Thread: Better detection of staled postmaster.pid
This is most likely just a request for brainstorm. It's been reported [1] that postmaster fails to start against staled postmaster.pid after (e.g.) power outage on Fedora, its due to init system parallelism and "some" other newly started process can already have allocated the same PID as the old postmaster had -- and in this case postmaster refuses to delete staled pidfile (which is expected as we need to be really careful). Don't you see some other possible check we could implement to guarantee that the PID mentioned in postmaster.pid does not hide concurrent postmaster? I can think of /proc/<CONCURRENT_PID>/cmdline parsing for possible '-D' option occurrence, but that is not terribly portable and it could be considered racy, or? Some acceptable hack we could use to tell to other processes that we are running particular data directory? [1] https://bugzilla.redhat.com/show_bug.cgi?id=1257334 Pavel
Pavel Raiskup <praiskup@redhat.com> wrote: > It's been reported [1] that postmaster fails to start against staled > postmaster.pid after (e.g.) power outage on Fedora, its due to init system > parallelism and "some" other newly started process can already have allocated > the same PID as the old postmaster had -- and in this case postmaster refuses > to delete staled pidfile (which is expected as we need to be really > careful). > > Don't you see some other possible check we could implement to guarantee that > the PID mentioned in postmaster.pid does not hide concurrent postmaster? Was the other newly started process another PostgreSQL cluster? Was it launched under the same OS user? (Those are the only conditions under which I've seen this.) I think it is wise to use a separate OS user for each cluster. If it's not a matter of multiple clusters running under the same OS user, please provide more deails, like the specific version and copy/paste of error messages and relevant log entries. -- Kevin Grittner EDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
Pavel Raiskup <praiskup@redhat.com> wrote:
> It's been reported [1] that postmaster fails to start against staled
> postmaster.pid after (e.g.) power outage on Fedora, its due to init system
> parallelism and "some" other newly started process can already have allocated
> the same PID as the old postmaster had -- and in this case postmaster refuses
> to delete staled pidfile (which is expected as we need to be really
> careful).
>
> Don't you see some other possible check we could implement to guarantee that
> the PID mentioned in postmaster.pid does not hide concurrent postmaster?
Most of this can be gleamed from the linked bug report...
Was the other newly started process another PostgreSQL cluster?
Yes
Was it launched under the same OS user? (Those are the only
conditions under which I've seen this.) I think it is wise to use
a separate OS user for each cluster.
Yes. Does the pid check that the owner of the pid file match the owner of the process? While seemingly good advice I'm not sure how it would prevent this scenario - likely due to lack of knowledge on my part.
If it's not a matter of multiple clusters running under the same OS
user, please provide more deails, like the specific version and
copy/paste of error messages and relevant log entries
See report. I get not having transient data linked to in these kinds of postings but the supplied description and official downstream project bug report seem like sufficient data work operate from even if only in a preliminary fashion.
The only obvious solution is to stop using (pid) as a primary key of sorts and use (pid, timecreated) instead. After a restart/reboot the timecreated would be guaranteed to have changed and no guessing would be involved. That seems invasive, though proper, for a problem largely limited to an uncommon distribution-specific setup that requires a unclean shutdown to occur.
David J.
Kevin Grittner <kgrittn@ymail.com> writes: > Pavel Raiskup <praiskup@redhat.com> wrote: >> It's been reported [1] that postmaster fails to start against staled >> postmaster.pid after (e.g.) power outage on Fedora, > Was the other newly started process another PostgreSQL cluster? > Was it launched under the same OS user? Yes, that's what the bugzilla report indicated. > (Those are the only > conditions under which I've seen this.) I think it is wise to use > a separate OS user for each cluster. That's my recommendation too. The only other thing you could do to prevent this would be to manually blow away postmaster.pid files, and please listen to this: that cure is a lot worse than the disease. It's almost never implemented safely (that is, in a way that guarantees the forcible removal can *only* happen at system boot and never later). regards, tom lane