// SpSchedulers will see FragmentTaskMessage for:
// - The scatter fragment(s) of a multi-part transaction (normal or sysproc)
// - Borrow tasks to do the local fragment work if this partition is the
// buddy of the MPI. Borrow tasks may include input dependency tables for
// aggregation fragments, or not, if it's a replicated table read.
// For multi-batch MP transactions, we'll need to look up the transaction state
// that gets created when the first batch arrives.
// During command log replay a new SP handle is going to be generated, but it really
// doesn't matter, it isn't going to be used for anything.
void handleFragmentTaskMessage(FragmentTaskMessage message) {
FragmentTaskMessage msg = message;
long newSpHandle;
if (m_isLeader) {
// Quick hack to make progress...we need to copy the FragmentTaskMessage
// before we start mucking with its state (SPHANDLE). We need to revisit
// all the messaging mess at some point.
msg =
new FragmentTaskMessage(
message.getInitiatorHSId(), message.getCoordinatorHSId(), message);
// Not going to use the timestamp from the new Ego because the multi-part timestamp is what
// should be used
TxnEgo ego = advanceTxnEgo();
newSpHandle = ego.getTxnId();
msg.setSpHandle(newSpHandle);
if (msg.getInitiateTask() != null) {
msg.getInitiateTask().setSpHandle(newSpHandle); // set the handle
msg.setInitiateTask(
msg.getInitiateTask()); // Trigger reserialization so the new handle is used
}
/*
* If there a replicas to send it to, forward it!
* Unless... it's read only AND not a sysproc. Read only sysprocs may expect to be sent
* everywhere.
* In that case don't propagate it to avoid a determinism check and extra messaging overhead
*/
if (m_sendToHSIds.length > 0 && (!msg.isReadOnly() || msg.isSysProcTask())) {
FragmentTaskMessage replmsg =
new FragmentTaskMessage(m_mailbox.getHSId(), m_mailbox.getHSId(), msg);
m_mailbox.send(m_sendToHSIds, replmsg);
DuplicateCounter counter;
/*
* Non-determinism should be impossible to happen with MP fragments.
* if you see "MP_DETERMINISM_ERROR" as procedure name in the crash logs
* something has horribly gone wrong.
*/
if (message.getFragmentTaskType() != FragmentTaskMessage.SYS_PROC_PER_SITE) {
counter =
new DuplicateCounter(
msg.getCoordinatorHSId(), msg.getTxnId(), m_replicaHSIds, "MP_DETERMINISM_ERROR");
} else {
counter =
new SysProcDuplicateCounter(
msg.getCoordinatorHSId(), msg.getTxnId(), m_replicaHSIds, "MP_DETERMINISM_ERROR");
}
m_duplicateCounters.put(new DuplicateCounterKey(msg.getTxnId(), newSpHandle), counter);
}
} else {
newSpHandle = msg.getSpHandle();
setMaxSeenTxnId(newSpHandle);
}
Iv2Trace.logFragmentTaskMessage(message, m_mailbox.getHSId(), newSpHandle, false);
doLocalFragmentOffer(msg);
}
/**
* Do the work necessary to turn the FragmentTaskMessage into a TransactionTask which can be
* queued to the TransactionTaskQueue. This is reused by both the normal message handling path and
* the repair path, and assumes that the caller has dealt with or ensured that the necessary ID,
* SpHandles, and replication issues are resolved.
*/
private void doLocalFragmentOffer(FragmentTaskMessage msg) {
TransactionState txn = m_outstandingTxns.get(msg.getTxnId());
boolean logThis = false;
// bit of a hack...we will probably not want to create and
// offer FragmentTasks for txn ids that don't match if we have
// something in progress already
if (txn == null) {
txn = new ParticipantTransactionState(msg.getSpHandle(), msg);
m_outstandingTxns.put(msg.getTxnId(), txn);
// Only want to send things to the command log if it satisfies this predicate
// AND we've never seen anything for this transaction before. We can't
// actually log until we create a TransactionTask, though, so just keep track
// of whether it needs to be done.
logThis = (msg.getInitiateTask() != null && !msg.getInitiateTask().isReadOnly());
}
// Check to see if this is the final task for this txn, and if so, if we can close it out early
// Right now, this just means read-only.
// NOTE: this overlaps slightly with CompleteTransactionMessage handling completion. It's so
// tiny
// that for now, meh, but if this scope grows then it should get refactored out
if (msg.isFinalTask() && txn.isReadOnly()) {
m_outstandingTxns.remove(msg.getTxnId());
}
TransactionTask task;
if (msg.isSysProcTask()) {
task =
new SysprocFragmentTask(
m_mailbox, (ParticipantTransactionState) txn, m_pendingTasks, msg, null);
} else {
task =
new FragmentTask(m_mailbox, (ParticipantTransactionState) txn, m_pendingTasks, msg, null);
}
if (logThis) {
if (!m_cl.log(msg.getInitiateTask(), msg.getSpHandle(), m_durabilityListener, task)) {
m_pendingTasks.offer(task);
} else {
/* Getting here means that the task is the first fragment of an MP txn and
* synchronous command logging is on, so create a backlog for future tasks of
* this MP arrived before it's marked durable.
*
* This is important for synchronous command logging and MP txn restart. Without
* this, a restarted MP txn may not be gated by logging of the first fragment.
*/
assert !m_mpsPendingDurability.containsKey(task.getTxnId());
m_mpsPendingDurability.put(task.getTxnId(), new ArrayDeque<TransactionTask>());
}
} else {
queueOrOfferMPTask(task);
}
}
private void logToDR(PartitionDRGateway drGateway) {
// Log invocation to DR
if (drGateway != null
&& !m_txnState.isForReplay()
&& !m_txnState.isReadOnly()
&& !m_completeMsg.isRollback()) {
FragmentTaskMessage fragment = (FragmentTaskMessage) m_txnState.getNotice();
Iv2InitiateTaskMessage initiateTask = fragment.getInitiateTask();
assert (initiateTask != null);
if (initiateTask == null) {
hostLog.error(
"Unable to log MP transaction to DR because of missing InitiateTaskMessage, "
+ "fragment: "
+ fragment.toString());
}
StoredProcedureInvocation invocation =
initiateTask.getStoredProcedureInvocation().getShallowCopy();
drGateway.onSuccessfulMPCall(
m_txnState.m_spHandle,
m_txnState.txnId,
m_txnState.uniqueId,
m_completeMsg.getHash(),
invocation,
m_txnState.getResults());
}
}
