On ApplyState, there's a check if the elapsed time exceeds a 50ms
threshold and it logs a warning. However the start time is captured when
the message is created prior to queueing. So if there's many ApplyState
or other messages already queued, the elapsed time also includes the time spent
in the queue, ie as a side effect includes the cumulative processing time
of each prior message in the queue. When a follower starts up, there can
be hundreds to thousands of catchup ApplyState messages and, eventually,
the cumulative processing times can add up to more than 50 ms, in which
case every subsequent ApplyState message trips the threshold with
increasing elapsed times, even though none of them actually took 50 ms
to process. Seeing hundreds to thousands of warnings with misleading
elapsed times looks ominous and leads users to think something is wrong.
Therefore I changed it to capture the start time just prior to calling
applyState so it captures just the processing time for that message. I
also removed the startTime field from ApplyState. This class is
Serializable but it is only ever sent locally to self and is never
serialized so there's no backwards compatibility concerns.
Change-Id: I9493734b5307d6dd5d723e5fe416ba97915dfc63
Signed-off-by: Tom Pantelis <tpanteli@brocade.com>
(cherry picked from commit
57d7e4788a488d992b9868d44ebc392b06e317c5)
} else if (message instanceof ApplyState){
ApplyState applyState = (ApplyState) message;
- long elapsedTime = (System.nanoTime() - applyState.getStartTime());
- if(elapsedTime >= APPLY_STATE_DELAY_THRESHOLD_IN_NANOS){
- LOG.warn("ApplyState took more time than expected. Elapsed Time = {} ms ApplyState = {}",
- TimeUnit.NANOSECONDS.toMillis(elapsedTime), applyState);
- }
+ long startTime = System.nanoTime();
if(LOG.isDebugEnabled()) {
LOG.debug("{}: Applying state for log index {} data {}",
applyState(applyState.getClientActor(), applyState.getIdentifier(),
applyState.getReplicatedLogEntry().getData());
+ long elapsedTime = System.nanoTime() - startTime;
+ if(elapsedTime >= APPLY_STATE_DELAY_THRESHOLD_IN_NANOS){
+ LOG.debug("ApplyState took more time than expected. Elapsed Time = {} ms ApplyState = {}",
+ TimeUnit.NANOSECONDS.toMillis(elapsedTime), applyState);
+ }
+
if (!hasFollowers()) {
// for single node, the capture should happen after the apply state
// as we delete messages from the persistent journal which have made it to the snapshot
private final ActorRef clientActor;
private final String identifier;
private final ReplicatedLogEntry replicatedLogEntry;
- private final long startTime;
public ApplyState(ActorRef clientActor, String identifier,
ReplicatedLogEntry replicatedLogEntry) {
this.clientActor = clientActor;
this.identifier = identifier;
this.replicatedLogEntry = replicatedLogEntry;
- this.startTime = System.nanoTime();
}
public ActorRef getClientActor() {
return replicatedLogEntry;
}
- public long getStartTime() {
- return startTime;
- }
-
@Override
public String toString() {
return "ApplyState{" +
"identifier='" + identifier + '\'' +
", replicatedLogEntry.index =" + replicatedLogEntry.getIndex() +
- ", startTime=" + startTime +
'}';
}
}