/* applies oplog from "now" until endOpTime using the applier threads for initial sync*/
void SyncTail::_applyOplogUntil(OperationContext* txn, const OpTime& endOpTime) {
unsigned long long bytesApplied = 0;
unsigned long long entriesApplied = 0;
while (true) {
OpQueue ops;
OperationContextImpl ctx;
while (!tryPopAndWaitForMore(&ops, getGlobalReplicationCoordinator())) {
// nothing came back last time, so go again
if (ops.empty()) continue;
// Check if we reached the end
const BSONObj currentOp = ops.back();
const OpTime currentOpTime = currentOp["ts"]._opTime();
// When we reach the end return this batch
if (currentOpTime == endOpTime) {
break;
}
else if (currentOpTime > endOpTime) {
severe() << "Applied past expected end " << endOpTime << " to " << currentOpTime
<< " without seeing it. Rollback?" << rsLog;
fassertFailedNoTrace(18693);
}
// apply replication batch limits
if (ops.getSize() > replBatchLimitBytes)
break;
if (ops.getDeque().size() > replBatchLimitOperations)
break;
};
if (ops.empty()) {
severe() << "got no ops for batch...";
fassertFailedNoTrace(18692);
}
const BSONObj lastOp = ops.back().getOwned();
// Tally operation information
bytesApplied += ops.getSize();
entriesApplied += ops.getDeque().size();
multiApply(ops.getDeque());
OpTime lastOpTime = applyOpsToOplog(&ops.getDeque());
// if the last op applied was our end, return
if (lastOpTime == endOpTime) {
LOG(1) << "SyncTail applied " << entriesApplied
<< " entries (" << bytesApplied << " bytes)"
<< " and finished at opTime " << endOpTime.toStringPretty();
return;
}
} // end of while (true)
}
void run() {
Client::initThread("ReplBatcher");
OperationContextImpl txn;
auto replCoord = ReplicationCoordinator::get(&txn);
while (!_inShutdown.load()) {
Timer batchTimer;
OpQueue ops;
// tryPopAndWaitForMore returns true when we need to end a batch early
while (!_syncTail->tryPopAndWaitForMore(&txn, &ops) &&
(ops.getSize() < replBatchLimitBytes) && !_inShutdown.load()) {
int now = batchTimer.seconds();
// apply replication batch limits
if (!ops.empty()) {
if (now > replBatchLimitSeconds)
break;
if (ops.getDeque().size() > replBatchLimitOperations)
break;
}
const int slaveDelaySecs = durationCount<Seconds>(replCoord->getSlaveDelaySecs());
if (!ops.empty() && slaveDelaySecs > 0) {
const BSONObj lastOp = ops.back().raw;
const unsigned int opTimestampSecs = lastOp["ts"].timestamp().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;
}
}
if (MONGO_FAIL_POINT(rsSyncApplyStop)) {
break;
}
// keep fetching more ops as long as we haven't filled up a full batch yet
}
// For pausing replication in tests
while (MONGO_FAIL_POINT(rsSyncApplyStop) && !_inShutdown.load()) {
sleepmillis(0);
}
stdx::unique_lock<stdx::mutex> lk(_mutex);
while (!_ops.empty()) {
// Block until the previous batch has been taken.
if (_inShutdown.load())
return;
_cv.wait(lk);
}
_ops = std::move(ops);
_cv.notify_all();
}
}
/* tail an oplog. ok to return, will be re-called. */
void SyncTail::oplogApplication() {
ReplicationCoordinator* replCoord = getGlobalReplicationCoordinator();
ApplyBatchFinalizer finalizer(replCoord);
OperationContextImpl txn;
OpTime originalEndOpTime(getMinValid(&txn).end);
while (!inShutdown()) {
OpQueue ops;
Timer batchTimer;
int lastTimeChecked = 0;
do {
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
// (always checked in the first iteration of this do-while loop, because
// ops is empty)
if (ops.empty() || now > lastTimeChecked) {
BackgroundSync* bgsync = BackgroundSync::get();
if (bgsync->getInitialSyncRequestedFlag()) {
// got a resync command
return;
}
lastTimeChecked = now;
// can we become secondary?
// we have to check this before calling mgr, as we must be a secondary to
// become primary
tryToGoLiveAsASecondary(&txn, replCoord);
}
const int slaveDelaySecs = durationCount<Seconds>(replCoord->getSlaveDelaySecs());
if (!ops.empty() && slaveDelaySecs > 0) {
const BSONObj lastOp = ops.back();
const unsigned int opTimestampSecs = lastOp["ts"].timestamp().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;
}
}
if (MONGO_FAIL_POINT(rsSyncApplyStop)) {
break;
}
// keep fetching more ops as long as we haven't filled up a full batch yet
} while (!tryPopAndWaitForMore(&txn, &ops, replCoord) && // tryPopAndWaitForMore returns
// true when we need to end a
// batch early
(ops.getSize() < replBatchLimitBytes) &&
!inShutdown());
// For pausing replication in tests
while (MONGO_FAIL_POINT(rsSyncApplyStop)) {
sleepmillis(0);
if (inShutdown())
return;
}
if (ops.empty()) {
continue;
}
const BSONObj lastOp = ops.back();
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);
}
}
/* applies oplog from "now" until endOpTime using the applier threads for initial sync*/
void InitialSync::_applyOplogUntil(OperationContext* txn, const OpTime& endOpTime) {
unsigned long long bytesApplied = 0;
unsigned long long entriesApplied = 0;
while (true) {
OpQueue ops;
auto replCoord = repl::ReplicationCoordinator::get(txn);
while (!tryPopAndWaitForMore(txn, &ops)) {
// nothing came back last time, so go again
if (ops.empty())
continue;
// Check if we reached the end
const BSONObj currentOp = ops.back().raw;
const OpTime currentOpTime =
fassertStatusOK(28772, OpTime::parseFromOplogEntry(currentOp));
// When we reach the end return this batch
if (currentOpTime == endOpTime) {
break;
} else if (currentOpTime > endOpTime) {
severe() << "Applied past expected end " << endOpTime << " to " << currentOpTime
<< " without seeing it. Rollback?";
fassertFailedNoTrace(18693);
}
// apply replication batch limits
if (ops.getSize() > replBatchLimitBytes)
break;
if (ops.getDeque().size() > replBatchLimitOperations)
break;
};
if (ops.empty()) {
severe() << "got no ops for batch...";
fassertFailedNoTrace(18692);
}
const BSONObj lastOp = ops.back().raw.getOwned();
// Tally operation information
bytesApplied += ops.getSize();
entriesApplied += ops.getDeque().size();
const OpTime lastOpTime = multiApply(txn, ops);
replCoord->setMyLastAppliedOpTime(lastOpTime);
setNewTimestamp(lastOpTime.getTimestamp());
if (inShutdown()) {
return;
}
// if the last op applied was our end, return
if (lastOpTime == endOpTime) {
LOG(1) << "SyncTail applied " << entriesApplied << " entries (" << bytesApplied
<< " bytes) and finished at opTime " << endOpTime;
return;
}
} // end of while (true)
}
/* 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);
auto minValidBoundaries = getMinValid(&txn);
OpTime originalEndOpTime(minValidBoundaries.end);
OpTime lastWriteOpTime{replCoord->getMyLastOptime()};
while (!inShutdown()) {
OpQueue ops;
do {
if (BackgroundSync::get()->getInitialSyncRequestedFlag()) {
// got a resync command
return;
}
tryToGoLiveAsASecondary(&txn, replCoord, minValidBoundaries, lastWriteOpTime);
// 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});
lastWriteOpTime = multiApply(&txn, ops);
setNewTimestamp(lastWriteOpTime.getTimestamp());
setMinValid(&txn, end, DurableRequirement::None);
minValidBoundaries.start = {};
minValidBoundaries.end = end;
finalizer.record(lastWriteOpTime);
}
}
