Re: Patch to implement pg_current_logfile() function

On Tue, 28 Jun 2016 11:06:24 +0200
Gilles Darold <gilles.darold@dalibo.com> wrote:

> Le 07/04/2016 08:30, Karl O. Pinc a écrit :

> > "src/backend/postmaster/syslogger.c expects to see fopen() fail
> > with
> ENFILE and EMFILE.  What will you do if you get these?"
>
>     - Nothing, if the problem occurs during the log rotate call, then
> log rotation file is disabled so logfile_writename() will not be
> called. Case where the problem occurs between the rotation call and
> the logfile_writename() call is possible but I don't think that it
> will be useful to try again. In this last case log filename will be
> updated during next rotation.

The case I'm interested in is when the rotation call succeeds but
you get ENFILE or EMFILE in logfile_writename() and current_logfiles
is not updated.

This looks like an ugly problem that only happens
sporadically under load.  If I've set log
rotation to, say, 1 week, and I'm post-processing my logs based
on the current_logfiles content, and the logs rotate but
current_logfiles is not updated, then I lose a week of log
post-processing.

I'm experimenting with some code that retries writing current_logfiles,
if it failed due to ENFILE or EMFILE, the next time a log message
is written.  If that's too often it'd be easy enough to add
a backoff counter that retries progressively less frequently
based on a "clock tick" per log message write.

However, per my last email, it'll be Tuesday before I really get
back to this.   Let me know if, instead, you want to jump in
and write the code, have a better idea, think this is a really
stupid approach, or have other reasons why I should abandon
this plan.

Regards,

Karl <kop@meme.com>
Free Software:  "You don't pay back, you pay forward."                -- Robert A. Heinlein