X-Git-Url: https://git.opendaylight.org/gerrit/gitweb?a=blobdiff_plain;f=opendaylight%2Fmd-sal%2Fsal-distributed-datastore%2Fsrc%2Fmain%2Fjava%2Forg%2Fopendaylight%2Fcontroller%2Fcluster%2Fdatastore%2FShard.java;h=d53cb48e501c75c25e3532ce2c71e481717612ad;hb=68fb550b416dddd0a50e0110add0a4ae9b706758;hp=0fa27706e19382c0b84cc44b93d890ee9f0d1c8e;hpb=34bc6ec632529a0dfe419aa7404bb42a456fbc96;p=controller.git 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 0fa27706e1..81449c5747 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 @@ -10,72 +10,70 @@ package org.opendaylight.controller.cluster.datastore; import akka.actor.ActorRef; import akka.actor.ActorSelection; -import akka.actor.PoisonPill; +import akka.actor.Cancellable; import akka.actor.Props; -import akka.event.Logging; -import akka.event.LoggingAdapter; 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.util.concurrent.CheckedFuture; import com.google.common.util.concurrent.FutureCallback; import com.google.common.util.concurrent.Futures; import com.google.common.util.concurrent.ListenableFuture; -import com.google.protobuf.ByteString; -import com.google.protobuf.InvalidProtocolBufferException; +import java.io.IOException; +import java.util.HashMap; +import java.util.Map; +import java.util.concurrent.ExecutionException; +import java.util.concurrent.TimeUnit; +import javax.annotation.Nonnull; import org.opendaylight.controller.cluster.common.actor.CommonConfig; 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.BatchedModificationsReply; +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.CreateTransaction; import org.opendaylight.controller.cluster.datastore.messages.CreateTransactionReply; -import org.opendaylight.controller.cluster.datastore.messages.EnableNotification; -import org.opendaylight.controller.cluster.datastore.messages.ForwardedCommitTransaction; +import org.opendaylight.controller.cluster.datastore.messages.ForwardedReadyTransaction; import org.opendaylight.controller.cluster.datastore.messages.PeerAddressResolved; -import org.opendaylight.controller.cluster.datastore.messages.ReadData; -import org.opendaylight.controller.cluster.datastore.messages.ReadDataReply; +import org.opendaylight.controller.cluster.datastore.messages.ReadyTransactionReply; 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.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.node.NormalizedNodeToNodeCodec; -import org.opendaylight.controller.cluster.raft.ConfigParams; -import org.opendaylight.controller.cluster.raft.DefaultConfigParamsImpl; +import org.opendaylight.controller.cluster.datastore.utils.Dispatchers; +import org.opendaylight.controller.cluster.datastore.utils.MessageTracker; +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.base.messages.CaptureSnapshotReply; +import org.opendaylight.controller.cluster.raft.RaftActorRecoveryCohort; +import org.opendaylight.controller.cluster.raft.RaftActorSnapshotCohort; +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.md.sal.common.api.data.AsyncDataChangeListener; -import org.opendaylight.controller.md.sal.common.api.data.ReadFailedException; import org.opendaylight.controller.md.sal.dom.store.impl.InMemoryDOMDataStore; import org.opendaylight.controller.md.sal.dom.store.impl.InMemoryDOMDataStoreFactory; -import org.opendaylight.controller.protobuff.messages.common.NormalizedNodeMessages; -import org.opendaylight.controller.sal.core.spi.data.DOMStoreReadTransaction; -import org.opendaylight.controller.sal.core.spi.data.DOMStoreThreePhaseCommitCohort; -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.yangtools.yang.model.api.SchemaContext; +import scala.concurrent.duration.Duration; import scala.concurrent.duration.FiniteDuration; -import java.util.ArrayList; -import java.util.HashMap; -import java.util.List; -import java.util.Map; -import java.util.concurrent.ExecutionException; -import java.util.concurrent.TimeUnit; - /** * A Shard represents a portion of the logical data tree
*

@@ -84,84 +82,99 @@ import java.util.concurrent.TimeUnit; */ public class Shard extends RaftActor { - private static final ConfigParams configParams = new ShardConfigParams(); + private static final Object TX_COMMIT_TIMEOUT_CHECK_MESSAGE = "txCommitTimeoutCheck"; - public static final String DEFAULT_NAME = "default"; + @VisibleForTesting + static final String DEFAULT_NAME = "default"; // The state of this Shard private final InMemoryDOMDataStore store; - private final Map - modificationToCohort = new HashMap<>(); + /// The name of this shard + private final String name; + + private final ShardStats shardMBean; - private final LoggingAdapter LOG = - Logging.getLogger(getContext().system(), this); + private DatastoreContext datastoreContext; - // By default persistent will be true and can be turned off using the system - // property shard.persistent - private final boolean persistent; + private final ShardCommitCoordinator commitCoordinator; - /// The name of this shard - private final ShardIdentifier name; + private long transactionCommitTimeout; - private final ShardStats shardMBean; + private Cancellable txCommitTimeoutCheckSchedule; - private final List dataChangeListeners = new ArrayList<>(); + private final Optional roleChangeNotifier; - private final DatastoreContext datastoreContext; + private final MessageTracker appendEntriesReplyTracker; - private SchemaContext schemaContext; + private final ReadyTransactionReply READY_TRANSACTION_REPLY = new ReadyTransactionReply( + Serialization.serializedActorPath(getSelf())); - private ActorRef createSnapshotTransaction; + private final DOMTransactionFactory domTransactionFactory; - private final Map transactionChains = new HashMap<>(); + private final ShardTransactionActorFactory transactionActorFactory; - private Shard(ShardIdentifier name, Map peerAddresses, - DatastoreContext datastoreContext, SchemaContext schemaContext) { - super(name.toString(), mapPeerAddresses(peerAddresses), Optional.of(configParams)); + private final ShardSnapshotCohort snapshotCohort; - this.name = name; - this.datastoreContext = datastoreContext; - this.schemaContext = schemaContext; + private final DataTreeChangeListenerSupport treeChangeSupport = new DataTreeChangeListenerSupport(this); + private final DataChangeListenerSupport changeSupport = new DataChangeListenerSupport(this); - String setting = System.getProperty("shard.persistent"); + protected Shard(final ShardIdentifier name, final Map peerAddresses, + final DatastoreContext datastoreContext, final SchemaContext schemaContext) { + super(name.toString(), new HashMap<>(peerAddresses), Optional.of(datastoreContext.getShardRaftConfig())); - this.persistent = !"false".equals(setting); + this.name = name.toString(); + this.datastoreContext = datastoreContext; + + setPersistence(datastoreContext.isPersistent()); - LOG.info("Shard created : {} persistent : {}", name, persistent); + LOG.info("Shard created : {}, persistent : {}", name, datastoreContext.isPersistent()); store = InMemoryDOMDataStoreFactory.create(name.toString(), null, datastoreContext.getDataStoreProperties()); - if(schemaContext != null) { + if (schemaContext != null) { store.onGlobalContextUpdated(schemaContext); } shardMBean = ShardMBeanFactory.getShardStatsMBean(name.toString(), datastoreContext.getDataStoreMXBeanType()); - shardMBean.setDataStoreExecutor(store.getDomStoreExecutor()); shardMBean.setNotificationManager(store.getDataChangeListenerNotificationManager()); + shardMBean.setShardActor(getSelf()); if (isMetricsCaptureEnabled()) { getContext().become(new MeteringBehavior(this)); } - } - private static Map mapPeerAddresses( - Map peerAddresses) { - Map map = new HashMap<>(); + domTransactionFactory = new DOMTransactionFactory(store, shardMBean, LOG, this.name); - for (Map.Entry entry : peerAddresses - .entrySet()) { - map.put(entry.getKey().toString(), entry.getValue()); - } + commitCoordinator = new ShardCommitCoordinator(domTransactionFactory, + TimeUnit.SECONDS.convert(5, TimeUnit.MINUTES), + datastoreContext.getShardTransactionCommitQueueCapacity(), self(), LOG, this.name); + + setTransactionCommitTimeout(); - return map; + // create a notifier actor for each cluster member + roleChangeNotifier = createRoleChangeNotifier(name.toString()); + + appendEntriesReplyTracker = new MessageTracker(AppendEntriesReply.class, + getRaftActorContext().getConfigParams().getIsolatedCheckIntervalInMillis()); + + transactionActorFactory = new ShardTransactionActorFactory(domTransactionFactory, datastoreContext, + new Dispatchers(context().system().dispatchers()).getDispatcherPath( + Dispatchers.DispatcherType.Transaction), self(), getContext(), shardMBean); + + snapshotCohort = new ShardSnapshotCohort(transactionActorFactory, store, LOG, this.name); + } + + private void setTransactionCommitTimeout() { + transactionCommitTimeout = TimeUnit.MILLISECONDS.convert( + datastoreContext.getShardTransactionCommitTimeoutInSeconds(), TimeUnit.SECONDS); } public static Props props(final ShardIdentifier name, - final Map peerAddresses, - DatastoreContext datastoreContext, SchemaContext schemaContext) { + 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"); @@ -170,422 +183,546 @@ public class Shard extends RaftActor { return Props.create(new ShardCreator(name, peerAddresses, datastoreContext, schemaContext)); } - @Override public void onReceiveRecover(Object message) { + private Optional createRoleChangeNotifier(String shardId) { + ActorRef shardRoleChangeNotifier = this.getContext().actorOf( + RoleChangeNotifier.getProps(shardId), shardId + "-notifier"); + return Optional.of(shardRoleChangeNotifier); + } + + @Override + public void postStop() { + LOG.info("Stopping Shard {}", persistenceId()); + + super.postStop(); + + if(txCommitTimeoutCheckSchedule != null) { + txCommitTimeoutCheckSchedule.cancel(); + } + + shardMBean.unregisterMBean(); + } + + @Override + public void onReceiveRecover(final Object message) throws Exception { if(LOG.isDebugEnabled()) { - LOG.debug("onReceiveRecover: Received message {} from {}", - message.getClass().toString(), - getSender()); + LOG.debug("{}: onReceiveRecover: Received message {} from {}", persistenceId(), + message.getClass().toString(), getSender()); } if (message instanceof RecoveryFailure){ - LOG.error(((RecoveryFailure) message).cause(), "Recovery failed because of this cause"); + 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(Object message) { - if(LOG.isDebugEnabled()) { - LOG.debug("onReceiveCommand: Received message {} from {}", - message.getClass().toString(), - getSender()); + @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()); } - if(message.getClass().equals(ReadDataReply.SERIALIZABLE_CLASS)) { - // This must be for install snapshot. Don't want to open this up and trigger - // deSerialization - self() - .tell(new CaptureSnapshotReply(ReadDataReply.getNormalizedNodeByteString(message)), - self()); - - createSnapshotTransaction = null; - // Send a PoisonPill instead of sending close transaction because we do not really need - // a response - getSender().tell(PoisonPill.getInstance(), self()); - - } 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 ForwardedCommitTransaction) { - handleForwardedCommit((ForwardedCommitTransaction) message); - } else if (message.getClass() - .equals(CreateTransaction.SERIALIZABLE_CLASS)) { - if (isLeader()) { - createTransaction(CreateTransaction.fromSerializable(message)); - } else if (getLeader() != null) { - getLeader().forward(message, getContext()); + try { + if (CreateTransaction.SERIALIZABLE_CLASS.isInstance(message)) { + handleCreateTransaction(message); + } else if (BatchedModifications.class.isInstance(message)) { + handleBatchedModifications((BatchedModifications)message); + } else if (message instanceof ForwardedReadyTransaction) { + handleForwardedReadyTransaction((ForwardedReadyTransaction) message); + } else if (CanCommitTransaction.SERIALIZABLE_CLASS.isInstance(message)) { + handleCanCommitTransaction(CanCommitTransaction.fromSerializable(message)); + } else if (CommitTransaction.SERIALIZABLE_CLASS.isInstance(message)) { + handleCommitTransaction(CommitTransaction.fromSerializable(message)); + } else if (AbortTransaction.SERIALIZABLE_CLASS.isInstance(message)) { + handleAbortTransaction(AbortTransaction.fromSerializable(message)); + } else if (CloseTransactionChain.SERIALIZABLE_CLASS.isInstance(message)) { + closeTransactionChain(CloseTransactionChain.fromSerializable(message)); + } else if (message instanceof RegisterChangeListener) { + changeSupport.onMessage((RegisterChangeListener) message, isLeader()); + } else if (message instanceof RegisterDataTreeChangeListener) { + treeChangeSupport.onMessage((RegisterDataTreeChangeListener) message, isLeader()); + } 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(message instanceof DatastoreContext) { + onDatastoreContext((DatastoreContext)message); + } else if(message instanceof RegisterRoleChangeListener){ + roleChangeNotifier.get().forward(message, context()); + } else if (message instanceof FollowerInitialSyncUpStatus){ + shardMBean.setFollowerInitialSyncStatus(((FollowerInitialSyncUpStatus) message).isInitialSyncDone()); + context().parent().tell(message, self()); } else { - getSender().tell(new akka.actor.Status.Failure(new IllegalStateException( - "Could not find leader so transaction cannot be created")), getSelf()); + super.onReceiveCommand(message); } - } else if (message instanceof PeerAddressResolved) { - PeerAddressResolved resolved = (PeerAddressResolved) message; - setPeerAddress(resolved.getPeerId().toString(), - resolved.getPeerAddress()); - } else { - super.onReceiveCommand(message); + } finally { + context.done(); } } - private void closeTransactionChain(CloseTransactionChain closeTransactionChain) { - DOMStoreTransactionChain chain = - transactionChains.remove(closeTransactionChain.getTransactionChainId()); + @Override + protected Optional getRoleChangeNotifier() { + return roleChangeNotifier; + } + + private void onDatastoreContext(DatastoreContext context) { + datastoreContext = context; + + commitCoordinator.setQueueCapacity(datastoreContext.getShardTransactionCommitQueueCapacity()); - if(chain != null) { - chain.close(); + setTransactionCommitTimeout(); + + if(datastoreContext.isPersistent() && !persistence().isRecoveryApplicable()) { + setPersistence(true); + } else if(!datastoreContext.isPersistent() && persistence().isRecoveryApplicable()) { + setPersistence(false); } - } - private ActorRef createTypedTransactionActor( - int transactionType, - ShardTransactionIdentifier transactionId, - String transactionChainId ) { + updateConfigParams(datastoreContext.getShardRaftConfig()); + } - DOMStoreTransactionFactory factory = store; + 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); - if(!transactionChainId.isEmpty()) { - factory = transactionChains.get(transactionChainId); - if(factory == null){ - DOMStoreTransactionChain transactionChain = store.createTransactionChain(); - transactionChains.put(transactionChainId, transactionChain); - factory = transactionChain; + doAbortTransaction(cohortEntry.getTransactionID(), null); } } + } - if(this.schemaContext == null){ - throw new NullPointerException("schemaContext should not be 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; } - if (transactionType - == TransactionProxy.TransactionType.READ_ONLY.ordinal()) { + // 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. - shardMBean.incrementReadOnlyTransactionCount(); + 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 + // or if cohortEntry has no modifications + // we can apply modification to the state immediately + if((!hasFollowers() && !persistence().isRecoveryApplicable()) || (!cohortEntry.hasModifications())){ + 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()); + } - return getContext().actorOf( - ShardTransaction.props(factory.newReadOnlyTransaction(), getSelf(), - schemaContext,datastoreContext, shardMBean), transactionId.toString()); + cohortEntry.updateLastAccessTime(); + } - } else if (transactionType - == TransactionProxy.TransactionType.READ_WRITE.ordinal()) { + 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()); + } - shardMBean.incrementReadWriteTransactionCount(); + return; + } - return getContext().actorOf( - ShardTransaction.props(factory.newReadWriteTransaction(), getSelf(), - schemaContext, datastoreContext, shardMBean), transactionId.toString()); + 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(); - } else if (transactionType - == TransactionProxy.TransactionType.WRITE_ONLY.ordinal()) { + sender.tell(CommitTransactionReply.INSTANCE.toSerializable(), getSelf()); - shardMBean.incrementWriteOnlyTransactionCount(); + shardMBean.incrementCommittedTransactionCount(); + shardMBean.setLastCommittedTransactionTime(System.currentTimeMillis()); - return getContext().actorOf( - ShardTransaction.props(factory.newWriteOnlyTransaction(), getSelf(), - schemaContext, datastoreContext, shardMBean), transactionId.toString()); - } else { - throw new IllegalArgumentException( - "Shard="+name + ":CreateTransaction message has unidentified transaction type=" - + transactionType); + } 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 createTransaction(CreateTransaction createTransaction) { - createTransaction(createTransaction.getTransactionType(), - createTransaction.getTransactionId(), createTransaction.getTransactionChainId()); + private void handleCanCommitTransaction(final CanCommitTransaction canCommit) { + LOG.debug("{}: Can committing transaction {}", persistenceId(), canCommit.getTransactionID()); + commitCoordinator.handleCanCommit(canCommit, getSender(), self()); } - private ActorRef createTransaction(int transactionType, String remoteTransactionId, String transactionChainId) { - - ShardTransactionIdentifier transactionId = - ShardTransactionIdentifier.builder() - .remoteTransactionId(remoteTransactionId) - .build(); - if(LOG.isDebugEnabled()) { - LOG.debug("Creating transaction : {} ", transactionId); + private void handleBatchedModifications(BatchedModifications batched) { + // This message is sent to prepare the modificationsa 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 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. + // + if(isLeader()) { + try { + boolean ready = commitCoordinator.handleTransactionModifications(batched); + if(ready) { + sender().tell(READY_TRANSACTION_REPLY, self()); + } else { + sender().tell(new BatchedModificationsReply(batched.getModifications().size()), self()); + } + } catch (Exception e) { + LOG.error("{}: Error handling BatchedModifications for Tx {}", persistenceId(), + batched.getTransactionID(), e); + getSender().tell(new akka.actor.Status.Failure(e), getSelf()); + } + } else { + ActorSelection leader = getLeader(); + if(leader != null) { + // TODO: what if this is not the first batch and leadership changed in between batched messages? + // We could check if the commitCoordinator already has a cached entry and forward all the previous + // batched modifications. + LOG.debug("{}: Forwarding BatchedModifications to leader {}", persistenceId(), leader); + leader.forward(batched, getContext()); + } else { + // TODO: rather than throwing an immediate exception, we could schedule a timer to try again to make + // it more resilient in case we're in the process of electing a new leader. + getSender().tell(new akka.actor.Status.Failure(new NoShardLeaderException(String.format( + "Could not find the leader for shard %s. This typically happens" + + " when the system is coming up or recovering and a leader is being elected. Try again" + + " later.", persistenceId()))), getSelf()); + } } - ActorRef transactionActor = - createTypedTransactionActor(transactionType, transactionId, transactionChainId); + } - getSender() - .tell(new CreateTransactionReply( - Serialization.serializedActorPath(transactionActor), - remoteTransactionId).toSerializable(), - getSelf()); + 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(), + (MutableCompositeModification) 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.LITHIUM_VERSION) { + ActorRef replyActorPath = getSelf(); + if(ready.getTxnClientVersion() < DataStoreVersions.HELIUM_1_VERSION) { + LOG.debug("{}: Creating BackwardsCompatibleThreePhaseCommitCohort", persistenceId()); + replyActorPath = getContext().actorOf(BackwardsCompatibleThreePhaseCommitCohort.props( + ready.getTransactionID())); + } - return transactionActor; + ReadyTransactionReply readyTransactionReply = + new ReadyTransactionReply(Serialization.serializedActorPath(replyActorPath), + ready.getTxnClientVersion()); + getSender().tell(ready.isReturnSerialized() ? readyTransactionReply.toSerializable() : + readyTransactionReply, getSelf()); + } else { + getSender().tell(READY_TRANSACTION_REPLY, getSelf()); + } } - private void syncCommitTransaction(DOMStoreWriteTransaction transaction) - throws ExecutionException, InterruptedException { - DOMStoreThreePhaseCommitCohort commitCohort = transaction.ready(); - commitCohort.preCommit().get(); - commitCohort.commit().get(); + 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); - private void commit(final ActorRef sender, Object serialized) { - Modification modification = MutableCompositeModification - .fromSerializable(serialized, schemaContext); - DOMStoreThreePhaseCommitCohort cohort = - modificationToCohort.remove(serialized); - if (cohort == null) { + // 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); - if(LOG.isDebugEnabled()) { - LOG.debug( - "Could not find cohort for modification : {}. Writing modification using a new transaction", - modification); - } - - DOMStoreWriteTransaction transaction = - store.newWriteOnlyTransaction(); + final ListenableFuture future = cohortEntry.getCohort().abort(); + final ActorRef self = getSelf(); - if(LOG.isDebugEnabled()) { - LOG.debug("Created new transaction {}", transaction.getIdentifier().toString()); - } + Futures.addCallback(future, new FutureCallback() { + @Override + public void onSuccess(final Void v) { + shardMBean.incrementAbortTransactionsCount(); - modification.apply(transaction); - try { - syncCommitTransaction(transaction); - } catch (InterruptedException | ExecutionException e) { - shardMBean.incrementFailedTransactionsCount(); - LOG.error("Failed to commit", e); - return; - } - //we want to just apply the recovery commit and return - shardMBean.incrementCommittedTransactionCount(); - return; - } + 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){ - LOG.error("Commit failed. Sender cannot be null"); - return; + if(sender != null) { + sender.tell(new akka.actor.Status.Failure(t), self); + } + } + }); } - - ListenableFuture future = cohort.commit(); - - Futures.addCallback(future, new FutureCallback() { - @Override - public void onSuccess(Void v) { - sender.tell(new CommitTransactionReply().toSerializable(), getSelf()); - shardMBean.incrementCommittedTransactionCount(); - shardMBean.setLastCommittedTransactionTime(System.currentTimeMillis()); - } - - @Override - public void onFailure(Throwable t) { - LOG.error(t, "An exception happened during commit"); - shardMBean.incrementFailedTransactionsCount(); - sender.tell(new akka.actor.Status.Failure(t), getSelf()); - } - }); - } - private void handleForwardedCommit(ForwardedCommitTransaction message) { - Object serializedModification = - message.getModification().toSerializable(); - - modificationToCohort - .put(serializedModification, message.getCohort()); - - if (persistent) { - this.persistData(getSender(), "identifier", - new CompositeModificationPayload(serializedModification)); + private void handleCreateTransaction(final Object message) { + if (isLeader()) { + createTransaction(CreateTransaction.fromSerializable(message)); + } else if (getLeader() != null) { + getLeader().forward(message, getContext()); } else { - this.commit(getSender(), serializedModification); + 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 updateSchemaContext(UpdateSchemaContext message) { - this.schemaContext = message.getSchemaContext(); - updateSchemaContext(message.getSchemaContext()); - store.onGlobalContextUpdated(message.getSchemaContext()); + private void closeTransactionChain(final CloseTransactionChain closeTransactionChain) { + domTransactionFactory.closeTransactionChain(closeTransactionChain.getTransactionChainId()); } - @VisibleForTesting void updateSchemaContext(SchemaContext schemaContext) { - store.onGlobalContextUpdated(schemaContext); - } + private ActorRef createTypedTransactionActor(int transactionType, + ShardTransactionIdentifier transactionId, String transactionChainId, + short clientVersion ) { - private void registerChangeListener( - RegisterChangeListener registerChangeListener) { + return transactionActorFactory.newShardTransaction(TransactionProxy.TransactionType.fromInt(transactionType), + transactionId, transactionChainId, clientVersion); + } - if(LOG.isDebugEnabled()) { - LOG.debug("registerDataChangeListener for {}", registerChangeListener - .getPath()); + private void createTransaction(CreateTransaction createTransaction) { + try { + ActorRef transactionActor = createTransaction(createTransaction.getTransactionType(), + createTransaction.getTransactionId(), createTransaction.getTransactionChainId(), + createTransaction.getVersion()); + + getSender().tell(new CreateTransactionReply(Serialization.serializedActorPath(transactionActor), + createTransaction.getTransactionId()).toSerializable(), getSelf()); + } catch (Exception e) { + getSender().tell(new akka.actor.Status.Failure(e), getSelf()); } + } - - ActorSelection dataChangeListenerPath = getContext() - .system().actorSelection( - registerChangeListener.getDataChangeListenerPath()); + private ActorRef createTransaction(int transactionType, String remoteTransactionId, + String transactionChainId, short clientVersion) { - // 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(isLeader()), getSelf()); + ShardTransactionIdentifier transactionId = new ShardTransactionIdentifier(remoteTransactionId); - // 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); + if(LOG.isDebugEnabled()) { + LOG.debug("{}: Creating transaction : {} ", persistenceId(), transactionId); + } - AsyncDataChangeListener> - listener = new DataChangeListenerProxy(schemaContext, dataChangeListenerPath); + ActorRef transactionActor = createTypedTransactionActor(transactionType, transactionId, + transactionChainId, clientVersion); - ListenerRegistration>> - registration = store.registerChangeListener(registerChangeListener.getPath(), - listener, registerChangeListener.getScope()); - ActorRef listenerRegistration = - getContext().actorOf( - DataChangeListenerRegistration.props(registration)); + return transactionActor; + } - if(LOG.isDebugEnabled()) { - LOG.debug( - "registerDataChangeListener sending reply, listenerRegistrationPath = {} " - , listenerRegistration.path().toString()); + private void commitWithNewTransaction(final Modification modification) { + DOMStoreWriteTransaction tx = store.newWriteOnlyTransaction(); + modification.apply(tx); + try { + snapshotCohort.syncCommitTransaction(tx); + shardMBean.incrementCommittedTransactionCount(); + shardMBean.setLastCommittedTransactionTime(System.currentTimeMillis()); + } catch (InterruptedException | ExecutionException e) { + shardMBean.incrementFailedTransactionsCount(); + LOG.error("{}: Failed to commit", persistenceId(), e); } + } - getSender() - .tell(new RegisterChangeListenerReply(listenerRegistration.path()), - getSelf()); + private void updateSchemaContext(final UpdateSchemaContext message) { + updateSchemaContext(message.getSchemaContext()); } - private boolean isMetricsCaptureEnabled(){ + @VisibleForTesting + void updateSchemaContext(final SchemaContext schemaContext) { + store.onGlobalContextUpdated(schemaContext); + } + + private boolean isMetricsCaptureEnabled() { CommonConfig config = new CommonConfig(getContext().system().settings().config()); return config.isMetricCaptureEnabled(); } - @Override protected void applyState(ActorRef clientActor, String identifier, - Object data) { - - if (data instanceof CompositeModificationPayload) { - Object modification = - ((CompositeModificationPayload) data).getModification(); - - if (modification != null) { - commit(clientActor, modification); - } else { - LOG.error( - "modification is null - this is very unexpected, clientActor = {}, identifier = {}", - identifier, clientActor.path().toString()); - } - - } else { - LOG.error("Unknown state received {} Class loader = {} CompositeNodeMod.ClassLoader = {}", data, data.getClass().getClassLoader(), CompositeModificationPayload.class.getClassLoader()); - } + @Override + protected RaftActorSnapshotCohort getRaftActorSnapshotCohort() { + return snapshotCohort; + } - // Update stats - ReplicatedLogEntry lastLogEntry = getLastLogEntry(); + @Override + @Nonnull + protected RaftActorRecoveryCohort getRaftActorRecoveryCohort() { + return new ShardRecoveryCoordinator(store, persistenceId(), LOG); + } - if (lastLogEntry != null) { - shardMBean.setLastLogIndex(lastLogEntry.getIndex()); - shardMBean.setLastLogTerm(lastLogEntry.getTerm()); + @Override + protected void onRecoveryComplete() { + //notify shard manager + getContext().parent().tell(new ActorInitialized(), getSelf()); + + // Being paranoid here - this method should only be called once but just in case... + if(txCommitTimeoutCheckSchedule == null) { + // Schedule a message to be periodically sent to check if the current in-progress + // transaction should be expired and aborted. + FiniteDuration period = Duration.create(transactionCommitTimeout / 3, TimeUnit.MILLISECONDS); + txCommitTimeoutCheckSchedule = getContext().system().scheduler().schedule( + period, period, getSelf(), + TX_COMMIT_TIMEOUT_CHECK_MESSAGE, getContext().dispatcher(), ActorRef.noSender()); } - - shardMBean.setCommitIndex(getCommitIndex()); - shardMBean.setLastApplied(getLastApplied()); - } - @Override protected void createSnapshot() { - if (createSnapshotTransaction == null) { - - // Create a transaction. We are really going to treat the transaction as a worker - // so that this actor does not get block building the snapshot - createSnapshotTransaction = createTransaction( - TransactionProxy.TransactionType.READ_ONLY.ordinal(), - "createSnapshot", ""); - - createSnapshotTransaction.tell( - new ReadData(YangInstanceIdentifier.builder().build()).toSerializable(), self()); + @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); + } } - } + else if (data instanceof CompositeModificationPayload) { + Object modification = ((CompositeModificationPayload) data).getModification(); - @VisibleForTesting @Override protected void applySnapshot(ByteString snapshot) { - // 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 + applyModificationToState(clientActor, identifier, modification); + } else if(data instanceof CompositeModificationByteStringPayload ){ + Object modification = ((CompositeModificationByteStringPayload) data).getModification(); - LOG.info("Applying snapshot"); - try { - DOMStoreWriteTransaction transaction = store.newWriteOnlyTransaction(); - NormalizedNodeMessages.Node serializedNode = NormalizedNodeMessages.Node.parseFrom(snapshot); - NormalizedNode node = new NormalizedNodeToNodeCodec(schemaContext) - .decode(YangInstanceIdentifier.builder().build(), serializedNode); - - // delete everything first - transaction.delete(YangInstanceIdentifier.builder().build()); - - // Add everything from the remote node back - transaction.write(YangInstanceIdentifier.builder().build(), node); - syncCommitTransaction(transaction); - } catch (InvalidProtocolBufferException | InterruptedException | ExecutionException e) { - LOG.error(e, "An exception occurred when applying snapshot"); - } finally { - LOG.info("Done applying snapshot"); + applyModificationToState(clientActor, identifier, modification); + } else { + LOG.error("{}: Unknown state received {} Class loader = {} CompositeNodeMod.ClassLoader = {}", + persistenceId(), data, data.getClass().getClassLoader(), + CompositeModificationPayload.class.getClassLoader()); } } - @Override protected void onStateChanged() { - for (ActorSelection dataChangeListener : dataChangeListeners) { - dataChangeListener - .tell(new EnableNotification(isLeader()), getSelf()); + private void applyModificationToState(ActorRef clientActor, String identifier, Object modification) { + if(modification == null) { + LOG.error( + "{}: modification is null - this is very unexpected, clientActor = {}, identifier = {}", + persistenceId(), identifier, clientActor != null ? clientActor.path().toString() : null); + } else if(clientActor == null) { + // There's no clientActor to which to send a commit reply so we must be applying + // replicated state from the leader. + commitWithNewTransaction(MutableCompositeModification.fromSerializable(modification)); + } else { + // This must be the OK to commit after replication consensus. + finishCommit(clientActor, identifier); } + } - shardMBean.setRaftState(getRaftState().name()); - shardMBean.setCurrentTerm(getCurrentTerm()); + @Override + protected void onStateChanged() { + boolean isLeader = isLeader(); + changeSupport.onLeadershipChange(isLeader); + treeChangeSupport.onLeadershipChange(isLeader); // 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", - entry.getKey(), getId()); - } - entry.getValue().close(); + if (!isLeader) { + if(LOG.isDebugEnabled()) { + LOG.debug( + "{}: onStateChanged: Closing all transaction chains because shard {} is no longer the leader", + persistenceId(), getId()); } - transactionChains.clear(); + domTransactionFactory.closeAllTransactionChains(); } } - @Override protected void onLeaderChanged(String oldLeader, String newLeader) { - shardMBean.setLeader(newLeader); + @Override + public String persistenceId() { + return this.name; } - @Override public String persistenceId() { - return this.name.toString(); + @VisibleForTesting + ShardCommitCoordinator getCommitCoordinator() { + return commitCoordinator; } - private static class ShardConfigParams extends DefaultConfigParamsImpl { - public static final FiniteDuration HEART_BEAT_INTERVAL = - new FiniteDuration(500, TimeUnit.MILLISECONDS); - - @Override public FiniteDuration getHeartBeatInterval() { - return HEART_BEAT_INTERVAL; - } - } - private static class ShardCreator implements Creator { private static final long serialVersionUID = 1L; final ShardIdentifier name; - final Map peerAddresses; + final Map peerAddresses; final DatastoreContext datastoreContext; final SchemaContext schemaContext; - ShardCreator(ShardIdentifier name, Map peerAddresses, - DatastoreContext datastoreContext, SchemaContext schemaContext) { + ShardCreator(final ShardIdentifier name, final Map peerAddresses, + final DatastoreContext datastoreContext, final SchemaContext schemaContext) { this.name = name; this.peerAddresses = peerAddresses; this.datastoreContext = datastoreContext; @@ -598,26 +735,13 @@ public class Shard extends RaftActor { } } - @VisibleForTesting NormalizedNode readStore() throws ExecutionException, InterruptedException { - DOMStoreReadTransaction transaction = store.newReadOnlyTransaction(); - - CheckedFuture>, ReadFailedException> future = - transaction.read(YangInstanceIdentifier.builder().build()); - - NormalizedNode node = future.get().get(); - - transaction.close(); - - return node; + @VisibleForTesting + public InMemoryDOMDataStore getDataStore() { + return store; } - @VisibleForTesting void writeToStore(YangInstanceIdentifier id, NormalizedNode node) - throws ExecutionException, InterruptedException { - DOMStoreWriteTransaction transaction = store.newWriteOnlyTransaction(); - - transaction.write(id, node); - - syncCommitTransaction(transaction); + @VisibleForTesting + ShardStats getShardMBean() { + return shardMBean; } - }