@Override
public void run(SiteProcedureConnection siteConnection) {
hostLog.debug("STARTING: " + this);
if (!m_txnState.isReadOnly()) {
// the truncation point token SHOULD be part of m_txn. However, the
// legacy interaces don't work this way and IV2 hasn't changed this
// ownership yet. But truncateUndoLog is written assuming the right
// eventual encapsulation.
siteConnection.truncateUndoLog(
m_completeMsg.isRollback(),
m_txnState.getBeginUndoToken(),
m_txnState.m_spHandle,
m_txnState.getUndoLog());
}
if (!m_completeMsg.isRestart()) {
doCommonSPICompleteActions();
// Log invocation to DR
logToDR(siteConnection.getDRGateway());
hostLog.debug("COMPLETE: " + this);
} else {
// If we're going to restart the transaction, then reset the begin undo token so the
// first FragmentTask will set it correctly. Otherwise, don't set the Done state or
// flush the queue; we want the TransactionTaskQueue to stay blocked on this TXN ID
// for the restarted fragments.
m_txnState.setBeginUndoToken(Site.kInvalidUndoToken);
hostLog.debug("RESTART: " + this);
}
final CompleteTransactionResponseMessage resp =
new CompleteTransactionResponseMessage(m_completeMsg);
resp.m_sourceHSId = m_initiator.getHSId();
m_initiator.deliver(resp);
}
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());
}
}
}
public static void logCompleteTransactionMessage(
CompleteTransactionMessage ctask, long localHSId) {
if (iv2log.isTraceEnabled()) {
String logmsg = "rxCompMsg %s from %s txnId %s %s %s";
iv2log.trace(
String.format(
logmsg,
CoreUtils.hsIdToString(localHSId),
CoreUtils.hsIdToString(ctask.m_sourceHSId),
txnIdToString(ctask.getTxnId()),
ctask.isRollback() ? "ROLLBACK" : "COMMIT",
ctask.isRestart() ? "RESTART" : ""));
}
}
@Override
public void runForRejoin(SiteProcedureConnection siteConnection, TaskLog taskLog)
throws IOException {
if (!m_txnState.isReadOnly() && !m_completeMsg.isRollback()) {
// ENG-5276: Need to set the last committed spHandle so that the rejoining site gets the
// accurate
// per-partition txnId set for the next snapshot. Normally, this is done through undo log
// truncation.
// Since the task is not run here, we need to set the last committed spHandle explicitly.
//
// How does this work?
// - Blocking rejoin with idle cluster: The spHandle is updated here with the spHandle of the
// stream
// snapshot that transfers the rejoin data. So the snapshot right after rejoin should have
// the spHandle
// passed here.
// - Live rejoin with idle cluster: Same as blocking rejoin.
// - Live rejoin with workload: Transactions will be logged and replayed afterward. The
// spHandle will be
// updated when they commit and truncate undo logs. So at the end of replay,
// the spHandle should have the latest value. If all replayed transactions rolled back,
// the spHandle is still guaranteed to be the spHandle of the stream snapshot that
// transfered the
// rejoin data, which is the correct value.
siteConnection.setSpHandleForSnapshotDigest(m_txnState.m_spHandle);
}
if (!m_completeMsg.isRestart()) {
// future: offer to siteConnection.IBS for replay.
doCommonSPICompleteActions();
}
if (!m_txnState.isReadOnly()) {
// We need to log the restarting message to the task log so we'll replay the whole
// stream faithfully
taskLog.logTask(m_completeMsg);
}
final CompleteTransactionResponseMessage resp =
new CompleteTransactionResponseMessage(m_completeMsg);
resp.setIsRecovering(true);
resp.m_sourceHSId = m_initiator.getHSId();
m_initiator.deliver(resp);
}
@Override
public void runFromTaskLog(SiteProcedureConnection siteConnection) {
if (!m_txnState.isReadOnly()) {
// the truncation point token SHOULD be part of m_txn. However, the
// legacy interaces don't work this way and IV2 hasn't changed this
// ownership yet. But truncateUndoLog is written assuming the right
// eventual encapsulation.
siteConnection.truncateUndoLog(
m_completeMsg.isRollback(),
m_txnState.getBeginUndoToken(),
m_txnState.m_spHandle,
m_txnState.getUndoLog());
}
if (!m_completeMsg.isRestart()) {
// this call does the right thing with a null TransactionTaskQueue
doCommonSPICompleteActions();
logToDR(siteConnection.getDRGateway());
} else {
m_txnState.setBeginUndoToken(Site.kInvalidUndoToken);
}
}
// Offer a new message to the repair log. This will truncate
// the repairLog if the message includes a truncation hint.
public void deliver(VoltMessage msg) {
if (!m_isLeader && msg instanceof Iv2InitiateTaskMessage) {
final Iv2InitiateTaskMessage m = (Iv2InitiateTaskMessage) msg;
// We can't repair read only SP transactions. Just don't log them to the repair log.
if (m.isReadOnly()) {
return;
}
m_lastSpHandle = m.getSpHandle();
truncate(m.getTruncationHandle(), IS_SP);
m_logSP.add(new Item(IS_SP, m, m.getSpHandle(), m.getTxnId()));
} else if (msg instanceof FragmentTaskMessage) {
final FragmentTaskMessage m = (FragmentTaskMessage) msg;
// We can't repair read only SP transactions. Just don't log them to the repair log.
if (m.isReadOnly()) {
return;
}
truncate(m.getTruncationHandle(), IS_MP);
// only log the first fragment of a procedure (and handle 1st case)
if (m.getTxnId() > m_lastMpHandle || m_lastMpHandle == Long.MAX_VALUE) {
m_logMP.add(new Item(IS_MP, m, m.getSpHandle(), m.getTxnId()));
m_lastMpHandle = m.getTxnId();
m_lastSpHandle = m.getSpHandle();
}
} else if (msg instanceof CompleteTransactionMessage) {
// a CompleteTransactionMessage which indicates restart is not the end of the
// transaction. We don't want to log it in the repair log.
CompleteTransactionMessage ctm = (CompleteTransactionMessage) msg;
// We can't repair read only SP transactions. Just don't log them to the repair log.
// Restart transaction do not need to be repaired here, don't log them as well.
if (ctm.isReadOnly() || ctm.isRestart()) {
return;
}
truncate(ctm.getTruncationHandle(), IS_MP);
m_logMP.add(new Item(IS_MP, ctm, ctm.getSpHandle(), ctm.getTxnId()));
// Restore will send a complete transaction message with a lower mp transaction id because
// the restore transaction precedes the loading of the right mp transaction id from the
// snapshot
// Hence Math.max
m_lastMpHandle = Math.max(m_lastMpHandle, ctm.getTxnId());
m_lastSpHandle = ctm.getSpHandle();
} else if (msg instanceof DumpMessage) {
String who = CoreUtils.hsIdToString(m_HSId);
tmLog.warn(
"Repair log dump for site: "
+ who
+ ", isLeader: "
+ m_isLeader
+ ", "
+ who
+ ": lastSpHandle: "
+ m_lastSpHandle
+ ", lastMpHandle: "
+ m_lastMpHandle);
for (Iv2RepairLogResponseMessage il : contents(0l, false)) {
tmLog.warn("[Repair log contents]" + who + ": msg: " + il);
}
} else if (msg instanceof RepairLogTruncationMessage) {
final RepairLogTruncationMessage truncateMsg = (RepairLogTruncationMessage) msg;
truncate(truncateMsg.getHandle(), IS_SP);
}
}
void replayFromTaskLog() throws IOException {
// not yet time to catch-up.
if (m_rejoinState != kStateReplayingRejoin) {
return;
}
// replay 10:1 in favor of replay
for (int i = 0; i < 10; ++i) {
if (m_rejoinTaskLog.isEmpty()) {
break;
}
TransactionInfoBaseMessage tibm = m_rejoinTaskLog.getNextMessage();
if (tibm == null) {
break;
}
// Apply the readonly / sysproc filter. With Iv2 read optimizations,
// reads should not reach here; the cost of post-filtering shouldn't
// be particularly high (vs pre-filtering).
if (filter(tibm)) {
continue;
}
if (tibm instanceof Iv2InitiateTaskMessage) {
Iv2InitiateTaskMessage m = (Iv2InitiateTaskMessage) tibm;
SpProcedureTask t =
new SpProcedureTask(m_initiatorMailbox, m.getStoredProcedureName(), null, m, null);
t.runFromTaskLog(this);
} else if (tibm instanceof FragmentTaskMessage) {
FragmentTaskMessage m = (FragmentTaskMessage) tibm;
if (global_replay_mpTxn == null) {
global_replay_mpTxn = new ParticipantTransactionState(m.getTxnId(), m);
} else if (global_replay_mpTxn.txnId != m.getTxnId()) {
VoltDB.crashLocalVoltDB(
"Started a MP transaction during replay before completing " + " open transaction.",
false,
null);
}
FragmentTask t = new FragmentTask(m_initiatorMailbox, m, global_replay_mpTxn);
t.runFromTaskLog(this);
} else if (tibm instanceof CompleteTransactionMessage) {
// Needs improvement: completes for sysprocs aren't filterable as sysprocs.
// Only complete transactions that are open...
if (global_replay_mpTxn != null) {
CompleteTransactionMessage m = (CompleteTransactionMessage) tibm;
CompleteTransactionTask t =
new CompleteTransactionTask(global_replay_mpTxn, null, m, null);
if (!m.isRestart()) {
global_replay_mpTxn = null;
}
t.runFromTaskLog(this);
}
} else {
VoltDB.crashLocalVoltDB(
"Can not replay message type " + tibm + " during live rejoin. Unexpected error.",
false,
null);
}
}
// exit replay being careful not to exit in the middle of a multi-partititon
// transaction. The SPScheduler doesn't have a valid transaction state for a
// partially replayed MP txn and in case of rollback the scheduler's undo token
// is wrong. Run MP txns fully kStateRejoining or fully kStateRunning.
if (m_rejoinTaskLog.isEmpty() && global_replay_mpTxn == null) {
setReplayRejoinComplete();
}
}
