- * 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 String name;
private final ShardStats shardMBean;
- private final List dataChangeListeners = Lists.newArrayList();
-
- private final List delayedListenerRegistrations =
- Lists.newArrayList();
-
private DatastoreContext datastoreContext;
- private SchemaContext schemaContext;
-
- private int createSnapshotTransactionCounter;
-
private final ShardCommitCoordinator commitCoordinator;
private long transactionCommitTimeout;
@@ -128,54 +118,45 @@ 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 final DataTreeChangeListenerSupport treeChangeSupport = new DataTreeChangeListenerSupport(this);
+ private final DataChangeListenerSupport changeSupport = new DataChangeListenerSupport(this);
- private final DOMTransactionFactory transactionFactory;
- private final String txnDispatcherPath;
+ private ShardSnapshot restoreFromSnapshot;
- protected Shard(final ShardIdentifier name, final Map peerAddresses,
- final DatastoreContext datastoreContext, final SchemaContext schemaContext) {
- super(name.toString(), new HashMap<>(peerAddresses), Optional.of(datastoreContext.getShardRaftConfig()));
+ private final ShardTransactionMessageRetrySupport messageRetrySupport;
- this.name = name.toString();
- this.datastoreContext = datastoreContext;
- this.schemaContext = schemaContext;
- 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();
setPersistence(datastoreContext.isPersistent());
LOG.info("Shard created : {}, persistent : {}", name, datastoreContext.isPersistent());
- store = InMemoryDOMDataStoreFactory.create(name.toString(), null,
- datastoreContext.getDataStoreProperties());
-
- 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.setShardActor(getSelf());
+ shardMBean.setShard(this);
if (isMetricsCaptureEnabled()) {
getContext().become(new MeteringBehavior(this));
}
- transactionFactory = new DOMTransactionFactory(store, shardMBean, LOG, this.name);
-
- commitCoordinator = new ShardCommitCoordinator(transactionFactory,
- TimeUnit.SECONDS.convert(5, TimeUnit.MINUTES),
- datastoreContext.getShardTransactionCommitQueueCapacity(), self(), LOG, this.name);
+ commitCoordinator = new ShardCommitCoordinator(store,
+ datastoreContext.getShardCommitQueueExpiryTimeoutInMillis(),
+ datastoreContext.getShardTransactionCommitQueueCapacity(), LOG, this.name);
setTransactionCommitTimeout();
@@ -185,23 +166,19 @@ public class Shard extends RaftActor {
appendEntriesReplyTracker = new MessageTracker(AppendEntriesReply.class,
getRaftActorContext().getConfigParams().getIsolatedCheckIntervalInMillis());
- recoveryCoordinator = new ShardRecoveryCoordinator(store, persistenceId(), LOG);
+ transactionActorFactory = new ShardTransactionActorFactory(store, datastoreContext,
+ new Dispatchers(context().system().dispatchers()).getDispatcherPath(
+ Dispatchers.DispatcherType.Transaction), self(), getContext(), shardMBean);
+
+ snapshotCohort = new ShardSnapshotCohort(builder.getId().getMemberName(), transactionActorFactory, store,
+ LOG, this.name);
+
+ messageRetrySupport = new ShardTransactionMessageRetrySupport(this);
}
private void setTransactionCommitTimeout() {
transactionCommitTimeout = TimeUnit.MILLISECONDS.convert(
- datastoreContext.getShardTransactionCommitTimeoutInSeconds(), TimeUnit.SECONDS);
- }
-
- 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));
+ datastoreContext.getShardTransactionCommitTimeoutInSeconds(), TimeUnit.SECONDS) / 2;
}
private Optional createRoleChangeNotifier(String shardId) {
@@ -216,91 +193,113 @@ 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 (BatchedModifications.class.isInstance(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());
@@ -316,75 +315,90 @@ public class Shard extends RaftActor {
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(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
- // 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()));
+ 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
@@ -393,8 +407,15 @@ 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.
@@ -404,41 +425,40 @@ public class Shard extends RaftActor {
LOG.error(ex.getMessage());
sender.tell(new akka.actor.Status.Failure(ex), getSelf());
}
-
- return;
+ } else {
+ finishCommit(sender, transactionID, 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().commit().get();
-
- sender.tell(CommitTransactionReply.INSTANCE.toSerializable(), getSelf());
+ private void handleCanCommitTransaction(final CanCommitTransaction canCommit) {
+ LOG.debug("{}: Can committing transaction {}", persistenceId(), canCommit.getTransactionID());
- shardMBean.incrementCommittedTransactionCount();
- shardMBean.setLastCommittedTransactionTime(System.currentTimeMillis());
+ 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());
+ }
+ }
+ }
+ 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());
-
- LOG.error("{}, An exception occurred while committing transaction {}", persistenceId(),
- transactionID, e);
- shardMBean.incrementFailedTransactionsCount();
- } finally {
- commitCoordinator.currentTransactionComplete(transactionID, true);
}
}
- private void handleCanCommitTransaction(final CanCommitTransaction canCommit) {
- LOG.debug("{}: Can committing transaction {}", persistenceId(), canCommit.getTransactionID());
- commitCoordinator.handleCanCommit(canCommit, getSender(), self());
- }
-
private void handleBatchedModifications(BatchedModifications batched) {
- // This message is sent to prepare the modificationsa transaction directly on the Shard as an
+ // 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
@@ -450,63 +470,93 @@ 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()) {
+ 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 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) {
+ LOG.debug("{}: handleReadyLocalTransaction for {}", persistenceId(), message.getTransactionID());
+
+ boolean isLeaderActive = isLeaderActive();
+ if (isLeader() && isLeaderActive) {
try {
- BatchedModificationsReply reply = commitCoordinator.handleTransactionModifications(batched);
- sender().tell(reply, self());
+ commitCoordinator.handleReadyLocalTransaction(message, getSender(), this, store.getSchemaContext());
} catch (Exception e) {
- LOG.error("{}: Error handling BatchedModifications for Tx {}", persistenceId(),
- batched.getTransactionID(), 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(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(message, getSender(),
+ "Could not commit transaction " + message.getTransactionID());
} 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());
+ LOG.debug("{}: Forwarding ReadyLocalTransaction to leader {}", persistenceId(), leader);
+ message.setRemoteVersion(getCurrentBehavior().getLeaderPayloadVersion());
+ leader.forward(message, getContext());
}
}
}
- 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.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());
+ private void handleForwardedReadyTransaction(ForwardedReadyTransaction forwardedReady) {
+ LOG.debug("{}: handleForwardedReadyTransaction for {}", persistenceId(), forwardedReady.getTransactionID());
+ 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);
- getSender().tell(ready.isReturnSerialized() ? READY_TRANSACTION_REPLY.toSerializable() :
- READY_TRANSACTION_REPLY, getSelf());
+ ReadyLocalTransaction readyLocal = new ReadyLocalTransaction(forwardedReady.getTransactionID(),
+ forwardedReady.getTransaction().getSnapshot(), forwardedReady.isDoImmediateCommit());
+ readyLocal.setRemoteVersion(getCurrentBehavior().getLeaderPayloadVersion());
+ leader.forward(readyLocal, getContext());
+ }
}
}
@@ -514,39 +564,8 @@ public class Shard extends RaftActor {
doAbortTransaction(abort.getTransactionID(), getSender());
}
- void doAbortTransaction(final String transactionID, final ActorRef sender) {
- final CohortEntry cohortEntry = commitCoordinator.getCohortEntryIfCurrent(transactionID);
- if(cohortEntry != null) {
- LOG.debug("{}: Aborting transaction {}", persistenceId(), transactionID);
-
- // We don't remove the cached cohort entry here (ie pass false) in case the Tx was
- // aborted during replication in which case we may still commit locally if replication
- // succeeds.
- commitCoordinator.currentTransactionComplete(transactionID, false);
-
- final ListenableFuture 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);
- }
- }
- });
- }
+ void doAbortTransaction(final TransactionIdentifier transactionID, final ActorRef sender) {
+ commitCoordinator.handleAbort(transactionID, sender, this);
}
private void handleCreateTransaction(final Object message) {
@@ -555,179 +574,81 @@ 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());
}
}
private void closeTransactionChain(final CloseTransactionChain closeTransactionChain) {
- transactionFactory.closeTransactionChain(closeTransactionChain.getTransactionChainId());
- }
-
- private ActorRef createTypedTransactionActor(int transactionType,
- ShardTransactionIdentifier transactionId, String transactionChainId,
- short clientVersion ) {
-
- DOMStoreTransaction transaction = transactionFactory.newTransaction(
- TransactionProxy.TransactionType.fromInt(transactionType), transactionId.toString(),
- transactionChainId);
-
- return createShardTransaction(transaction, transactionId, clientVersion);
- }
-
- 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());
-
+ 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());
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, TransactionIdentifier transactionId) {
+ LOG.debug("{}: Creating transaction : {} ", persistenceId(), transactionId);
+ return transactionActorFactory.newShardTransaction(TransactionType.fromInt(transactionType),
+ transactionId);
}
- private void commitWithNewTransaction(final Modification modification) {
- DOMStoreWriteTransaction tx = store.newWriteOnlyTransaction();
- modification.apply(tx);
+ private void commitWithNewTransaction(final BatchedModifications modification) {
+ ReadWriteShardDataTreeTransaction tx = store.newReadWriteTransaction(modification.getTransactionID());
+ 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);
- }
-
- 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), 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());
+ store.updateSchemaContext(schemaContext);
}
- private boolean isMetricsCaptureEnabled(){
+ private boolean isMetricsCaptureEnabled() {
CommonConfig config = new CommonConfig(getContext().system().settings().config());
return config.isMetricCaptureEnabled();
}
@Override
- protected
- void startLogRecoveryBatch(final int maxBatchSize) {
- recoveryCoordinator.startLogRecoveryBatch(maxBatchSize);
- }
-
- @Override
- protected void appendRecoveredLogEntry(final Payload data) {
- recoveryCoordinator.appendRecoveredLogPayload(data);
- }
-
- @Override
- protected void applyRecoverySnapshot(final byte[] snapshotBytes) {
- recoveryCoordinator.applyRecoveredSnapshot(snapshotBytes);
+ @VisibleForTesting
+ public RaftActorSnapshotCohort getRaftActorSnapshotCohort() {
+ return snapshotCohort;
}
@Override
- protected void applyCurrentLogRecoveryBatch() {
- recoveryCoordinator.applyCurrentLogRecoveryBatch();
+ @Nonnull
+ protected RaftActorRecoveryCohort getRaftActorRecoveryCohort() {
+ return new ShardRecoveryCoordinator(store, store.getSchemaContext(),
+ restoreFromSnapshot != null ? restoreFromSnapshot.getSnapshot() : null, persistenceId(), LOG);
}
@Override
protected void onRecoveryComplete() {
- recoveryCoordinator = null;
+ restoreFromSnapshot = null;
//notify shard manager
getContext().parent().tell(new ActorInitialized(), getSelf());
@@ -744,31 +665,26 @@ 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());
}
}
- 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 = {}",
@@ -776,7 +692,11 @@ public class Shard extends RaftActor {
} 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));
+
+ // The only implementation we know of is BatchedModifications, which also carries a transaction
+ // identifier -- which we really need that.
+ Preconditions.checkArgument(modification instanceof BatchedModifications);
+ commitWithNewTransaction((BatchedModifications)modification);
} else {
// This must be the OK to commit after replication consensus.
finishCommit(clientActor, identifier);
@@ -784,74 +704,67 @@ public class Shard extends RaftActor {
}
@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();
+ protected void onStateChanged() {
+ boolean isLeader = isLeader();
+ boolean hasLeader = hasLeader();
+ changeSupport.onLeadershipChange(isLeader, hasLeader);
+ treeChangeSupport.onLeadershipChange(isLeader, hasLeader);
- NormalizedNode, ?> node = SerializationUtils.deserializeNormalizedNode(snapshotBytes);
+ // 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());
+ }
- // delete everything first
- transaction.delete(DATASTORE_ROOT);
+ store.closeAllTransactionChains();
+ }
- // 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());
+ if(hasLeader && !isIsolatedLeader()) {
+ messageRetrySupport.retryMessages();
}
}
@Override
- protected void onStateChanged() {
- boolean isLeader = isLeader();
- for (ActorSelection dataChangeListener : dataChangeListeners) {
- dataChangeListener.tell(new EnableNotification(isLeader), getSelf());
- }
+ protected void onLeaderChanged(String oldLeader, String newLeader) {
+ shardMBean.incrementLeadershipChangeCount();
- if(isLeader) {
- for(DelayedListenerRegistration reg: delayedListenerRegistrations) {
- if(!reg.isClosed()) {
- reg.setDelegate(doChangeListenerRegistration(reg.getRegisterChangeListener()));
- }
- }
+ 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