Re: Patch review for logging hooks (CF 2012-01)

Excerpts from Andrew Dunstan's message of mié ene 18 13:27:40 -0300 2012:
>
> On 01/18/2012 11:12 AM, Martin Pihlak wrote:
> > On 01/18/2012 03:56 AM, Fujii Masao wrote:
> >> or syslog process (if you use syslog). So ISTM that there is no
> >> guarantee that the order of log messages processed by the
> >> hook is same as that of messages written to the log file. For
> >> example, imagine the case where two backends call EmitErrorReport()
> >> at the same time. Is this OK? If not, the hook might need to be
> >> in syslogger.
> > For high volume logging I'd avoid going through the syslogger. One
> > big issue with syslogger is that it creates a choke point - everything
> > has to pass through it, and if it cannot keep up it starts stalling
> > the backends. Also, in EmitErrorReport the hook gets to have access
> > to the actual ErrorData structure -- that makes filtering and looking
> > at message content much simpler.
>
> Hmm, interesting. I don't think I've encountered a situation where
> backends would actually stall.

You have to have really high velocity for this to happen.  At least one
customer of ours has suffered this problem (I vaguely recall a second
case but I'm not really sure), having had to switch to syslog (which
uses lossy sockets, with the advantage that it doesn't cause stalls).

> But in any case, I don't think we have to
> be that deterministic. The only thing that needs to be absolutely
> guaranteed is that the log messages from a given backend are in order.
> Some slight fuzz between backends seems acceptable.

Agreed.

--
Álvaro Herrera <alvherre@commandprompt.com>
The PostgreSQL Company - Command Prompt, Inc.
PostgreSQL Replication, Consulting, Custom Development, 24x7 support