@Override
protected void onStateChanged() {
boolean isLeader = 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) {
if (LOG.isDebugEnabled()) {
LOG.debug(
"{}: onStateChanged: Closing all transaction chains because shard {} is no longer the leader",
persistenceId(),
getId());
}
store.closeAllTransactionChains();
commitCoordinator.abortPendingTransactions(
"The transacton was aborted due to inflight leadership change.", this);
}
if (hasLeader && !isIsolatedLeader()) {
messageRetrySupport.retryMessages();
}
}
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 {
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());
}
}
}
private void handleReadyLocalTransaction(final ReadyLocalTransaction message) {
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 ReadyLocalTransaction for Tx {}",
persistenceId(),
message.getTransactionID(),
e);
getSender().tell(new akka.actor.Status.Failure(e), getSelf());
}
} else {
ActorSelection leader = getLeader();
if (!isLeaderActive || leader == null) {
messageRetrySupport.addMessageToRetry(
message, getSender(), "Could not commit transaction " + message.getTransactionID());
} else {
LOG.debug("{}: Forwarding ReadyLocalTransaction to leader {}", persistenceId(), leader);
message.setRemoteVersion(getCurrentBehavior().getLeaderPayloadVersion());
leader.forward(message, getContext());
}
}
}
private void handleBatchedModifications(BatchedModifications batched) {
// This message is sent to prepare the modifications transaction directly on the Shard as an
// optimization to avoid the extra overhead of a separate ShardTransaction actor. On the last
// BatchedModifications message, the caller sets the ready flag in the message indicating
// modifications are complete. The reply contains the cohort actor path (this actor) for the
// caller
// to initiate the 3-phase commit. This also avoids the overhead of sending an additional
// ReadyTransaction message.
// If we're not the leader then forward to the leader. This is a safety measure - we shouldn't
// normally get here if we're not the leader as the front-end (TransactionProxy) should
// determine
// the primary/leader shard. However with timing and caching on the front-end, there's a small
// window where it could have a stale leader during leadership transitions.
//
boolean isLeaderActive = isLeaderActive();
if (isLeader() && isLeaderActive) {
handleBatchedModificationsLocal(batched, getSender());
} else {
ActorSelection leader = getLeader();
if (!isLeaderActive || leader == null) {
messageRetrySupport.addMessageToRetry(
batched, getSender(), "Could not commit transaction " + batched.getTransactionID());
} else {
// 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());
}
}
}
@Override
protected void onLeaderChanged(String oldLeader, String newLeader) {
shardMBean.incrementLeadershipChangeCount();
if (hasLeader() && !isIsolatedLeader()) {
messageRetrySupport.retryMessages();
}
}
@Override
public void postStop() {
LOG.info("Stopping Shard {}", persistenceId());
super.postStop();
messageRetrySupport.close();
if (txCommitTimeoutCheckSchedule != null) {
txCommitTimeoutCheckSchedule.cancel();
}
commitCoordinator.abortPendingTransactions("Transaction aborted due to shutdown.", this);
shardMBean.unregisterMBean();
}
@Override
public void onReceiveCommand(final Object message) throws Exception {
MessageTracker.Context context = appendEntriesReplyTracker.received(message);
if (context.error().isPresent()) {
LOG.trace(
"{} : AppendEntriesReply failed to arrive at the expected interval {}",
persistenceId(),
context.error());
}
try {
if (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 (message instanceof ReadyLocalTransaction) {
handleReadyLocalTransaction((ReadyLocalTransaction) 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(), 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)) {
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) {
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 {
super.onReceiveCommand(message);
}
} finally {
context.done();
}
}
