Обсуждение: WAL partition filling up after high WAL activity

Hi Everyone,

On 11/07/2011 05:18 PM, Richard Yen wrote:

My biggest question is: we know from the docs that there should be no more than (2 + checkpoint_completion_target) * checkpoint_segments + 1 files.  For us, that would mean no more than 48 files, which equates to 384MB--far lower than the 9.7GB partition size.  **Why would WAL use up so much disk space?**

That's only true if things are operating normally.  There are at least two ways this can fail to be a proper upper limit on space used:

1) You are archiving to a second system, and the archiving isn't keeping up.  Things that haven't been archived can't be re-used, so more disk space is used.

2) Disk I/O is slow, and the checkpoint writes take a significant period of time.  The internal scheduling assumes each individual write will happen without too much delay.  That assumption can easily be untrue on a busy system.  The worst I've seen now are checkpoints that take 6 hours to sync, where the time is supposed to be a few seconds.  Disk space in that case was a giant multiple of checkpoint_segments.  (The source of that problem is very much improved in PostgreSQL 9.1)

The info needed to figure out which category you're in would appear after tuning log_checkpoints on in the postgresql.conf ; you only need to reload the server config after that, doesn't require a restart.  I would guess you have realy long sync times there.

As for what to do about it, checkpoint_segments=16 is a low setting.  You might as well set it to a large number, say 128, and let checkpoints get driven by time instead.  The existing limit isn't working effectively anyway, and having more segments lets the checkpoint spreading code work more evenly.

-- 
Greg Smith   2ndQuadrant US    greg@2ndQuadrant.com   Baltimore, MD
PostgreSQL Training, Services, and 24x7 Support  www.2ndQuadrant.us

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

On 11/09/2011 05:06 PM, Greg Smith wrote:
> On 11/07/2011 05:18 PM, Richard Yen wrote:
>> My biggest question is: we know from the docs that there should be no
>> more than (2 + checkpoint_completion_target) * checkpoint_segments + 1
>> files.  For us, that would mean no more than 48 files, which equates
>> to 384MB--far lower than the 9.7GB partition size.  **Why would WAL
>> use up so much disk space?**
>>
>
> That's only true if things are operating normally.  There are at least
> two ways this can fail to be a proper upper limit on space used:
>
> 1) You are archiving to a second system, and the archiving isn't keeping
> up.  Things that haven't been archived can't be re-used, so more disk
> space is used.
>
> 2) Disk I/O is slow, and the checkpoint writes take a significant period
> of time.  The internal scheduling assumes each individual write will
> happen without too much delay.  That assumption can easily be untrue on
> a busy system.  The worst I've seen now are checkpoints that take 6
> hours to sync, where the time is supposed to be a few seconds.  Disk
> space in that case was a giant multiple of checkpoint_segments.  (The
> source of that problem is very much improved in PostgreSQL 9.1)
>


Hello

We have a similar case in june but we did not find the cause of our
problem. More details and information:
http://archives.postgresql.org/pgsql-docs/2011-06/msg00007.php

Your explanation in 2) sounds like a good candidate for the problem we
had. As I said in june, I think we need to improve the documentation in
this area. A note in the documentation about what you have explained in
2) with maybe some hints about how to find out if this is happening will
be a great improvement.

We did not understand why we experienced this problem in june when
creating a GIN index on a tsvector column. But we found out that a lot
of the tsvector data was generated from "garbage" data (base64 encoding
of huge attachments). When we generated the right tsvector data, the
creation of the GIN index ran smoothly and the problem with extra WAL
files disappeared.

PS.- In our case, the disk space used by all the extra WAL files was
almost the equivalent to the 17GB of our GIN index.

regards,
- --
 Rafael Martinez Guerrero
 Center for Information Technology
 University of Oslo, Norway

 PGP Public Key: http://folk.uio.no/rafael/
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.10 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org/

iEYEARECAAYFAk688LoACgkQBhuKQurGihTbvQCfaSBdYNF2oOtErcx/e4u0Zw1J
pLIAn2Ztdbuz33es2uw8ddSIjj8UXe3s
=olkD
-----END PGP SIGNATURE-----

On 11/11/2011 04:54 AM, Rafael Martinez wrote:
> Your explanation in 2) sounds like a good candidate for the problem we
> had. As I said in june, I think we need to improve the documentation in
> this area. A note in the documentation about what you have explained in
> 2) with maybe some hints about how to find out if this is happening will
> be a great improvement.
>

A new counter was added to pg_stat_bgwriter in PostgreSQL 9.1 that
tracks when the problem I described happens.  It's hard to identify it
specifically without a source code change of some sort.  Initially I
added new logging to the server code to identify the issue before the
new counter was there.  The only thing you can easily look at that tends
to correlate well with the worst problems here is the output from
turning log_checkpoint on.  Specifically, the "sync" times going way up
is a sign there's a problem with write speed.

As for the documentation, not much has really changed from when you
brought this up on the docs list.  The amount of WAL files that can be
created by a "short-term peak" is unlimited, which is why there's no
better limit listed than that.  Some of the underlying things that make
the problem worse are operating system level issues, not ones in the
database itself; the PostgreSQL documentation doesn't try to wander too
far into that level.  There are also a large number of things you can do
at the application level that will generate a lot of WAL activity.  It
would be impractical to list all of them in the checkpoint documentation
though.

On reviewing this section of the docs again, one thing that we could do
is make the "WAL Configuration" section talk more about log_checkpoints
and interpreting its output.  Right now there's no mention of that
parameter in the section that talks about parameters to configure; there
really should be.

--
Greg Smith   2ndQuadrant US    greg@2ndQuadrant.com   Baltimore, MD
PostgreSQL Training, Services, and 24x7 Support  www.2ndQuadrant.us