/* tail an oplog. ok to return, will be re-called. */
void SyncTail::oplogApplication() {
OpQueueBatcher batcher(this);
OperationContextImpl txn;
auto replCoord = ReplicationCoordinator::get(&txn);
ApplyBatchFinalizer finalizer(replCoord);
OpTime originalEndOpTime(getMinValid(&txn).end);
while (!inShutdown()) {
OpQueue ops;
do {
if (BackgroundSync::get()->getInitialSyncRequestedFlag()) {
// got a resync command
return;
}
tryToGoLiveAsASecondary(&txn, replCoord);
// Blocks up to a second waiting for a batch to be ready to apply. If one doesn't become
// ready in time, we'll loop again so we can do the above checks periodically.
ops = batcher.getNextBatch(Seconds(1));
} while (!inShutdown() && ops.empty());
if (inShutdown())
return;
invariant(!ops.empty());
const BSONObj lastOp = ops.back().raw;
if (lastOp.isEmpty()) {
// This means that the network thread has coalesced and we have processed all of its
// data.
invariant(ops.getDeque().size() == 1);
if (replCoord->isWaitingForApplierToDrain()) {
replCoord->signalDrainComplete(&txn);
}
continue; // This wasn't a real op. Don't try to apply it.
}
handleSlaveDelay(lastOp);
// Set minValid to the last OpTime that needs to be applied, in this batch or from the
// (last) failed batch, whichever is larger.
// This will cause this node to go into RECOVERING state
// if we should crash and restart before updating finishing.
const OpTime start(getLastSetTimestamp(), OpTime::kUninitializedTerm);
// Take the max of the first endOptime (if we recovered) and the end of our batch.
const auto lastOpTime = fassertStatusOK(28773, OpTime::parseFromOplogEntry(lastOp));
// Setting end to the max of originalEndOpTime and lastOpTime (the end of the batch)
// ensures that we keep pushing out the point where we can become consistent
// and allow reads. If we recover and end up doing smaller batches we must pass the
// originalEndOpTime before we are good.
//
// For example:
// batch apply, 20-40, end = 40
// batch failure,
// restart
// batch apply, 20-25, end = max(25, 40) = 40
// batch apply, 25-45, end = 45
const OpTime end(std::max(originalEndOpTime, lastOpTime));
// This write will not journal/checkpoint.
setMinValid(&txn, {start, end});
OpTime finalOpTime = multiApply(&txn, ops);
setNewTimestamp(finalOpTime.getTimestamp());
setMinValid(&txn, end, DurableRequirement::None);
finalizer.record(finalOpTime);
}
}
/* tail an oplog. ok to return, will be re-called. */
void SyncTail::oplogApplication() {
while( 1 ) {
OpQueue ops;
verify( !Lock::isLocked() );
Timer batchTimer;
int lastTimeChecked = 0;
// always fetch a few ops first
// tryPopAndWaitForMore returns true when we need to end a batch early
while (!tryPopAndWaitForMore(&ops) &&
(ops.getSize() < replBatchLimitBytes)) {
if (theReplSet->isPrimary()) {
massert(16620, "there are ops to sync, but I'm primary", ops.empty());
return;
}
int now = batchTimer.seconds();
// apply replication batch limits
if (!ops.empty()) {
if (now > replBatchLimitSeconds)
break;
if (ops.getDeque().size() > replBatchLimitOperations)
break;
}
// occasionally check some things
if (ops.empty() || now > lastTimeChecked) {
lastTimeChecked = now;
// can we become secondary?
// we have to check this before calling mgr, as we must be a secondary to
// become primary
if (!theReplSet->isSecondary()) {
OpTime minvalid;
theReplSet->tryToGoLiveAsASecondary(minvalid);
}
// normally msgCheckNewState gets called periodically, but in a single node repl set
// there are no heartbeat threads, so we do it here to be sure. this is relevant if the
// singleton member has done a stepDown() and needs to come back up.
if (theReplSet->config().members.size() == 1 &&
theReplSet->myConfig().potentiallyHot()) {
Manager* mgr = theReplSet->mgr;
// When would mgr be null? During replsettest'ing.
if (mgr) mgr->send(boost::bind(&Manager::msgCheckNewState, theReplSet->mgr));
sleepsecs(1);
// There should never be ops to sync in a 1-member set, anyway
return;
}
}
const int slaveDelaySecs = theReplSet->myConfig().slaveDelay;
if (!ops.empty() && slaveDelaySecs > 0) {
const BSONObj& lastOp = ops.getDeque().back();
const unsigned int opTimestampSecs = lastOp["ts"]._opTime().getSecs();
// Stop the batch as the lastOp is too new to be applied. If we continue
// on, we can get ops that are way ahead of the delay and this will
// make this thread sleep longer when handleSlaveDelay is called
// and apply ops much sooner than we like.
if (opTimestampSecs > static_cast<unsigned int>(time(0) - slaveDelaySecs)) {
break;
}
}
}
// For pausing replication in tests
while (MONGO_FAIL_POINT(rsSyncApplyStop)) {
sleepmillis(0);
}
const BSONObj& lastOp = ops.getDeque().back();
setOplogVersion(lastOp);
handleSlaveDelay(lastOp);
// Set minValid to the last op to be applied in this next batch.
// This will cause this node to go into RECOVERING state
// if we should crash and restart before updating the oplog
theReplSet->setMinValid(lastOp);
multiApply(ops.getDeque(), multiSyncApply);
applyOpsToOplog(&ops.getDeque());
}
}