[
https://issues.apache.org/jira/browse/MESOS-2484?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14359114#comment-14359114
]
Cody Maloney commented on MESOS-2484:
-------------------------------------
Since we continually spawn new pids for little one of things, the blacklist of
"used" pids would grow infinitely (Probably stay less than a couple MB for
thousands of entries, but still...)
It should be fairly straightforward long term to just make it so when we do a
process::terminate() that coordinates with Clock to remove all outstanding
clocks destined for that process. If there is some sort of race where another
process has the pid cached and is still sending stuff to it, that is probably a
bug in the code for shutting down a particular process.
I have a patch to fix waiter specifically which I'll post the review for
shortly / once it finishes passing my test build.
> libprocess Clock messages delivered
> ------------------------------------
>
> Key: MESOS-2484
> URL: https://issues.apache.org/jira/browse/MESOS-2484
> Project: Mesos
> Issue Type: Bug
> Components: libprocess
> Reporter: Cody Maloney
>
> Found in / discussed at: https://reviews.apache.org/r/30587/#rc118737-72676
> When a process is terminated, any outstanding delay() destined for that
> process aren't terminated, meaning they arrive whenever the clock happens to
> get there. With uniquely named processes this isn't an issue, but with names
> that are reused (master), it could potentially lead to odd test flakiness,
> and is artifacts carrying across / between tests which shouldn't be.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
