On 10 March 2014 15:32, hubert depesz lubaczewski <depesz@depesz.com> wrote:
> On Thu, Mar 06, 2014 at 06:03:54PM +0100, hubert depesz lubaczewski wrote:
>> On Thu, Mar 06, 2014 at 12:02:50PM -0500, Tom Lane wrote:
>> > hubert depesz lubaczewski <depesz@depesz.com> writes:
>> > > I didn't have a chance to do it. Can try if there is a way to get trace
>> > > *without* making core (sorry, my c/gdb knowledge is very, very limited).
>> >
>> > Sure, you just attach to the process:
>> >
>> > $ gdb /path/to/postgres PID-of-process
>> > gdb> bt
>> > gdb> quit
>> >
>> > This is usually preferable to forcing a core dump.
>>
>> Thank you. If the problem will strike again, I will do it on all (or
>> most, depending how fast I can make it) backends.
>
> The problem did happen again, and we were able to find a fix (I think).
> For some reason we had a table with over 50000 (yes, 50 thousand)
> indexes on it. This table was a bucardo internals table, so maybe it was
> something in bucardo (we are using it to migrate hundreds of tables to
> another machine, so maybe it has something to do with it.
This sort of thing is the reason why I'd want to see index maintenance
nodes in explain (analyse) plans, so that it's possible to gauge their
contribution to the overall duration of a DML statement.
--
Thom