Re: [COMMITTERS] pgsql: Teach tuplesort.c about "top N" sorting, in which only the first

Magnus Hagander wrote:
> Alvaro Herrera wrote:
>> Jim Nasby wrote:
>>
>>> If we really want to make the logfile the approved method for  
>>> monitoring performance, then why do we have the stats infrastructure  
>>> at all? It could all be replaced with logging output and pgfouine.
>> First we'd have to fix the usability problem of our redirect_stderr
>> stuff for pgfouine (multiline messages from different backends are
>> mixed, according to Guillaume).
> 
> I'd like to sign on to the list of people saying that logging isn't the
> best way to do performance measuring. Providing a way to get at the
> counters in realtime for monitoring or graphing or whatever is what
> AFAIK everybody else do, and it's for a good reason - it fits in to
> existing monitoring/management solutions. It may be that our current
> stats system isn't the best way to do this, but if it isn't that just
> means we have to come up with a better way.


Speaking of which, it might be interesting to actually show these values
in the stats collector. I was thinking three cols for each database
(probably the best level?) that counts each of those three counters. If
you have a lot of sorts (percentage-wise) spilling to disk, it is often
something you want to investigate, so exposing it that way seems like a
good thing.

Comments on that? Oh, and is it too late to sneak this into 8.3 claiming
it's an extension to the addition of the new sort feature? ;-)

//Magnus