Re: [GENERAL] pg_xlog on a hot_standby slave filling up

Sent from my iPhone

> On Jun 17, 2015, at 03:22, Christoph Berg <cb(at)df7cb(dot)de> wrote:
>
> Re: Jeff Frost 2015-06-16 <67E2F20A-6A2E-484E-BF97-544F1FC66566(at)pgexperts(dot)com>
>
> Yes, SSL, with default settings. I can confirm your
> wal-file-from-before analysis:
>
> Jun 16 07:14:59 synthesis postgres[32525]: [8-1] 2015-06-16 07:14:59 GMT LOG: unexpected pageaddr 33/39000000 in log segment 000000010000003300000040, offset 0
> Jun 16 07:15:00 synthesis postgres[11514]: [3-1] 2015-06-16 07:15:00 GMT LOG: started streaming WAL from primary at 33/40000000 on timeline 1
>
> -rw------- 1 postgres postgres 16777216 Jun 16 09:14 00000001000000330000003F
>
> Jun 16 17:55:01 synthesis postgres[32525]: [9-1] 2015-06-16 17:55:01 GMT LOG: unexpected pageaddr 33/5A000000 in log segment 000000010000003300000060, offset 0
> Jun 16 17:55:02 synthesis postgres[24337]: [3-1] 2015-06-16 17:55:02 GMT LOG: started streaming WAL from primary at 33/60000000 on timeline 1
>
> -rw------- 1 postgres postgres 16777216 Jun 16 19:55 00000001000000330000005F
>
> Jun 17 04:35:02 synthesis postgres[24337]: [4-1] 2015-06-17 04:35:02 GMT FATAL: could not send data to WAL stream: server closed the connection unexpectedly
> Jun 17 04:35:02 synthesis postgres[24337]: [4-2] This probably means the server terminated abnormally
> Jun 17 04:35:02 synthesis postgres[24337]: [4-3] before or while processing the request.
> Jun 17 04:35:02 synthesis postgres[24337]: [4-4]
> Jun 17 04:35:04 synthesis postgres[32525]: [10-1] 2015-06-17 04:35:04 GMT LOG: unexpected pageaddr 33/7B000000 in log segment 000000010000003300000080, offset 0
> Jun 17 04:35:05 synthesis postgres[4756]: [5-1] 2015-06-17 04:35:05 GMT LOG: started streaming WAL from primary at 33/80000000 on timeline 1
>
> -rw------- 1 postgres postgres 16777216 Jun 17 06:35 00000001000000330000007F
>
> There's a 1:1 correspondence with log and leaked files.

We thought it was related to the ssl renegotiation limit, but reducing it didn't seem to make it happen more often.

The problem was that I couldn't seem to make a reproducible test case with pgbench and two servers, so it seems there is slightly more at play.

I believe setting the ssl renegotiation limit to 0 made it stop. Can you confirm?

Have you been able to reproduce synthetically?