[cinder][tooz]Lock-files are remained
cjeanner at redhat.com
Tue Oct 15 15:24:55 UTC 2019
On 10/15/19 4:47 PM, Ben Nemec wrote:
> On 10/15/19 7:48 AM, Herve Beraud wrote:
>> I proposed some patches through heat templates and puppet-cinder to
>> remove lock files older than 1 week and avoid file system growing.
>> This is a solution based on a cron job, to fix that on stable
>> branches, in a second time I'll help the fasteners project to fix the
>> root cause by reviewing and testing the proposed patch (lock based on
>> file offset). In next versions I hope we will use a patched fasteners
>> and so we could drop the cron based solution.
>> Please can you give /reviews/feedbacks:
>> - https://review.opendev.org/688413
>> - https://review.opendev.org/688414
>> - https://review.opendev.org/688415
> I'm rather hesitant to recommend this. It looks like the change is only
> removing the -delete lock files, which are a fraction of the total lock
> files created by Cinder, and I don't particularly want to encourage
> people to start monkeying with the lock files while a service is
> running. Even with this limited set of deletions, I think we need a
> Cinder person to look and verify that we aren't making bad assumptions
> about how the locks are used.
I'm also not that happy with the cron way - but apparently it might
create some issues in some setup with the current way things are done:
- inodes aren't infinit on ext* FS (ext3, ext4, blah) - see bellow for
- perfs might be bad (see bellow for context)
So one way or another, cleanup is needed.
> In essence, I don't think this is going to meaningfully reduce the
> amount of leftover lock files and it sets a bad precedent for how to
> handle them.
The filter is strict - on purpose, to address this specific issue. Of
course, we might want to loosen it, but... do we really want that? IF
we're to go with the cron thingy of course. Some more thinking is needed
> Personally, I'd rather see a boot-time service added for each OpenStack
> service that goes out and wipes the lock file directory before starting
> the service.
Well.... former sysops here: don't count on regular reboot - once a year
is a fair average - and it's usually due to some power cut... Sad world,
I know ;).
So a "boot-time cleanup" will help. A little. And wouldn't hurt anyway.
So +1 for that idea, but I wouldn't rely only on it. And there might be
some issues (see bellow)
> On a more general note, I'm going to challenge the assertion that
> "Customer file system growing slowly and so customer risk to facing some
> issues to file system usage after a long period." I have yet to hear an
> actual bug report from the leftover lock files. Every time this comes up
> it's because someone noticed a lot of lock files and thought we were
> leaking them. I've never heard anyone report an actual functional or
> performance problem as a result of the lock files. I don't think we
> should "fix" this until someone reports that it's actually broken.
> Especially because previous attempts have all resulted in very real bugs
> that did break people.
I'm not on your side here. Waiting to get a fire before thinking about
correction and counter-measures isn't good. Since we know there's an
issue, and that, eventually, it might be a really big one, it would be
good to address it before it explodes in our face.
The disk *space* is probably not the issue. File with 1k, on a couple of
gigas, it's good.
But there are other concerns:
- inodes. Yes, we're supposed to get things on XFS, and that dude
doesn't have inodes. But some ops might want to rely on the good(?) old
ext3, or ext4. There, we might get some issues, and pretty quickly
depending on the speed of lock creation (so, linked to cinder actions in
this case). Or it might be some NFS with an ext4 FS.
- rm limits: you probably never ever hit "rm argument list limit". But
it does exist, and I did hit it (like 15 years ago - maybe it's sorted
out, but I have some doubts). This means that rm will NOT be able to
cope with the directory content after a certain amount (which is huge,
right, but still... we're talking about some long-lasting process
filling a directory).
Of course, "find" might be the right tool in such case, but it will take
a long, long time to delete (thinking about the "boot-time cleanup
proposal" for instance).
- performances: it might impact the system, for instance if one has some
backup process running and wanting to eat the /var/lib/cinder directory:
if the op doesn't know about this issue, they might get some surprises
with long, long, loooong running backups.
With a target on some ext4 - hello Inodes!
Sooo... yeah. I think this issue should be addressed. Really. But I +1
the fact that it should be done "the right way", whatever it is. The
"cron" might be the wrong one. Or not. We need some more feedbacks on
> Maybe we should have oslo.concurrency drop a file named _README (or
> something else likely to sort first in the file listing) into the
> configured lock_path that explains why the files are there and the
> proper way to deal with them.
Hmmm... who reads README anyway? Like, really. Better getting some
cleanup un place to avoid questions ;).
>> Le lun. 30 sept. 2019 à 03:35, Rikimaru Honjo
>> <honjo.rikimaru at ntt-tx.co.jp <mailto:honjo.rikimaru at ntt-tx.co.jp>> a
>> écrit :
>> On 2019/09/28 1:44, Ben Nemec wrote:
>> > On 9/23/19 11:42 PM, Rikimaru Honjo wrote:
>> >> Hi Eric,
>> >> On 2019/09/20 23:10, Eric Harney wrote:
>> >>> On 9/20/19 1:52 AM, Rikimaru Honjo wrote:
>> >>>> Hi,
>> >>>> I'm using Queens cinder with the following setting.
>> >>>> ---------------------------------
>> >>>> [coordination]
>> >>>> backend_url = file://$state_path
>> >>>> ---------------------------------
>> >>>> As a result, the files like the following were remained under
>> the state path after some operations.
>> >>>> cinder-63dacb3d-bd4d-42bb-88fe-6e4180164765-delete_volume
>> >>>> cinder-32c426af-82b4-41de-b637-7d76fed69e83-delete_snapshot
>> >>>> In my understanding, these are lock-files created for
>> synchronization by tooz.
>> >>>> But, these lock-files were not deleted after finishing
>> >>>> Is this behaviour correct?
>> >>>> 
>> >>>> e.g. Delete volume, Delete snapshot
>> >>> This is a known bug that's described here:
>> >>> https://github.com/harlowja/fasteners/issues/26
>> >>> (The fasteners library is used by tooz, which is used by Cinder
>> for managing these lock files.)
>> >>> There's an old Cinder bug for it here:
>> >>> https://bugs.launchpad.net/cinder/+bug/1432387
>> >>> but that's marked as "Won't Fix" because Cinder needs it to be
>> fixed in the underlying libraries.
>> >> Thank you for your explanation.
>> >> I understood the state.
>> >> But, I have one more question.
>> >> Can I think this bug doesn't affect synchronization?
>> > It does not. In fact, it's important to not remove lock files
>> while a service is running or you can end up with synchronization
>> > To clean up the leftover lock files, we generally recommend
>> clearing the lock_path for each service on reboot before the
>> services have started.
>> Thank you for your information.
>> I think that I understood this issue completely.
>> Best Regards,
>> >> Best regards,
>> >>> Thanks,
>> >>> Eric
>> -- _/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/_/
>> Rikimaru Honjo
>> E-mail:honjo.rikimaru at ntt-tx.co.jp
>> <mailto:E-mail%3Ahonjo.rikimaru at ntt-tx.co.jp>
>> Hervé Beraud
>> Senior Software Engineer
>> Red Hat - Openstack Oslo
>> irc: hberaud
>> -----BEGIN PGP SIGNATURE-----
>> -----END PGP SIGNATURE-----
Cédric Jeanneret (He/Him/His)
Software Engineer - OpenStack Platform
Red Hat EMEA
More information about the openstack-discuss