dimitri@2ndQuadrant.fr wrote:
>
> I witnessed a case where I had no time to extract information about it, so
> it's all from memory, sorry about that.
>
> We had several Hot Standby nodes in 9.3.x used for load balancing read only
> queries. All queries were stuck, and pg_locks showed they were refused a
> lock against pg_catalog.pg_statistics.
My WAG is that this is related to the standby replaying the
AccessExclusive lock obtained to truncate pg_statistics. That seems
consistent with the presented symptoms.
I would have assumed that the actual truncation shouldn't take a
particularly long time, so user processes on the standby wouldn't be
locked for long enough to cause visible trouble. Maybe the table was
originally very bloated and a lot of pages got removed, leading the
replay process to take some time. It is strange that it would be locked
for long enough to allow for manually querying pg_locks and
pg_stat_activity.
--
Álvaro Herrera http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services
