Tom Lane schrieb:
> Peter Hinse <loco@d0pefish.de> writes:
>> More info: the statement is an INSERT with some huge subselects, running
>> every night on a PGSQL 8.3.6 on CentOS 4.7 x86_64. In 97% of all
>> occasions, the job terminates in about 1-2 minutes - however, sometimes
>> it just hangs. If terminated with kill <pid> and restarted, it always
>> terminates.
>
> Define "huge" --- you mean a lot of relations in the query? If you have
> enough to trigger GEQO optimization, it could be that it's sometimes
> picking a bad plan. You might try raising the geqo threshold to more
> relations than that; or if this results in unacceptably long planning
> time, increase geqo_effort instead.
After downgrading to 8.3.5, the massive problems we had with our system
are gone. With 8.3.6, we had these hanging jobs in 100% of all runs
(this is a DWH, the job calculates the ranking for users). We tried
several settings for GEQO, that made it even worse - some jobs, that
never had any problems, started to act weird (did not terminate) with
the same lseek in strace.
We will investigate this issue with some more test machines.
Regards,
Peter
