Thread: high throughput 9.3, master/slave question

Hi,

I have a postgres server that is fairly high traffic. (about 5MB/second
in writes to SSD backed Netapp NFS storage).  The outbound network
traffic for this machine averages 61MB/second, but gets over
250MB/second during peal times.  (includes data to slave + user requests)

Initially, I had my application servers using the slave for short, read
only queries, but this turned out to be highly unstable. The slave would
start refusing connections, and the logs would fill with:

ERROR:  canceling statement due to conflict with recovery

I've tried these 2 settings:

max_standby_archive_delay = -1
max_standby_streaming_delay = -1

But then I starting getting these:
DETAIL:  User transaction caused buffer deadlock with recovery.

Read requests come in at anywhere between 200 and 1000/second.

I was wondering if there is some combination of configuration settings
that would safely let me use the slave for read only queries?

-Matt

On Fri, Feb 28, 2014 at 1:54 PM, Matthew Chambers <mchambers@wetafx.co.nz> wrote:

Initially, I had my application servers using the slave for short, read only queries, but this turned out to be highly unstable. The slave would start refusing connections, and the logs would fill with:

ERROR:  canceling statement due to conflict with recovery

I've tried these 2 settings:

max_standby_archive_delay = -1
max_standby_streaming_delay = -1

But then I starting getting these:
DETAIL:  User transaction caused buffer deadlock with recovery.

Read requests come in at anywhere between 200 and 1000/second.

I was wondering if there is some combination of configuration settings that would safely let me use the slave for read only queries?

Have you tried setting max_standby_archive|streaming_delay to several minutes (or whatever makes sense for your typical query durations), rather than disabling those settings with -1?