RE: WAL Log using all my disk space!

From: "Mikheev, Vadim" <vmikheev(at)SECTORBASE(dot)COM>
To: "'webb sprague'" <wsprague(at)o1(dot)com>, pgsql-general(at)postgresql(dot)org
Subject: RE: WAL Log using all my disk space!
Date: 2001-04-27 20:45:39
Message-ID: 8F4C99C66D04D4118F580090272A7A234D33DD@sectorbase1.sectorbase.com
Views: Raw Message | Whole Thread | Download mbox | Resend email
Thread:
Lists: pgsql-general

> I do a large bulk copy once a day (100,000 records of Radius data),
> tearing down indices, truncating a large table that contains summary
> information, and rebuilding everything after the copy. Over the course
> of this operation, I can generate up to 1.5 gigs of WAL data in
> pg_xlog. Sometimes (like just now), I will run out of disk space and
> the postmaster will crash. I try to restart it, and it errors out.
> Then I delete all the WAL logs, try to restart, and (surprise) it errors
> out again.

(Removing WAL logs is not good idea).

> I tried to set some of the of the WAL parameters in postgres.conf like
> so:
>
> wal_buffers = 4 # min 4

(More is better and doesn't affect disk space usage).

> wal_files = 8 # range 0-64

Ops. With wal_files > 0 server pre-allocates log files in advance!
Should be used only if disk space is not problem...

> I would like to recover without an initdb, but if that isn't
> possible, I would definitely like to avoid this problem in the
> future.

So, are you able to restart? If not - send us startup server log.
You should be able to remove some of preallocated wal_files but
I need in numbers from server log to say what is safe to remove.

Now how to reduce disk space usage.
First to keep in mind - server removes old (useless) log files at
checkpoint time. Second - log file becomes useless (from transaction
subsystem POV) if it keeps no record from any running transaction.
Third - unfortunately (from my POV), we requires two checkpoint in
log files now, so we do not remove files with records between last
two checkpoints.
Recommendation: try to split your bulk operation into a few transactions
with smaller write traffic and run CHECKPOINT commands between them.
You could also try to change checkpoint_segments and/or checkpoint_timeout
params, but imho explicit CHECKPOINT is better for bulk ops, because of
it will not affect normal operations.

Vadim

Responses

Browse pgsql-general by date

  From Date Subject
Next Message Vince Vielhaber 2001-04-27 20:50:26 Re: [HACKERS] While we're on the subject of searches...
Previous Message Jan Wieck 2001-04-27 20:12:27 Re: PostgreSQL and mySQL database size question