Re: strange hot_standby behaviour

From: Lonni J Friedman <netllama(at)gmail(dot)com>
To: pfote <pfote(at)ypsilon(dot)net>
Cc: pgsql-general(at)postgresql(dot)org
Subject: Re: strange hot_standby behaviour
Date: 2012-10-01 14:56:16
Message-ID: CAP=oouHRZuqN7xqz7VEG25hWUd9mDQtu4oarfg0_GWEkzuDHgg@mail.gmail.com
Views: Raw Message | Whole Thread | Download mbox | Resend email
Thread:
Lists: pgsql-general

On Mon, Oct 1, 2012 at 7:28 AM, pfote <pfote(at)ypsilon(dot)net> wrote:
> Hi,
>
> I had a very strange effect on the weekend that smells like a bug, so i'd
> like so share it.
>
> Setup:
> machine A: 16 CPU Cores (modern), 128GB RAM, nice 6-drive SAS Raid-10
> machines B, C: 8 Cores (substantially older than A), 48GB Ram, some scsi
> Raid, substantially slower than A
>
> The workload is about 80% - 90% SELECTs with heavy sorting and grouping, the
> remaining are INSERTs/UPDATEs/DELETEs.
> So In the original setup A is the master, B and C are hot standby's that
> process some of the SELECTs, but by far the most processing is done on the
> master (A). pg version is 9.0.6. CPU utilization is about 80% on the master
> and between 90-100% in the standby's, so it's decided to upgrade to the
> latest 9.2 to profit from the latest performance enhancements.
>
> So B gets upgraded to 9.2.1-1.pgdg60+1 (from pgapt.debian.org) and becomes
> master, then A becomes a hot_standby slave that takes all the SELECTs (and C
> becomes another hot_standby). In the beginning everything works as expected,
> CPU utilization drops from 80% to about 50-60%, selects run faster,
> everything looks smoother (some queries drop from >5s to <1s due to 9.2s
> index-only-scan feature). Its friday, everyone is happy.
>
> About 16 hours later, saturday morning around 6:00, A suddenly goes wild and
> has a CPU utilization of 100% without a change in the workload, out of the
> blue. Queries that used to take <1s suddenly take 5-10s, "explain analyze"
> plans of these queries havn't change a bit though. Switching the workload
> off causes the server to become idle. (while I'm writing this I realize we
> haven't tried to restart A). Instead, $boss decides to twitch back to the
> original setup, so B gets dropped, A becomes master and gets 100% of the
> workload (all SELECTs/INSERTs/UPDATEs/DELETEs), and everything becomes just
> like friday, CPU usage drops to 50-60%, everything runs smothly.
>
> I'm not sure yet if this is replication related or a 9.2.1 problem. Any
> Ideas?

This could be just about anything. Which OS are you running? Did you
check any logs when everything went crazy?

In response to

Responses

Browse pgsql-general by date

  From Date Subject
Next Message Vick Khera 2012-10-01 14:57:45 Re: shared memory settings
Previous Message Carrington, Matthew (Produban) 2012-10-01 14:50:41 Re: pg_upgrade: out of memory