenceId(), - message.getClass().toString(), getSender()); - } + protected void handleRecover(final Object message) { + LOG.debug("{}: onReceiveRecover: Received message {} from {}", persistenceId(), message.getClass(), + 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(); - } + super.handleRecover(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()); - } + protected void handleNonRaftCommand(final Object message) { + try (final MessageTracker.Context context = appendEntriesReplyTracker.received(message)) { + final Optional maybeError = context.error(); + if (maybeError.isPresent()) { + LOG.trace("{} : AppendEntriesReply failed to arrive at the expected interval {}", persistenceId(), + maybeError.get()); + } - try { - if (CreateTransaction.SERIALIZABLE_CLASS.isInstance(message)) { + if (CreateTransaction.isSerializedType(message)) { handleCreateTransaction(message); + } else if (message instanceof BatchedModifications) { + handleBatchedModifications((BatchedModifications)message); } else if (message instanceof ForwardedReadyTransaction) { handleForwardedReadyTransaction((ForwardedReadyTransaction) message); - } else if (CanCommitTransaction.SERIALIZABLE_CLASS.isInstance(message)) { + } else if (message instanceof ReadyLocalTransaction) { + handleReadyLocalTransaction((ReadyLocalTransaction)message); + } else if (CanCommitTransaction.isSerializedType(message)) { handleCanCommitTransaction(CanCommitTransaction.fromSerializable(message)); - } else if (CommitTransaction.SERIALIZABLE_CLASS.isInstance(message)) { + } else if (CommitTransaction.isSerializedType(message)) { handleCommitTransaction(CommitTransaction.fromSerializable(message)); - } else if (AbortTransaction.SERIALIZABLE_CLASS.isInstance(message)) { + } else if (AbortTransaction.isSerializedType(message)) { handleAbortTransaction(AbortTransaction.fromSerializable(message)); - } else if (CloseTransactionChain.SERIALIZABLE_CLASS.isInstance(message)) { + } else if (CloseTransactionChain.isSerializedType(message)) { closeTransactionChain(CloseTransactionChain.fromSerializable(message)); } else if (message instanceof RegisterChangeListener) { - registerChangeListener((RegisterChangeListener) message); + changeSupport.onMessage((RegisterChangeListener) message, isLeader(), hasLeader()); + } else if (message instanceof RegisterDataTreeChangeListener) { + treeChangeSupport.onMessage((RegisterDataTreeChangeListener) message, isLeader(), hasLeader()); } 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 if (TX_COMMIT_TIMEOUT_CHECK_MESSAGE.equals(message)) { + commitCoordinator.checkForExpiredTransactions(transactionCommitTimeout, this); } else if(message instanceof DatastoreContext) { onDatastoreContext((DatastoreContext)message); } else if(message instanceof RegisterRoleChangeListener){ roleChangeNotifier.get().forward(message, context()); - } else if (message instanceof FollowerInitialSyncUpStatus){ + } else if (message instanceof FollowerInitialSyncUpStatus) { shardMBean.setFollowerInitialSyncStatus(((FollowerInitialSyncUpStatus) message).isInitialSyncDone()); context().parent().tell(message, self()); + } else if(GET_SHARD_MBEAN_MESSAGE.equals(message)){ + sender().tell(getShardMBean(), self()); + } else if(message instanceof GetShardDataTree) { + sender().tell(store.getDataTree(), self()); + } else if(message instanceof ServerRemoved){ + context().parent().forward(message, context()); + } else if(ShardTransactionMessageRetrySupport.TIMER_MESSAGE_CLASS.isInstance(message)) { + messageRetrySupport.onTimerMessage(message); + } else if (message instanceof DataTreeCohortActorRegistry.CohortRegistryCommand) { + commitCoordinator.processCohortRegistryCommand(getSender(), + (DataTreeCohortActorRegistry.CohortRegistryCommand) message); } else { - super.onReceiveCommand(message); + super.handleNonRaftCommand(message); } - } finally { - context.done(); } } + private boolean hasLeader() { + return getLeaderId() != null; + } + + public int getPendingTxCommitQueueSize() { + return commitCoordinator.getQueueSize(); + } + + public int getCohortCacheSize() { + return commitCoordinator.getCohortCacheSize(); + } + @Override protected Optional getRoleChangeNotifier() { return roleChangeNotifier; } - private void onDatastoreContext(DatastoreContext context) { + @Override + protected LeaderStateChanged newLeaderStateChanged(String memberId, String leaderId, short leaderPayloadVersion) { + return isLeader() ? new ShardLeaderStateChanged(memberId, leaderId, store.getDataTree(), leaderPayloadVersion) + : new ShardLeaderStateChanged(memberId, leaderId, leaderPayloadVersion); + } + + protected void onDatastoreContext(DatastoreContext context) { datastoreContext = context; commitCoordinator.setQueueCapacity(datastoreContext.getShardTransactionCommitQueueCapacity()); setTransactionCommitTimeout(); - if(datastoreContext.isPersistent() && - dataPersistenceProvider instanceof NonPersistentRaftDataProvider) { - dataPersistenceProvider = new PersistentDataProvider(); - } else if(!datastoreContext.isPersistent() && - dataPersistenceProvider instanceof PersistentDataProvider) { - dataPersistenceProvider = new NonPersistentRaftDataProvider(); + if(datastoreContext.isPersistent() && !persistence().isRecoveryApplicable()) { + setPersistence(true); + } else if(!datastoreContext.isPersistent() && persistence().isRecoveryApplicable()) { + setPersistence(false); } updateConfigParams(datastoreContext.getShardRaftConfig()); } - 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); + private static boolean isEmptyCommit(final DataTreeCandidate candidate) { + return ModificationType.UNMODIFIED.equals(candidate.getRootNode().getModificationType()); + } - doAbortTransaction(cohortEntry.getTransactionID(), null); - } + void continueCommit(final CohortEntry cohortEntry) { + final DataTreeCandidate candidate = cohortEntry.getCandidate(); + + // If we do not have any followers and we are not using persistence + // or if cohortEntry has no modifications + // we can apply modification to the state immediately + if ((!hasFollowers() && !persistence().isRecoveryApplicable()) || isEmptyCommit(candidate)) { + applyModificationToState(cohortEntry.getReplySender(), cohortEntry.getTransactionID(), candidate); + } else { + persistData(cohortEntry.getReplySender(), cohortEntry.getTransactionID(), + DataTreeCandidatePayload.create(candidate)); } } 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; + if (isLeader()) { + if(!commitCoordinator.handleCommit(new StringIdentifier(commit.getTransactionID()), getSender(), this)) { + shardMBean.incrementFailedTransactionsCount(); + } + } else { + ActorSelection leader = getLeader(); + if (leader == null) { + messageRetrySupport.addMessageToRetry(commit, getSender(), + "Could not commit transaction " + commit.getTransactionID()); + } else { + LOG.debug("{}: Forwarding CommitTransaction to leader {}", persistenceId(), leader); + leader.forward(commit, getContext()); + } } + } - // 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. + private void finishCommit(@Nonnull final ActorRef sender, @Nonnull final Identifier transactionID, + @Nonnull final CohortEntry cohortEntry) { + 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().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())); + try { + cohortEntry.commit(); + } catch(ExecutionException e) { + // We may get a "store tree and candidate base differ" IllegalStateException from commit under + // certain edge case scenarios so we'll try to re-apply the candidate from scratch as a last + // resort. Eg, we're a follower and a tx payload is replicated but the leader goes down before + // applying it to the state. We then become the leader and a second tx is pre-committed and + // replicated. When consensus occurs, this will cause the first tx to be applied as a foreign + // candidate via applyState prior to the second tx. Since the second tx has already been + // pre-committed, when it gets here to commit it will get an IllegalStateException. + + // FIXME - this is not an ideal way to handle this scenario. This is temporary - a cleaner + // solution will be forthcoming. + if(e.getCause() instanceof IllegalStateException) { + LOG.debug("{}: commit failed for transaction {} - retrying as foreign candidate", persistenceId(), + transactionID, e); + store.applyForeignCandidate(transactionID, cohortEntry.getCandidate()); + } else { + throw e; + } } + + sender.tell(CommitTransactionReply.instance(cohortEntry.getClientVersion()).toSerializable(), getSelf()); + + shardMBean.incrementCommittedTransactionCount(); + shardMBean.setLastCommittedTransactionTime(System.currentTimeMillis()); + } catch (Exception e) { - LOG.error("{} An exception occurred while preCommitting transaction {}", - persistenceId(), cohortEntry.getTransactionID(), e); + sender.tell(new akka.actor.Status.Failure(e), getSelf()); + + LOG.error("{}, An exception occurred while committing transaction {}", persistenceId(), + transactionID, e); shardMBean.incrementFailedTransactionsCount(); - getSender().tell(new akka.actor.Status.Failure(e), getSelf()); + } finally { + commitCoordinator.currentTransactionComplete(transactionID, true); } - - cohortEntry.updateLastAccessTime(); } - private void finishCommit(@Nonnull final ActorRef sender, final @Nonnull String transactionID) { + private void finishCommit(@Nonnull final ActorRef sender, final @Nonnull Identifier 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) { + 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 @@ -406,8 +408,15 @@ public class Shard extends RaftActor { // transaction. cohortEntry = commitCoordinator.getAndRemoveCohortEntry(transactionID); if(cohortEntry != null) { - commitWithNewTransaction(cohortEntry.getModification()); - sender.tell(CommitTransactionReply.INSTANCE.toSerializable(), getSelf()); + try { + store.applyForeignCandidate(transactionID, cohortEntry.getCandidate()); + } catch (DataValidationFailedException e) { + shardMBean.incrementFailedTransactionsCount(); + LOG.error("{}: Failed to re-apply transaction {}", persistenceId(), transactionID, e); + } + + sender.tell(CommitTransactionReply.instance(cohortEntry.getClientVersion()).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. @@ -417,382 +426,230 @@ public class Shard extends RaftActor { 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); + } else { + finishCommit(sender, transactionID, cohortEntry); } } 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. - if(ready.getTxnClientVersion() < DataStoreVersions.HELIUM_1_VERSION) { - LOG.debug("{}: Creating BackwardsCompatibleThreePhaseCommitCohort", persistenceId()); - ActorRef replyActorPath = getContext().actorOf(BackwardsCompatibleThreePhaseCommitCohort.props( - ready.getTransactionID())); - - ReadyTransactionReply readyTransactionReply = - new ReadyTransactionReply(Serialization.serializedActorPath(replyActorPath)); - getSender().tell(ready.isReturnSerialized() ? readyTransactionReply.toSerializable() : - readyTransactionReply, getSelf()); + if (isLeader()) { + commitCoordinator.handleCanCommit(new StringIdentifier(canCommit.getTransactionID()), getSender(), this); } else { - - getSender().tell(ready.isReturnSerialized() ? READY_TRANSACTION_REPLY.toSerializable() : - READY_TRANSACTION_REPLY, getSelf()); + ActorSelection leader = getLeader(); + if (leader == null) { + messageRetrySupport.addMessageToRetry(canCommit, getSender(), + "Could not canCommit transaction " + canCommit.getTransactionID()); + } else { + LOG.debug("{}: Forwarding CanCommitTransaction to leader {}", persistenceId(), leader); + leader.forward(canCommit, getContext()); + } } } - private void handleAbortTransaction(final AbortTransaction abort) { - doAbortTransaction(abort.getTransactionID(), getSender()); + protected void handleBatchedModificationsLocal(BatchedModifications batched, ActorRef sender) { + try { + commitCoordinator.handleBatchedModifications(batched, sender, this, store.getSchemaContext()); + } catch (Exception e) { + LOG.error("{}: Error handling BatchedModifications for Tx {}", persistenceId(), + batched.getTransactionID(), e); + sender.tell(new akka.actor.Status.Failure(e), getSelf()); + } } - 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 future = cohortEntry.getCohort().abort(); - final ActorRef self = getSelf(); - - Futures.addCallback(future, new FutureCallback() { - @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); + private void handleBatchedModifications(BatchedModifications batched) { + // This message is sent to prepare the modifications transaction directly on the Shard as an + // optimization to avoid the extra overhead of a separate ShardTransaction actor. On the last + // BatchedModifications message, the caller sets the ready flag in the message indicating + // modifications are complete. The reply contains the cohort actor path (this actor) for the caller + // to initiate the 3-phase commit. This also avoids the overhead of sending an additional + // ReadyTransaction message. - if(sender != null) { - sender.tell(new akka.actor.Status.Failure(t), self); - } + // If we're not the leader then forward to the leader. This is a safety measure - we shouldn't + // normally get here if we're not the leader as the front-end (TransactionProxy) should determine + // the primary/leader shard. However with timing and caching on the front-end, there's a small + // window where it could have a stale leader during leadership transitions. + // + boolean isLeaderActive = isLeaderActive(); + if (isLeader() && isLeaderActive) { + handleBatchedModificationsLocal(batched, getSender()); + } else { + ActorSelection leader = getLeader(); + if (!isLeaderActive || leader == null) { + messageRetrySupport.addMessageToRetry(batched, getSender(), + "Could not commit transaction " + batched.getTransactionID()); + } else { + // If this is not the first batch and leadership changed in between batched messages, + // we need to reconstruct previous BatchedModifications from the transaction + // DataTreeModification, honoring the max batched modification count, and forward all the + // previous BatchedModifications to the new leader. + Collection newModifications = commitCoordinator.createForwardedBatchedModifications( + batched, datastoreContext.getShardBatchedModificationCount()); + + LOG.debug("{}: Forwarding {} BatchedModifications to leader {}", persistenceId(), + newModifications.size(), leader); + + for(BatchedModifications bm: newModifications) { + leader.forward(bm, getContext()); } - }); + } } } - 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 boolean failIfIsolatedLeader(ActorRef sender) { + if(isIsolatedLeader()) { + sender.tell(new akka.actor.Status.Failure(new NoShardLeaderException(String.format( + "Shard %s was the leader but has lost contact with all of its followers. Either all" + + " other follower nodes are down or this node is isolated by a network partition.", + persistenceId()))), getSelf()); + return true; } - } - private void closeTransactionChain(final CloseTransactionChain closeTransactionChain) { - DOMStoreTransactionChain chain = - transactionChains.remove(closeTransactionChain.getTransactionChainId()); - - if(chain != null) { - chain.close(); - } + return false; } - private ActorRef createTypedTransactionActor(int transactionType, - ShardTransactionIdentifier transactionId, String transactionChainId, - short clientVersion ) { + protected boolean isIsolatedLeader() { + return getRaftState() == RaftState.IsolatedLeader; + } - DOMStoreTransactionFactory factory = store; + private void handleReadyLocalTransaction(final ReadyLocalTransaction message) { + LOG.debug("{}: handleReadyLocalTransaction for {}", persistenceId(), message.getTransactionID()); - if(!transactionChainId.isEmpty()) { - factory = transactionChains.get(transactionChainId); - if(factory == null){ - DOMStoreTransactionChain transactionChain = store.createTransactionChain(); - transactionChains.put(transactionChainId, transactionChain); - factory = transactionChain; + boolean isLeaderActive = isLeaderActive(); + if (isLeader() && isLeaderActive) { + try { + commitCoordinator.handleReadyLocalTransaction(message, getSender(), this, store.getSchemaContext()); + } catch (Exception e) { + LOG.error("{}: Error handling ReadyLocalTransaction for Tx {}", persistenceId(), + message.getTransactionID(), e); + getSender().tell(new akka.actor.Status.Failure(e), getSelf()); + } + } else { + ActorSelection leader = getLeader(); + if (!isLeaderActive || leader == null) { + messageRetrySupport.addMessageToRetry(message, getSender(), + "Could not commit transaction " + message.getTransactionID()); + } else { + LOG.debug("{}: Forwarding ReadyLocalTransaction to leader {}", persistenceId(), leader); + message.setRemoteVersion(getCurrentBehavior().getLeaderPayloadVersion()); + leader.forward(message, getContext()); } } + } - if(this.schemaContext == null) { - throw new IllegalStateException("SchemaContext is not set"); - } - - if (transactionType == TransactionProxy.TransactionType.WRITE_ONLY.ordinal()) { - - shardMBean.incrementWriteOnlyTransactionCount(); - - return createShardTransaction(factory.newWriteOnlyTransaction(), transactionId, clientVersion); - - } else if (transactionType == TransactionProxy.TransactionType.READ_WRITE.ordinal()) { - - shardMBean.incrementReadWriteTransactionCount(); + private void handleForwardedReadyTransaction(ForwardedReadyTransaction forwardedReady) { + LOG.debug("{}: handleForwardedReadyTransaction for {}", persistenceId(), forwardedReady.getTransactionID()); - return createShardTransaction(factory.newReadWriteTransaction(), transactionId, clientVersion); + boolean isLeaderActive = isLeaderActive(); + if (isLeader() && isLeaderActive) { + commitCoordinator.handleForwardedReadyTransaction(forwardedReady, getSender(), this, + store.getSchemaContext()); + } else { + ActorSelection leader = getLeader(); + if (!isLeaderActive || leader == null) { + messageRetrySupport.addMessageToRetry(forwardedReady, getSender(), + "Could not commit transaction " + forwardedReady.getTransactionID()); + } else { + LOG.debug("{}: Forwarding ForwardedReadyTransaction to leader {}", persistenceId(), leader); - } else if (transactionType == TransactionProxy.TransactionType.READ_ONLY.ordinal()) { + ReadyLocalTransaction readyLocal = new ReadyLocalTransaction(forwardedReady.getTransactionID(), + forwardedReady.getTransaction().getSnapshot(), forwardedReady.isDoImmediateCommit()); + readyLocal.setRemoteVersion(getCurrentBehavior().getLeaderPayloadVersion()); + leader.forward(readyLocal, getContext()); + } + } + } - shardMBean.incrementReadOnlyTransactionCount(); + private void handleAbortTransaction(final AbortTransaction abort) { + doAbortTransaction(abort.getTransactionID(), getSender()); + } - return createShardTransaction(factory.newReadOnlyTransaction(), transactionId, clientVersion); + void doAbortTransaction(final String transactionID, final ActorRef sender) { + commitCoordinator.handleAbort(new StringIdentifier(transactionID), sender, this); + } + private void handleCreateTransaction(final Object message) { + if (isLeader()) { + createTransaction(CreateTransaction.fromSerializable(message)); + } else if (getLeader() != null) { + getLeader().forward(message, getContext()); } else { - throw new IllegalArgumentException( - "Shard="+name + ":CreateTransaction message has unidentified transaction type=" - + transactionType); + getSender().tell(new akka.actor.Status.Failure(new NoShardLeaderException( + "Could not create a shard transaction", persistenceId())), getSelf()); } } - private ActorRef createShardTransaction(DOMStoreTransaction transaction, ShardTransactionIdentifier transactionId, - short clientVersion){ - return getContext().actorOf( - ShardTransaction.props(transaction, getSelf(), - schemaContext, datastoreContext, shardMBean, - transactionId.getRemoteTransactionId(), clientVersion) - .withDispatcher(txnDispatcherPath), - transactionId.toString()); - + private void closeTransactionChain(final CloseTransactionChain closeTransactionChain) { + store.closeTransactionChain(closeTransactionChain.getTransactionChainId()); } private void createTransaction(CreateTransaction createTransaction) { try { + if(TransactionType.fromInt(createTransaction.getTransactionType()) != TransactionType.READ_ONLY && + failIfIsolatedLeader(getSender())) { + return; + } + ActorRef transactionActor = createTransaction(createTransaction.getTransactionType(), - createTransaction.getTransactionId(), createTransaction.getTransactionChainId(), - createTransaction.getVersion()); + createTransaction.getTransactionId(), createTransaction.getTransactionChainId()); getSender().tell(new CreateTransactionReply(Serialization.serializedActorPath(transactionActor), - createTransaction.getTransactionId()).toSerializable(), getSelf()); + createTransaction.getTransactionId(), createTransaction.getVersion()).toSerializable(), getSelf()); } catch (Exception e) { getSender().tell(new akka.actor.Status.Failure(e), getSelf()); } } - private ActorRef createTransaction(int transactionType, String remoteTransactionId, - String transactionChainId, short clientVersion) { - - - ShardTransactionIdentifier transactionId = new ShardTransactionIdentifier(remoteTransactionId); - - if(LOG.isDebugEnabled()) { - LOG.debug("{}: Creating transaction : {} ", persistenceId(), transactionId); - } - - ActorRef transactionActor = createTypedTransactionActor(transactionType, transactionId, - transactionChainId, clientVersion); - - return transactionActor; - } - - private void syncCommitTransaction(final DOMStoreWriteTransaction transaction) - throws ExecutionException, InterruptedException { - DOMStoreThreePhaseCommitCohort commitCohort = transaction.ready(); - commitCohort.preCommit().get(); - commitCohort.commit().get(); + private ActorRef createTransaction(int transactionType, String transactionId, String transactionChainId) { + LOG.debug("{}: Creating transaction : {} ", persistenceId(), transactionId); + return transactionActorFactory.newShardTransaction(TransactionType.fromInt(transactionType), + transactionId, transactionChainId); } private void commitWithNewTransaction(final Modification modification) { - DOMStoreWriteTransaction tx = store.newWriteOnlyTransaction(); - modification.apply(tx); + ReadWriteShardDataTreeTransaction tx = store.newReadWriteTransaction(modification.toString(), null); + modification.apply(tx.getSnapshot()); try { - syncCommitTransaction(tx); + snapshotCohort.syncCommitTransaction(tx); shardMBean.incrementCommittedTransactionCount(); shardMBean.setLastCommittedTransactionTime(System.currentTimeMillis()); - } catch (InterruptedException | ExecutionException e) { + } catch (Exception e) { shardMBean.incrementFailedTransactionsCount(); LOG.error("{}: Failed to commit", persistenceId(), e); } } private void updateSchemaContext(final UpdateSchemaContext message) { - this.schemaContext = message.getSchemaContext(); updateSchemaContext(message.getSchemaContext()); - store.onGlobalContextUpdated(message.getSchemaContext()); } @VisibleForTesting void updateSchemaContext(final SchemaContext schemaContext) { - store.onGlobalContextUpdated(schemaContext); + store.updateSchemaContext(schemaContext); } - private void registerChangeListener(final RegisterChangeListener registerChangeListener) { - - LOG.debug("{}: registerDataChangeListener for {}", persistenceId(), registerChangeListener.getPath()); - - ListenerRegistration>> registration; - if(isLeader()) { - registration = doChangeListenerRegistration(registerChangeListener); - } else { - LOG.debug("{}: Shard is not the leader - delaying registration", persistenceId()); - - DelayedListenerRegistration delayedReg = - new DelayedListenerRegistration(registerChangeListener); - delayedListenerRegistrations.add(delayedReg); - registration = delayedReg; - } - - ActorRef listenerRegistration = getContext().actorOf( - DataChangeListenerRegistration.props(registration)); - - LOG.debug("{}: registerDataChangeListener sending reply, listenerRegistrationPath = {} ", - persistenceId(), listenerRegistration.path()); - - getSender().tell(new RegisterChangeListenerReply(listenerRegistration.path()), getSelf()); - } - - private ListenerRegistration>> doChangeListenerRegistration( - final RegisterChangeListener registerChangeListener) { - - ActorSelection dataChangeListenerPath = getContext().system().actorSelection( - registerChangeListener.getDataChangeListenerPath()); - - // Notify the listener if notifications should be enabled or not - // If this shard is the leader then it will enable notifications else - // it will not - dataChangeListenerPath.tell(new EnableNotification(true), getSelf()); - - // Now store a reference to the data change listener so it can be notified - // at a later point if notifications should be enabled or disabled - dataChangeListeners.add(dataChangeListenerPath); - - AsyncDataChangeListener> listener = - new DataChangeListenerProxy(dataChangeListenerPath); - - LOG.debug("{}: Registering for path {}", persistenceId(), registerChangeListener.getPath()); - - return store.registerChangeListener(registerChangeListener.getPath(), listener, - registerChangeListener.getScope()); - } - - private boolean isMetricsCaptureEnabled(){ + private boolean isMetricsCaptureEnabled() { CommonConfig config = new CommonConfig(getContext().system().settings().config()); return config.isMetricCaptureEnabled(); } @Override - protected - void startLogRecoveryBatch(final int maxBatchSize) { - currentLogRecoveryBatch = Lists.newArrayListWithCapacity(maxBatchSize); - - if(LOG.isDebugEnabled()) { - LOG.debug("{}: starting log recovery batch with max size {}", persistenceId(), maxBatchSize); - } - } - - @Override - protected void appendRecoveredLogEntry(final Payload data) { - if(data instanceof ModificationPayload) { - try { - currentLogRecoveryBatch.add(((ModificationPayload) data).getModification()); - } catch (ClassNotFoundException | IOException e) { - LOG.error("{}: Error extracting ModificationPayload", persistenceId(), e); - } - } else if (data instanceof CompositeModificationPayload) { - currentLogRecoveryBatch.add(((CompositeModificationPayload) data).getModification()); - } else if (data instanceof CompositeModificationByteStringPayload) { - currentLogRecoveryBatch.add(((CompositeModificationByteStringPayload) data).getModification()); - } else { - LOG.error("{}: Unknown state received {} during recovery", persistenceId(), data); - } - } - - @Override - protected void applyRecoverySnapshot(final byte[] snapshotBytes) { - if(recoveryCoordinator == null) { - recoveryCoordinator = new ShardRecoveryCoordinator(persistenceId(), schemaContext, - LOG, name.toString()); - } - - recoveryCoordinator.submit(snapshotBytes, store.newWriteOnlyTransaction()); - - if(LOG.isDebugEnabled()) { - LOG.debug("{}: submitted recovery sbapshot", persistenceId()); - } + @VisibleForTesting + public RaftActorSnapshotCohort getRaftActorSnapshotCohort() { + return snapshotCohort; } @Override - protected void applyCurrentLogRecoveryBatch() { - if(recoveryCoordinator == null) { - recoveryCoordinator = new ShardRecoveryCoordinator(persistenceId(), schemaContext, - LOG, name.toString()); - } - - recoveryCoordinator.submit(currentLogRecoveryBatch, store.newWriteOnlyTransaction()); - - if(LOG.isDebugEnabled()) { - LOG.debug("{}: submitted log recovery batch with size {}", persistenceId(), - currentLogRecoveryBatch.size()); - } + @Nonnull + protected RaftActorRecoveryCohort getRaftActorRecoveryCohort() { + return new ShardRecoveryCoordinator(store, store.getSchemaContext(), + restoreFromSnapshot != null ? restoreFromSnapshot.getSnapshot() : null, persistenceId(), LOG); } @Override protected void onRecoveryComplete() { - if(recoveryCoordinator != null) { - Collection txList = recoveryCoordinator.getTransactions(); - - if(LOG.isDebugEnabled()) { - LOG.debug("{}: recovery complete - committing {} Tx's", persistenceId(), txList.size()); - } - - for(DOMStoreWriteTransaction tx: txList) { - try { - syncCommitTransaction(tx); - shardMBean.incrementCommittedTransactionCount(); - } catch (InterruptedException | ExecutionException e) { - shardMBean.incrementFailedTransactionsCount(); - LOG.error("{}: Failed to commit", persistenceId(), e); - } - } - } - - recoveryCoordinator = null; - currentLogRecoveryBatch = null; - updateJournalStats(); + restoreFromSnapshot = null; //notify shard manager getContext().parent().tell(new ActorInitialized(), getSelf()); @@ -809,34 +666,26 @@ public class Shard extends RaftActor { } @Override - protected void applyState(final ActorRef clientActor, final String identifier, final Object data) { - - if(data instanceof ModificationPayload) { - try { - applyModificationToState(clientActor, identifier, ((ModificationPayload) data).getModification()); - } catch (ClassNotFoundException | IOException e) { - LOG.error("{}: Error extracting ModificationPayload", persistenceId(), e); + protected void applyState(final ActorRef clientActor, final Identifier identifier, final Object data) { + if (data instanceof DataTreeCandidatePayload) { + if (clientActor == null) { + // No clientActor indicates a replica coming from the leader + try { + store.applyForeignCandidate(identifier, ((DataTreeCandidatePayload)data).getCandidate()); + } catch (DataValidationFailedException | IOException e) { + LOG.error("{}: Error applying replica {}", persistenceId(), identifier, e); + } + } else { + // Replication consensus reached, proceed to commit + finishCommit(clientActor, identifier); } - } - else if (data instanceof CompositeModificationPayload) { - Object modification = ((CompositeModificationPayload) data).getModification(); - - applyModificationToState(clientActor, identifier, modification); - } else if(data instanceof CompositeModificationByteStringPayload ){ - Object modification = ((CompositeModificationByteStringPayload) data).getModification(); - - applyModificationToState(clientActor, identifier, modification); } else { - LOG.error("{}: Unknown state received {} Class loader = {} CompositeNodeMod.ClassLoader = {}", - persistenceId(), data, data.getClass().getClassLoader(), - CompositeModificationPayload.class.getClassLoader()); + LOG.error("{}: Unknown state received {} ClassLoader {}", persistenceId(), data, + data.getClass().getClassLoader()); } - - updateJournalStats(); - } - private void applyModificationToState(ActorRef clientActor, String identifier, Object modification) { + private void applyModificationToState(ActorRef clientActor, Identifier identifier, Object modification) { if(modification == null) { LOG.error( "{}: modification is null - this is very unexpected, clientActor = {}, identifier = {}", @@ -851,137 +700,84 @@ public class Shard extends RaftActor { } } - private void updateJournalStats() { - ReplicatedLogEntry lastLogEntry = getLastLogEntry(); - - if (lastLogEntry != null) { - shardMBean.setLastLogIndex(lastLogEntry.getIndex()); - shardMBean.setLastLogTerm(lastLogEntry.getTerm()); - } - - shardMBean.setCommitIndex(getCommitIndex()); - shardMBean.setLastApplied(getLastApplied()); - shardMBean.setInMemoryJournalDataSize(getRaftActorContext().getReplicatedLog().dataSize()); - } - - @Override - protected void createSnapshot() { - // Create a transaction actor. We are really going to treat the transaction as a worker - // so that this actor does not get block building the snapshot. THe transaction actor will - // after processing the CreateSnapshot message. - - ActorRef createSnapshotTransaction = createTransaction( - TransactionProxy.TransactionType.READ_ONLY.ordinal(), - "createSnapshot" + ++createSnapshotTransactionCounter, "", - DataStoreVersions.CURRENT_VERSION); - - createSnapshotTransaction.tell(CreateSnapshot.INSTANCE, self()); - } - - @VisibleForTesting - @Override - protected void applySnapshot(final byte[] snapshotBytes) { - // Since this will be done only on Recovery or when this actor is a Follower - // we can safely commit everything in here. We not need to worry about event notifications - // as they would have already been disabled on the follower - - LOG.info("{}: Applying snapshot", persistenceId()); - try { - DOMStoreWriteTransaction transaction = store.newWriteOnlyTransaction(); - - NormalizedNode node = SerializationUtils.deserializeNormalizedNode(snapshotBytes); - - // delete everything first - transaction.delete(DATASTORE_ROOT); - - // Add everything from the remote node back - transaction.write(DATASTORE_ROOT, node); - syncCommitTransaction(transaction); - } catch (InterruptedException | ExecutionException e) { - LOG.error("{}: An exception occurred when applying snapshot", persistenceId(), e); - } finally { - LOG.info("{}: Done applying snapshot", persistenceId()); - } - } - @Override protected void onStateChanged() { boolean isLeader = isLeader(); - for (ActorSelection dataChangeListener : dataChangeListeners) { - dataChangeListener.tell(new EnableNotification(isLeader), getSelf()); - } + boolean hasLeader = hasLeader(); + changeSupport.onLeadershipChange(isLeader, hasLeader); + treeChangeSupport.onLeadershipChange(isLeader, hasLeader); - if(isLeader) { - for(DelayedListenerRegistration reg: delayedListenerRegistrations) { - if(!reg.isClosed()) { - reg.setDelegate(doChangeListenerRegistration(reg.getRegisterChangeListener())); - } + // If this actor is no longer the leader close all the transaction chains + if (!isLeader) { + if(LOG.isDebugEnabled()) { + LOG.debug( + "{}: onStateChanged: Closing all transaction chains because shard {} is no longer the leader", + persistenceId(), getId()); } - delayedListenerRegistrations.clear(); + store.closeAllTransactionChains(); } - shardMBean.setRaftState(getRaftState().name()); - shardMBean.setCurrentTerm(getCurrentTerm()); + if(hasLeader && !isIsolatedLeader()) { + messageRetrySupport.retryMessages(); + } + } - // If this actor is no longer the leader close all the transaction chains - if(!isLeader){ - for(Map.Entry entry : transactionChains.entrySet()){ - if(LOG.isDebugEnabled()) { - LOG.debug( - "{}: onStateChanged: Closing transaction chain {} because shard {} is no longer the leader", - persistenceId(), entry.getKey(), getId()); + @Override + protected void onLeaderChanged(String oldLeader, String newLeader) { + shardMBean.incrementLeadershipChangeCount(); + + boolean hasLeader = hasLeader(); + if(hasLeader && !isLeader()) { + // Another leader was elected. If we were the previous leader and had pending transactions, convert + // them to transaction messages and send to the new leader. + ActorSelection leader = getLeader(); + if(leader != null) { + Collection messagesToForward = commitCoordinator.convertPendingTransactionsToMessages( + datastoreContext.getShardBatchedModificationCount()); + + if(!messagesToForward.isEmpty()) { + LOG.debug("{}: Forwarding {} pending transaction messages to leader {}", persistenceId(), + messagesToForward.size(), leader); + + for(Object message: messagesToForward) { + leader.tell(message, self()); + } } - entry.getValue().close(); + } else { + commitCoordinator.abortPendingTransactions( + "The transacton was aborted due to inflight leadership change and the leader address isn't available.", + this); } + } - transactionChains.clear(); + if(hasLeader && !isIsolatedLeader()) { + messageRetrySupport.retryMessages(); } } @Override - protected DataPersistenceProvider persistence() { - return dataPersistenceProvider; + protected void pauseLeader(Runnable operation) { + LOG.debug("{}: In pauseLeader, operation: {}", persistenceId(), operation); + commitCoordinator.setRunOnPendingTransactionsComplete(operation); } - @Override protected void onLeaderChanged(final String oldLeader, final String newLeader) { - shardMBean.setLeader(newLeader); - } - - @Override public String persistenceId() { - return this.name.toString(); + @Override + public String persistenceId() { + return this.name; } @VisibleForTesting - DataPersistenceProvider getDataPersistenceProvider() { - return dataPersistenceProvider; + ShardCommitCoordinator getCommitCoordinator() { + return commitCoordinator; } - private static class ShardCreator implements Creator { - - private static final long serialVersionUID = 1L; - - final ShardIdentifier name; - final Map peerAddresses; - final DatastoreContext datastoreContext; - final SchemaContext schemaContext; - - ShardCreator(final ShardIdentifier name, final Map peerAddresses, - final DatastoreContext datastoreContext, final SchemaContext schemaContext) { - this.name = name; - this.peerAddresses = peerAddresses; - this.datastoreContext = datastoreContext; - this.schemaContext = schemaContext; - } - - @Override - public Shard create() throws Exception { - return new Shard(name, peerAddresses, datastoreContext, schemaContext); - } + public DatastoreContext getDatastoreContext() { + return datastoreContext; } @VisibleForTesting - InMemoryDOMDataStore getDataStore() { + public ShardDataTree getDataStore() { return store; } @@ -990,44 +786,110 @@ public class Shard extends RaftActor { return shardMBean; } - private static class DelayedListenerRegistration implements - ListenerRegistration>> { + public static Builder builder() { + return new Builder(); + } - private volatile boolean closed; + public static abstract class AbstractBuilder, S extends Shard> { + private final Class shardClass; + private ShardIdentifier id; + private Map peerAddresses = Collections.emptyMap(); + private DatastoreContext datastoreContext; + private SchemaContext schemaContext; + private DatastoreSnapshot.ShardSnapshot restoreFromSnapshot; + private volatile boolean sealed; - private final RegisterChangeListener registerChangeListener; + protected AbstractBuilder(Class shardClass) { + this.shardClass = shardClass; + } - private volatile ListenerRegistration>> delegate; + protected void checkSealed() { + Preconditions.checkState(!sealed, "Builder isalready sealed - further modifications are not allowed"); + } - DelayedListenerRegistration(final RegisterChangeListener registerChangeListener) { - this.registerChangeListener = registerChangeListener; + @SuppressWarnings("unchecked") + private T self() { + return (T) this; } - void setDelegate( final ListenerRegistration>> registration) { - this.delegate = registration; + public T id(ShardIdentifier id) { + checkSealed(); + this.id = id; + return self(); } - boolean isClosed() { - return closed; + public T peerAddresses(Map peerAddresses) { + checkSealed(); + this.peerAddresses = peerAddresses; + return self(); } - RegisterChangeListener getRegisterChangeListener() { - return registerChangeListener; + public T datastoreContext(DatastoreContext datastoreContext) { + checkSealed(); + this.datastoreContext = datastoreContext; + return self(); } - @Override - public AsyncDataChangeListener> getInstance() { - return delegate != null ? delegate.getInstance() : null; + public T schemaContext(SchemaContext schemaContext) { + checkSealed(); + this.schemaContext = schemaContext; + return self(); } - @Override - public void close() { - closed = true; - if(delegate != null) { - delegate.close(); + public T restoreFromSnapshot(DatastoreSnapshot.ShardSnapshot restoreFromSnapshot) { + checkSealed(); + this.restoreFromSnapshot = restoreFromSnapshot; + return self(); + } + + public ShardIdentifier getId() { + return id; + } + + public Map getPeerAddresses() { + return peerAddresses; + } + + public DatastoreContext getDatastoreContext() { + return datastoreContext; + } + + public SchemaContext getSchemaContext() { + return schemaContext; + } + + public DatastoreSnapshot.ShardSnapshot getRestoreFromSnapshot() { + return restoreFromSnapshot; + } + + public TreeType getTreeType() { + switch (datastoreContext.getLogicalStoreType()) { + case CONFIGURATION: + return TreeType.CONFIGURATION; + case OPERATIONAL: + return TreeType.OPERATIONAL; } + + throw new IllegalStateException("Unhandled logical store type " + datastoreContext.getLogicalStoreType()); + } + + protected void verify() { + Preconditions.checkNotNull(id, "id 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"); + } + + public Props props() { + sealed = true; + verify(); + return Props.create(shardClass, this); + } + } + + public static class Builder extends AbstractBuilder { + private Builder() { + super(Shard.class); } } }