public void handleCompleteTransactionMessage(CompleteTransactionMessage message) {
if (m_isLeader) {
CompleteTransactionMessage replmsg = new CompleteTransactionMessage(message);
// Set the spHandle so that on repair the new master will set the max seen spHandle
// correctly
advanceTxnEgo();
replmsg.setSpHandle(getCurrentTxnId());
if (m_sendToHSIds.length > 0) {
m_mailbox.send(m_sendToHSIds, replmsg);
}
} else {
setMaxSeenTxnId(message.getSpHandle());
}
TransactionState txn = m_outstandingTxns.get(message.getTxnId());
// We can currently receive CompleteTransactionMessages for multipart procedures
// which only use the buddy site (replicated table read). Ignore them for
// now, fix that later.
if (txn != null) {
Iv2Trace.logCompleteTransactionMessage(message, m_mailbox.getHSId());
final CompleteTransactionTask task =
new CompleteTransactionTask(txn, m_pendingTasks, message, m_drGateway);
queueOrOfferMPTask(task);
// If this is a restart, then we need to leave the transaction state around
if (!message.isRestart()) {
m_outstandingTxns.remove(message.getTxnId());
}
}
}
// 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);
}
// BorrowTaskMessages encapsulate a FragmentTaskMessage along with
// input dependency tables. The MPI issues borrows to a local site
// to perform replicated reads or aggregation fragment work.
private void handleBorrowTaskMessage(BorrowTaskMessage message) {
// borrows do not advance the sp handle. The handle would
// move backwards anyway once the next message is received
// from the SP leader.
long newSpHandle = getCurrentTxnId();
Iv2Trace.logFragmentTaskMessage(
message.getFragmentTaskMessage(), m_mailbox.getHSId(), newSpHandle, true);
TransactionState txn = m_outstandingTxns.get(message.getTxnId());
if (txn == null) {
// If the borrow is the first fragment for a transaction, run it as
// a single partition fragment; Must not engage/pause this
// site on a MP transaction before the SP instructs to do so.
// Do not track the borrow task as outstanding - it completes
// immediately and is not a valid transaction state for
// full MP participation (it claims everything can run as SP).
txn = new BorrowTransactionState(newSpHandle, message);
}
if (message.getFragmentTaskMessage().isSysProcTask()) {
final SysprocFragmentTask task =
new SysprocFragmentTask(
m_mailbox,
(ParticipantTransactionState) txn,
m_pendingTasks,
message.getFragmentTaskMessage(),
message.getInputDepMap());
m_pendingTasks.offer(task);
} else {
final FragmentTask task =
new FragmentTask(
m_mailbox,
(ParticipantTransactionState) txn,
m_pendingTasks,
message.getFragmentTaskMessage(),
message.getInputDepMap());
m_pendingTasks.offer(task);
}
}
// SpScheduler expects to see InitiateTaskMessages corresponding to single-partition
// procedures only.
public void handleIv2InitiateTaskMessage(Iv2InitiateTaskMessage message) {
if (!message.isSinglePartition()) {
throw new RuntimeException(
"SpScheduler.handleIv2InitiateTaskMessage "
+ "should never receive multi-partition initiations.");
}
final String procedureName = message.getStoredProcedureName();
long newSpHandle;
long uniqueId = Long.MIN_VALUE;
Iv2InitiateTaskMessage msg = message;
if (m_isLeader || message.isReadOnly()) {
/*
* A short circuit read is a read where the client interface is local to
* this node. The CI will let a replica perform a read in this case and
* it does looser tracking of client handles since it can't be
* partitioned from the local replica.
*/
if (!m_isLeader
&& CoreUtils.getHostIdFromHSId(msg.getInitiatorHSId())
!= CoreUtils.getHostIdFromHSId(m_mailbox.getHSId())) {
VoltDB.crashLocalVoltDB("Only allowed to do short circuit reads locally", true, null);
}
/*
* If this is for CL replay or DR, update the unique ID generator
*/
if (message.isForReplay()) {
uniqueId = message.getUniqueId();
try {
m_uniqueIdGenerator.updateMostRecentlyGeneratedUniqueId(uniqueId);
} catch (Exception e) {
hostLog.fatal(e.getMessage());
hostLog.fatal("Invocation: " + message);
VoltDB.crashLocalVoltDB(e.getMessage(), true, e);
}
} else if (message.isForDR()) {
uniqueId = message.getStoredProcedureInvocation().getOriginalUniqueId();
// @LoadSinglepartitionTable does not have a valid uid
if (UniqueIdGenerator.getPartitionIdFromUniqueId(uniqueId) == m_partitionId) {
m_uniqueIdGenerator.updateMostRecentlyGeneratedUniqueId(uniqueId);
}
}
/*
* If this is CL replay use the txnid from the CL and also
* update the txnid to match the one from the CL
*/
if (message.isForReplay()) {
newSpHandle = message.getTxnId();
setMaxSeenTxnId(newSpHandle);
} else if (m_isLeader) {
TxnEgo ego = advanceTxnEgo();
newSpHandle = ego.getTxnId();
uniqueId = m_uniqueIdGenerator.getNextUniqueId();
} else {
/*
* The short circuit read case. Since we are not a master
* we can't create new transaction IDs, so reuse the last seen
* txnid. For a timestamp, might as well give a reasonable one
* for a read heavy workload so time isn't bursty.
*/
uniqueId =
UniqueIdGenerator.makeIdFromComponents(
Math.max(System.currentTimeMillis(), m_uniqueIdGenerator.lastUsedTime),
0,
m_uniqueIdGenerator.partitionId);
// Don't think it wise to make a new one for a short circuit read
newSpHandle = getCurrentTxnId();
}
// Need to set the SP handle on the received message
// Need to copy this or the other local sites handling
// the same initiate task message will overwrite each
// other's memory -- the message isn't copied on delivery
// to other local mailboxes.
msg =
new Iv2InitiateTaskMessage(
message.getInitiatorHSId(),
message.getCoordinatorHSId(),
m_repairLogTruncationHandle,
message.getTxnId(),
message.getUniqueId(),
message.isReadOnly(),
message.isSinglePartition(),
message.getStoredProcedureInvocation(),
message.getClientInterfaceHandle(),
message.getConnectionId(),
message.isForReplay());
msg.setSpHandle(newSpHandle);
// Also, if this is a vanilla single-part procedure, make the TXNID
// be the SpHandle (for now)
// Only system procedures are every-site, so we'll check through the SystemProcedureCatalog
if (SystemProcedureCatalog.listing.get(procedureName) == null
|| !SystemProcedureCatalog.listing.get(procedureName).getEverysite()) {
msg.setTxnId(newSpHandle);
msg.setUniqueId(uniqueId);
}
// Don't replicate reads, this really assumes that DML validation
// is going to be integrated soonish
if (m_isLeader && !msg.isReadOnly() && m_sendToHSIds.length > 0) {
Iv2InitiateTaskMessage replmsg =
new Iv2InitiateTaskMessage(
m_mailbox.getHSId(),
m_mailbox.getHSId(),
m_repairLogTruncationHandle,
msg.getTxnId(),
msg.getUniqueId(),
msg.isReadOnly(),
msg.isSinglePartition(),
msg.getStoredProcedureInvocation(),
msg.getClientInterfaceHandle(),
msg.getConnectionId(),
msg.isForReplay());
// Update the handle in the copy since the constructor doesn't set it
replmsg.setSpHandle(newSpHandle);
m_mailbox.send(m_sendToHSIds, replmsg);
DuplicateCounter counter =
new DuplicateCounter(
msg.getInitiatorHSId(),
msg.getTxnId(),
m_replicaHSIds,
msg.getStoredProcedureName());
m_duplicateCounters.put(new DuplicateCounterKey(msg.getTxnId(), newSpHandle), counter);
}
} else {
setMaxSeenTxnId(msg.getSpHandle());
newSpHandle = msg.getSpHandle();
uniqueId = msg.getUniqueId();
}
Iv2Trace.logIv2InitiateTaskMessage(message, m_mailbox.getHSId(), msg.getTxnId(), newSpHandle);
doLocalInitiateOffer(msg);
return;
}
