On 5.12.2011 00:06, Tom Lane wrote:
> Tomas Vondra <tv@fuzzy.cz> writes:
>> On 4.12.2011 22:16, Tom Lane wrote:
>>> Um ... why would you design it like that?
>
>> The backends are added to pg_stat_activity after the auth hook finishes,
>> which means possible race conditions (backends executed at about the
>> same time don't see each other in pg_stat_activity). So I use an
>> exclusive lock that's acquired before reading pg_stat_activity and
>> released after the pg_stat_activity is updated.
>> That's the only thing the library loaded using local_preload_libraries
>> does - it releases the lock.
>
> That's an unbelievably ugly, and dangerous, kluge. All you need is one
> backend not loading the second library (and remember,
> local_preload_libraries is user-settable) and you've just locked up the
> system.
Yes, but I wasn't aware of that - I thought local_preload_libraries is
defined in postgresql.conf so it seemed fine (yes, it was ugly but it
did not seem that dangerous).
> Why are you using pg_stat_activity for this anyway? Searching the
> ProcArray seems much safer ... see CountDBBackends for an example.
Because I've never user ProcArray before, but I use pg_stat_activity
quite frequently, so it seemed very natural. Thanks for pointing to
ProcArray/CountDBBackends, I'll see how to use that.
Tomas
