[
https://issues.apache.org/jira/browse/ARTEMIS-1011?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15893884#comment-15893884
]
Miroslav Novak commented on ARTEMIS-1011:
-----------------------------------------
> I'm not clear why you need the AtomicLong messageCountSnapshot. Couldn't you
> just invoke getMessageCount()?
Only got inspired by {{messagesAddedSnapshot.getAndSet(locaMessageAdded))}}. It
just makes code nicer because of the getAndSet() method. Otherwise I don't see
reason to use AtomicLong for messageCountSnapshot and messagesAddedSnapshot.
> Invoking getMessageCount() will lock the queue and therefore negatively
> impact performance for high-throughput use-cases. We may want to add some
> kind of optimization to getRate() to avoid that call if at all possible or
> perhaps avoiding the call to getRate() unless absolutely necessary.
Good point. Just a quick idea. Producer msg/s rate could be calculated just
from messageAdded as it was until now. If it does not meet the condition in if
statement in SlowConsumerReaperRunnable.run() line 3135 {{} else if (queueRate
< (threshold * consumersSet.size())) {}} then we would calculate producer msg/s
again from messageAdded and messageCount as suggested in description. This
should avoid calling getMessageCount() when there is a high load. I assume here
that high load means that there are lots of added messages.
> This changes the semantics of the method such that it no longer returns the
> rate of message production on the queue between invocations so it would
> probably be good to rename the method to something more accurate.
Yes :-)
Test scenario could be like:
* Start server with slow consumer policy set to KILL, stresshold 10 msg/s,
check period 5 seconds
* Send 100 messages to queue, wait 5 seconds
* Start slow consumer - consumes messages in rate 1 msg/s
Pass Criteria: Consumer was disconnected. This test will fail now.
> Slow consumer detection - producer msg/s rate for queue should take into
> account messages which are already in queue
> --------------------------------------------------------------------------------------------------------------------
>
> Key: ARTEMIS-1011
> URL: https://issues.apache.org/jira/browse/ARTEMIS-1011
> Project: ActiveMQ Artemis
> Issue Type: Bug
> Components: Broker
> Affects Versions: 1.5.3
> Reporter: Miroslav Novak
> Fix For: 2.0.0, 1.5.4
>
>
> There is still a problem how producer msg/s rate is calculated in
> {{QueueImpl.getRate()}} for slow consumer detection. It calculates only
> messages added during the last slow consumer check period. As this is used to
> figure out, in which msg/s rate the queue could serve the consumer then it
> should also take into account messages which are already in queue at the
> start of queueRateCheckTime period.
> Current implementation is problem for cases when messages are sent to queue
> in bursts, for example producer sends 1000s messages in a few seconds and
> then stops and will do that again in 1 hour. QueueImpl.getRate() method
> returns 0 msg/s for slow consumer check period set to for example 5 min and
> slow consumer detection will be skipped.
> I tried to fix it by following change to QueueImpl.getRate() method and seems
> to be ok, wdyt?
> {code}
> private final AtomicLong messageCountSnapshot = new AtomicLong(0);
> public float getRate() {
> long locaMessageAdded = getMessagesAdded();
> float timeSlice = ((System.currentTimeMillis() -
> queueRateCheckTime.getAndSet(System.currentTimeMillis())) / 1000.0f);
> if (timeSlice == 0) {
> messagesAddedSnapshot.getAndSet(locaMessageAdded);
> return 0.0f;
> }
> return BigDecimal.valueOf(((locaMessageAdded -
> messagesAddedSnapshot.getAndSet(locaMessageAdded)) +
> messageCountSnapshot.getAndSet(getMessageCount())) / timeSlice).setScale(2,
> BigDecimal.ROUND_UP).floatValue();
> }
> {code}
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
