On 2018-04-16 14:03:40 -0300, Alvaro Herrera wrote:
> Peter Geoghegan wrote:
>
> > This now seems unnecessary, since it's already evident from your "bt
> > full" output that the query involved in the crash was "select count(*)
> > from pg_buffercache where isdirty".
>
> Hmm, so is the Zabbix monitoring running that query frequently?
> Because, as I recall, pg_buffercache is pretty heavy on the system,
> since it needs to acquire all the bufmgr locks simultaneously?
>
> In other words, this seems a terrible query to be running in zabbix.
Can be extremely useful however, to predict how much longer your
workload's hot data set fits into cache. That's worth the cost in a
number of cases... Either way, a crash is clearly something separate.
> I have vague memories of somebody submitting a version of this code
> that returned approximate answers, good enough for monitoring ...
That might have been me, but I don't recall the details anymore...
Greetings,
Andres Freund