Setup:
* PostgreSQL 9.3.9
* 1 master, 1 replica
* Tiny database, under 0.5GB, completely cached in shared_buffers
* 90% read query traffic, which is handled by replica
* Traffic in the 1000's QPS.
The wierdness:
Periodically the master runs an "update all rows" query on the main
table in the database. When this update hits the replica via
replication stream, *some* (about 5%) of the queries which do seq scans
will stall for 22 to 32 seconds (these queries normally take about
75ms). Queries which do index scans seem not to be affected.
Thing is, the update all rows only takes 2.5 seconds to execute on the
master. So even if the update is blocking the seq scans on the replica
(and I can't see why it would), it should only block them for < 3 seconds.
Anyone seen anything like this?
--
Josh Berkus
PostgreSQL Experts Inc.
http://pgexperts.com