[
https://issues.apache.org/jira/browse/POOL-418?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17924818#comment-17924818
]
SunShuai commented on POOL-418:
-------------------------------
Hi [~ggregory]
I think it should be changed like this. Today I was testing version 2.12.1,
and I found this issue quite obvious. The following image can prove this point.
I set the maxWaitMillis to 120ms, but the total time was 240ms, with both
create and pollFirst taking 120ms each.
!DzZUFH2cNMwTAAAAAElFTkSuQmCC|width=688,height=179!
Here's a suggestion: Before executing pollFirst, check the value of
remainingWaitDuration. If it's negative, there's no need to execute pollFirst,
which can save one lock operation inside pollFirst.
Something like:
{code:java}
p = negativeDuration || remainingWaitDuration.isNegative() ?
idleObjects.takeFirst() : idleObjects.pollFirst(maxWaitDuration);{code}
[~ggregory]
> The maximum wait time for GenericObjectPool.borrowObject(*) may exceed
> expectations due to a spurious thread wakeup
> -------------------------------------------------------------------------------------------------------------------
>
> Key: POOL-418
> URL: https://issues.apache.org/jira/browse/POOL-418
> Project: Commons Pool
> Issue Type: Bug
> Reporter: SunShuai
> Assignee: Gary D. Gregory
> Priority: Minor
> Fix For: 2.12.1
>
> Attachments: 74C00A2F-EE3A-4660-BE58-66CA37F9808A.png,
> EE4BD798-26B6-4AE1-B114-5AC7342B31A6.png
>
>
> I found an issue when using jedis to operate Redis database, Here is the
> issue link -> [https://github.com/redis/jedis/issues/4014]. we feel that it's
> a problem with the Commons pool.
>
> In situations where the connection is tight, the waiting time can be up to
> twice the maxWaitMillis consumption.
> code: org.apache.commons.pool2.impl.GenericObjectPool#borrowObject(long) ->
> org.apache.commons.pool2.impl.GenericObjectPool#create
> The thread that failed to create the connection will wake up the thread that
> is waiting to be created. Threads that do not compete for resources will wait
> again, and the waiting time will still be the maxWaitMillis.
> If the first waiting time is ( maxWaitMillis - 1ms), plus the second full
> waiting time, then the full waiting time will be twice the expected time.
>
> The waiting time we hope for this time is the total waiting time minus the
> time already waited.
>
> such as: Duration remainingWait =
> localMaxWaitDuration.minus(Duration.between(localStartInstant,
> Instant.now()));
> if (remainingWait.isNegative()) {
> create = Boolean.FALSE;
> } else {
> wait(makeObjectCountLock, remainingWait);
> }
>
> We are not sure if this issue has been addressed and look forward to your
> reply.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
