distributedlog-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Sijie Guo (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (DL-145) Fix the flaky testServiceTimeout
Date Sun, 18 Dec 2016 16:44:58 GMT

    [ https://issues.apache.org/jira/browse/DL-145?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15759085#comment-15759085
] 

Sijie Guo commented on DL-145:
------------------------------

The change "Be able to close the writer within a timeout period" is to timeout the closing
operation if closing a writer takes too long, so we can remove the stream from acquired map.

For the race condition that you described above, it the write requests are not error out immediately
when closing writer. so they are different problems.

> Fix the flaky testServiceTimeout
> --------------------------------
>
>                 Key: DL-145
>                 URL: https://issues.apache.org/jira/browse/DL-145
>             Project: DistributedLog
>          Issue Type: Test
>          Components: distributedlog-service
>    Affects Versions: 0.4.0
>            Reporter: Liang Xie
>            Assignee: Liang Xie
>
> The TestDistributedLogService#testServiceTimeout case is not stable, e.g. https://builds.apache.org/job/distributedlog-precommit-pullrequest/22/com.twitter$distributedlog-service/testReport/com.twitter.distributedlog.service/TestDistributedLogService/testServiceTimeout/
> It could be reproduced on my box occasionally, and the failures were stable if i tuned
the ServiceTimeoutMs from 200 to 150, and always passed if tuned to a larger value, e.g. 1000(btw,
my disk is SSD type)
> After digging into it, shows it related with starting a new log segment corner case.
> For a good case, once service time out occurs, steam status : ERROR -> CLOSING ->
CLOSED, calling Abortables.asyncAbort will trigger the cached logsegment be aborted, then
writeOp will be injected an exception, e.g. write cancel exception.
> For a bad case, since no log records be written before, so there'll be an async start
new log segment, once the timeout occurs, the segment starting still not be done, so no cache,
then asyncAbort has no change to abort that segment.
> I think change the test timeout value to a larger one should be fine for this special
test corner case.
> will attach a minor patch later.  Any suggestions are welcome.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Mime
View raw message