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=9cb015cfaf35796bdf98692d8d98b9b67dd86a13;hp=c8f2b1b8d98a45e86a827bb7613fb94f850b4243;hb=f276ae33b951d173b51c467bb7bb1a5f5cf9a1e6;hpb=a89f83014714207a4ccb704eb75050d758266d71 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 c8f2b1b8d9..9cb015cfaf 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 @@ -12,21 +12,20 @@ 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.util.concurrent.FutureCallback; -import com.google.common.util.concurrent.Futures; -import com.google.common.util.concurrent.ListenableFuture; import java.io.IOException; -import java.util.HashMap; +import java.util.Collection; +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.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.exceptions.NoShardLeaderException; @@ -35,7 +34,6 @@ import org.opendaylight.controller.cluster.datastore.identifiers.ShardTransactio 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; @@ -44,7 +42,10 @@ 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.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.ReadyLocalTransaction; import org.opendaylight.controller.cluster.datastore.messages.RegisterChangeListener; @@ -52,24 +53,24 @@ import org.opendaylight.controller.cluster.datastore.messages.RegisterDataTreeCh 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.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.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.messages.ServerRemoved; import org.opendaylight.controller.cluster.raft.protobuff.client.messages.CompositeModificationByteStringPayload; import org.opendaylight.controller.cluster.raft.protobuff.client.messages.CompositeModificationPayload; -import org.opendaylight.yangtools.yang.data.api.schema.tree.DataTree; 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; @@ -81,14 +82,24 @@ import scala.concurrent.duration.FiniteDuration; *

*/ public class Shard extends RaftActor { - - private static final Object TX_COMMIT_TIMEOUT_CHECK_MESSAGE = "txCommitTimeoutCheck"; - @VisibleForTesting - static final Object GET_SHARD_MBEAN_MESSAGE = "getShardMBeanMessage"; + 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 ShardDataTree store; @@ -117,23 +128,30 @@ public class Shard extends RaftActor { private final DataTreeChangeListenerSupport treeChangeSupport = new DataTreeChangeListenerSupport(this); private final DataChangeListenerSupport changeSupport = new DataChangeListenerSupport(this); - protected Shard(final ShardIdentifier name, final Map peerAddresses, - final DatastoreContext datastoreContext, final SchemaContext schemaContext) { - super(name.toString(), new HashMap<>(peerAddresses), Optional.of(datastoreContext.getShardRaftConfig()), - DataStoreVersions.CURRENT_VERSION); - this.name = name.toString(); - this.datastoreContext = datastoreContext; + private ShardSnapshot restoreFromSnapshot; + + private final ShardTransactionMessageRetrySupport messageRetrySupport; + + 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(); setPersistence(datastoreContext.isPersistent()); LOG.info("Shard created : {}, persistent : {}", name, datastoreContext.isPersistent()); - store = new ShardDataTree(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.setShardActor(getSelf()); + shardMBean.setShard(this); if (isMetricsCaptureEnabled()) { getContext().become(new MeteringBehavior(this)); @@ -141,7 +159,7 @@ public class Shard extends RaftActor { commitCoordinator = new ShardCommitCoordinator(store, datastoreContext.getShardCommitQueueExpiryTimeoutInMillis(), - datastoreContext.getShardTransactionCommitQueueCapacity(), self(), LOG, this.name); + datastoreContext.getShardTransactionCommitQueueCapacity(), LOG, this.name); setTransactionCommitTimeout(); @@ -156,6 +174,8 @@ public class Shard extends RaftActor { Dispatchers.DispatcherType.Transaction), self(), getContext(), shardMBean); snapshotCohort = new ShardSnapshotCohort(transactionActorFactory, store, LOG, this.name); + + messageRetrySupport = new ShardTransactionMessageRetrySupport(this); } private void setTransactionCommitTimeout() { @@ -163,17 +183,6 @@ public class Shard extends RaftActor { datastoreContext.getShardTransactionCommitTimeoutInSeconds(), TimeUnit.SECONDS) / 2; } - 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 Optional createRoleChangeNotifier(String shardId) { ActorRef shardRoleChangeNotifier = this.getContext().actorOf( RoleChangeNotifier.getProps(shardId), shardId + "-notifier"); @@ -186,75 +195,65 @@ 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()); - } - - if (message instanceof RecoveryFailure){ - LOG.error("{}: Recovery failed because of this cause", - persistenceId(), ((RecoveryFailure) message).cause()); + protected void handleRecover(final Object message) { + LOG.debug("{}: onReceiveRecover: Received message {} from {}", persistenceId(), message.getClass(), + getSender()); - // 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 (BatchedModifications.class.isInstance(message)) { + } else if (message instanceof BatchedModifications) { handleBatchedModifications((BatchedModifications)message); } else if (message instanceof ForwardedReadyTransaction) { - commitCoordinator.handleForwardedReadyTransaction((ForwardedReadyTransaction) message, - getSender(), this); + handleForwardedReadyTransaction((ForwardedReadyTransaction) message); } else if (message instanceof ReadyLocalTransaction) { handleReadyLocalTransaction((ReadyLocalTransaction)message); - } else if (CanCommitTransaction.SERIALIZABLE_CLASS.isInstance(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) { - changeSupport.onMessage((RegisterChangeListener) message, isLeader()); + changeSupport.onMessage((RegisterChangeListener) message, isLeader(), hasLeader()); } else if (message instanceof RegisterDataTreeChangeListener) { - treeChangeSupport.onMessage((RegisterDataTreeChangeListener) message, isLeader()); + 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){ @@ -264,14 +263,30 @@ public class Shard extends RaftActor { 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 { - 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; @@ -279,12 +294,11 @@ public class Shard extends RaftActor { @Override protected LeaderStateChanged newLeaderStateChanged(String memberId, String leaderId, short leaderPayloadVersion) { - return new ShardLeaderStateChanged(memberId, leaderId, - isLeader() ? Optional.of(store.getDataTree()) : Optional.absent(), - leaderPayloadVersion); + return isLeader() ? new ShardLeaderStateChanged(memberId, leaderId, store.getDataTree(), leaderPayloadVersion) + : new ShardLeaderStateChanged(memberId, leaderId, leaderPayloadVersion); } - private void onDatastoreContext(DatastoreContext context) { + protected void onDatastoreContext(DatastoreContext context) { datastoreContext = context; commitCoordinator.setQueueCapacity(datastoreContext.getShardTransactionCommitQueueCapacity()); @@ -300,26 +314,12 @@ public class Shard extends RaftActor { updateConfigParams(datastoreContext.getShardRaftConfig()); } - private void handleTransactionCommitTimeoutCheck() { - CohortEntry cohortEntry = commitCoordinator.getCurrentCohortEntry(); - if(cohortEntry != null) { - if(cohortEntry.isExpired(transactionCommitTimeout)) { - LOG.warn("{}: Current transaction {} has timed out after {} ms - aborting", - persistenceId(), cohortEntry.getTransactionID(), transactionCommitTimeout); - - doAbortTransaction(cohortEntry.getTransactionID(), null); - } - } - - commitCoordinator.cleanupExpiredCohortEntries(); - } - private static boolean isEmptyCommit(final DataTreeCandidate candidate) { return ModificationType.UNMODIFIED.equals(candidate.getRootNode().getModificationType()); } - void continueCommit(final CohortEntry cohortEntry) throws Exception { - final DataTreeCandidate candidate = cohortEntry.getCohort().getCandidate(); + 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 @@ -328,13 +328,24 @@ public class Shard extends RaftActor { applyModificationToState(cohortEntry.getReplySender(), cohortEntry.getTransactionID(), candidate); } else { Shard.this.persistData(cohortEntry.getReplySender(), cohortEntry.getTransactionID(), - DataTreeCandidatePayload.create(candidate)); + DataTreeCandidatePayload.create(candidate)); } } private void handleCommitTransaction(final CommitTransaction commit) { - if(!commitCoordinator.handleCommit(commit.getTransactionID(), getSender(), this)) { - shardMBean.incrementFailedTransactionsCount(); + if (isLeader()) { + if(!commitCoordinator.handleCommit(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()); + } } } @@ -342,12 +353,29 @@ public class Shard extends RaftActor { 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(); + 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.toSerializable(), getSelf()); + sender.tell(CommitTransactionReply.instance(cohortEntry.getClientVersion()).toSerializable(), getSelf()); shardMBean.incrementCommittedTransactionCount(); shardMBean.setLastCommittedTransactionTime(System.currentTimeMillis()); @@ -378,13 +406,14 @@ public class Shard extends RaftActor { cohortEntry = commitCoordinator.getAndRemoveCohortEntry(transactionID); if(cohortEntry != null) { try { - store.applyForeignCandidate(transactionID, cohortEntry.getCohort().getCandidate()); + store.applyForeignCandidate(transactionID, cohortEntry.getCandidate()); } catch (DataValidationFailedException e) { shardMBean.incrementFailedTransactionsCount(); LOG.error("{}: Failed to re-apply transaction {}", persistenceId(), transactionID, e); } - sender.tell(CommitTransactionReply.INSTANCE.toSerializable(), getSelf()); + 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. @@ -401,16 +430,29 @@ public class Shard extends RaftActor { private void handleCanCommitTransaction(final CanCommitTransaction canCommit) { LOG.debug("{}: Can committing transaction {}", persistenceId(), canCommit.getTransactionID()); - commitCoordinator.handleCanCommit(canCommit.getTransactionID(), getSender(), this); + + if (isLeader()) { + commitCoordinator.handleCanCommit(canCommit.getTransactionID(), getSender(), this); + } else { + 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 noLeaderError(Object message) { - // 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()); + protected void handleBatchedModificationsLocal(BatchedModifications batched, ActorRef sender) { + try { + commitCoordinator.handleBatchedModifications(batched, sender, this); + } catch (Exception e) { + LOG.error("{}: Error handling BatchedModifications for Tx {}", persistenceId(), + batched.getTransactionID(), e); + sender.tell(new akka.actor.Status.Failure(e), getSelf()); + } } private void handleBatchedModifications(BatchedModifications batched) { @@ -426,44 +468,91 @@ public class Shard extends RaftActor { // 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 { - commitCoordinator.handleBatchedModifications(batched, getSender(), this); - } catch (Exception e) { - LOG.error("{}: Error handling BatchedModifications for Tx {}", persistenceId(), - batched.getTransactionID(), e); - getSender().tell(new akka.actor.Status.Failure(e), getSelf()); - } + boolean isLeaderActive = isLeaderActive(); + if (isLeader() && isLeaderActive) { + handleBatchedModificationsLocal(batched, getSender()); } 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()); + if (!isLeaderActive || leader == null) { + messageRetrySupport.addMessageToRetry(batched, getSender(), + "Could not commit transaction " + batched.getTransactionID()); } else { - noLeaderError(batched); + // 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 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; + } + + return false; + } + + protected boolean isIsolatedLeader() { + return getRaftState() == RaftState.IsolatedLeader; + } + private void handleReadyLocalTransaction(final ReadyLocalTransaction message) { - if (isLeader()) { + LOG.debug("{}: handleReadyLocalTransaction for {}", persistenceId(), message.getTransactionID()); + + boolean isLeaderActive = isLeaderActive(); + if (isLeader() && isLeaderActive) { try { commitCoordinator.handleReadyLocalTransaction(message, getSender(), this); } catch (Exception e) { - LOG.error("{}: Error handling LocalModifications for Tx {}", persistenceId(), + 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 (leader != null) { - LOG.debug("{}: Forwarding LocalModifications to leader {}", persistenceId(), leader); + 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()); + } + } + } + + private void handleForwardedReadyTransaction(ForwardedReadyTransaction forwardedReady) { + LOG.debug("{}: handleForwardedReadyTransaction for {}", persistenceId(), forwardedReady.getTransactionID()); + + boolean isLeaderActive = isLeaderActive(); + if (isLeader() && isLeaderActive) { + commitCoordinator.handleForwardedReadyTransaction(forwardedReady, getSender(), this); + } else { + ActorSelection leader = getLeader(); + if (!isLeaderActive || leader == null) { + messageRetrySupport.addMessageToRetry(forwardedReady, getSender(), + "Could not commit transaction " + forwardedReady.getTransactionID()); } else { - noLeaderError(message); + LOG.debug("{}: Forwarding ForwardedReadyTransaction to leader {}", persistenceId(), leader); + + ReadyLocalTransaction readyLocal = new ReadyLocalTransaction(forwardedReady.getTransactionID(), + forwardedReady.getTransaction().getSnapshot(), forwardedReady.isDoImmediateCommit()); + readyLocal.setRemoteVersion(getCurrentBehavior().getLeaderPayloadVersion()); + leader.forward(readyLocal, getContext()); } } } @@ -473,38 +562,7 @@ public class Shard extends RaftActor { } 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); - - if(sender != null) { - sender.tell(new akka.actor.Status.Failure(t), self); - } - } - }); - } + commitCoordinator.handleAbort(transactionID, sender, this); } private void handleCreateTransaction(final Object message) { @@ -513,10 +571,8 @@ public class Shard extends RaftActor { } 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()); + getSender().tell(new akka.actor.Status.Failure(new NoShardLeaderException( + "Could not create a shard transaction", persistenceId())), getSelf()); } } @@ -525,28 +581,31 @@ public class Shard extends RaftActor { } private ActorRef createTypedTransactionActor(int transactionType, - ShardTransactionIdentifier transactionId, String transactionChainId, - short clientVersion ) { + ShardTransactionIdentifier transactionId, String transactionChainId) { return transactionActorFactory.newShardTransaction(TransactionType.fromInt(transactionType), - transactionId, transactionChainId, clientVersion); + transactionId, transactionChainId); } 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) { + String transactionChainId) { ShardTransactionIdentifier transactionId = new ShardTransactionIdentifier(remoteTransactionId); @@ -556,7 +615,7 @@ public class Shard extends RaftActor { } ActorRef transactionActor = createTypedTransactionActor(transactionType, transactionId, - transactionChainId, clientVersion); + transactionChainId); return transactionActor; } @@ -589,19 +648,22 @@ public class Shard extends RaftActor { } @Override - protected RaftActorSnapshotCohort getRaftActorSnapshotCohort() { + @VisibleForTesting + public RaftActorSnapshotCohort getRaftActorSnapshotCohort() { return snapshotCohort; } @Override @Nonnull protected RaftActorRecoveryCohort getRaftActorRecoveryCohort() { - return new ShardRecoveryCoordinator(store, persistenceId(), LOG); + return new ShardRecoveryCoordinator(store, store.getSchemaContext(), + restoreFromSnapshot != null ? restoreFromSnapshot.getSnapshot() : null, persistenceId(), LOG); } @Override protected void onRecoveryComplete() { - store.recoveryDone(); + restoreFromSnapshot = null; + //notify shard manager getContext().parent().tell(new ActorInitialized(), getSelf()); @@ -630,12 +692,6 @@ public class Shard extends RaftActor { // Replication consensus reached, proceed to commit finishCommit(clientActor, identifier); } - } else 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(); @@ -669,8 +725,9 @@ public class Shard extends RaftActor { @Override protected void onStateChanged() { boolean isLeader = isLeader(); - changeSupport.onLeadershipChange(isLeader); - treeChangeSupport.onLeadershipChange(isLeader); + boolean hasLeader = hasLeader(); + changeSupport.onLeadershipChange(isLeader, hasLeader); + treeChangeSupport.onLeadershipChange(isLeader, hasLeader); // If this actor is no longer the leader close all the transaction chains if (!isLeader) { @@ -682,11 +739,49 @@ public class Shard extends RaftActor { store.closeAllTransactionChains(); } + + if(hasLeader && !isIsolatedLeader()) { + messageRetrySupport.retryMessages(); + } } @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()); + } + } + } else { + commitCoordinator.abortPendingTransactions( + "The transacton was aborted due to inflight leadership change and the leader address isn't available.", + this); + } + } + + if(hasLeader && !isIsolatedLeader()) { + messageRetrySupport.retryMessages(); + } + } + + @Override + protected void pauseLeader(Runnable operation) { + LOG.debug("{}: In pauseLeader, operation: {}", persistenceId(), operation); + commitCoordinator.setRunOnPendingTransactionsComplete(operation); } @Override @@ -699,37 +794,124 @@ public class Shard extends RaftActor { return commitCoordinator; } + public DatastoreContext getDatastoreContext() { + return datastoreContext; + } - private static class ShardCreator implements Creator { + @VisibleForTesting + public ShardDataTree getDataStore() { + return store; + } - private static final long serialVersionUID = 1L; + @VisibleForTesting + ShardStats getShardMBean() { + return shardMBean; + } + + public static Builder builder() { + return new Builder(); + } + + 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; + + protected AbstractBuilder(Class shardClass) { + this.shardClass = shardClass; + } - final ShardIdentifier name; - final Map peerAddresses; - final DatastoreContext datastoreContext; - final SchemaContext schemaContext; + protected void checkSealed() { + Preconditions.checkState(!sealed, "Builder isalready sealed - further modifications are not allowed"); + } + + @SuppressWarnings("unchecked") + private T self() { + return (T) this; + } + + public T id(ShardIdentifier id) { + checkSealed(); + this.id = id; + return self(); + } - ShardCreator(final ShardIdentifier name, final Map peerAddresses, - final DatastoreContext datastoreContext, final SchemaContext schemaContext) { - this.name = name; + public T peerAddresses(Map peerAddresses) { + checkSealed(); this.peerAddresses = peerAddresses; + return self(); + } + + public T datastoreContext(DatastoreContext datastoreContext) { + checkSealed(); this.datastoreContext = datastoreContext; + return self(); + } + + public T schemaContext(SchemaContext schemaContext) { + checkSealed(); this.schemaContext = schemaContext; + return self(); } - @Override - public Shard create() throws Exception { - return new Shard(name, peerAddresses, datastoreContext, schemaContext); + public T restoreFromSnapshot(DatastoreSnapshot.ShardSnapshot restoreFromSnapshot) { + checkSealed(); + this.restoreFromSnapshot = restoreFromSnapshot; + return self(); } - } - @VisibleForTesting - public ShardDataTree getDataStore() { - return store; + 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); + } } - @VisibleForTesting - ShardStats getShardMBean() { - return shardMBean; + public static class Builder extends AbstractBuilder { + private Builder() { + super(Shard.class); + } } }