Обсуждение: problem (bug?) when killing client program

I don't know if this is a bug or feature, but here goes the report.

When trying to backup a table that lives on a database with the same
name, I ended issuing the wrong command and, instead, started a
full dump of a multigigabyte database with pg_dump.

A couple seconds later, just pressed Ctrl-C. The pg_dump, as desired,
quit. After that, the follwing line was found on `ps` output:

postgres 22332 26.4  4.5 96264 69980 pts/2   R    10:32   3:24 postgres:
postgres <dbname> [local] COPY

The pg_dump process was confirmed killed and the uptime output, at the
same time was:

 10:41:02  up 14 days, 19:56,  3 users,  load average: 20.11, 11.48, 10.28

Realizing that the things weren't quite right, sent a SIGTERM (kill with
no flags) to the PID 22332.

What happened? The entire backend crashed and burned in enormous flames.

We're using postgres 7.4.1 on a Linux/i386 SMP machine.

Ok, it wasn't nice to start a full backup when it wasn't the real intention,
but shouldn't the backend process deal with these situations without
crashing down and causing minors data loss in the recovery??

TIA.

"Marcus Andree S. Magalhaes" <marcus.magalhaes@vlinfo.com.br> writes:
> Realizing that the things weren't quite right, sent a SIGTERM (kill with
> no flags) to the PID 22332.
> What happened? The entire backend crashed and burned in enormous flames.

SIGTERM should have been fine.  Are you certain you didn't do kill -9 ?

I tried the same thing here, and the backend shut down just fine, with
only this entry in the postmaster log:
    FATAL:  terminating connection due to administrator command
So I think you didn't really do SIGTERM.

            regards, tom lane

Yes, I sent a SIGTERM...

Anyway, the fact that a postgres process was alive & kicking & eating
CPU power after pg_dump exited is a normal behaviour?

Best regards.

<snip>
>
> SIGTERM should have been fine.  Are you certain you didn't do kill -9 ?
>
> I tried the same thing here, and the backend shut down just fine, with
> only this entry in the postmaster log:
>     FATAL:  terminating connection due to administrator command
> So I think you didn't really do SIGTERM.
>
>             regards, tom lane

"Marcus Andree S. Magalhaes" <marcus.magalhaes@vlinfo.com.br> writes:
> Anyway, the fact that a postgres process was alive & kicking & eating
> CPU power after pg_dump exited is a normal behaviour?

Yes; it'll finish out the current command before noticing that the
client connection was dropped.  In this case the current command
happened to be a long-running one ...

If you can reproduce the crash-on-SIGTERM behavior I'd like to know how.
*That*, if real, is a bug.

            regards, tom lane