Re: WAL on zfs Settings

Laurent Asorne <[email protected]> Wed, 13 Feb 2019 00:03:11 +0100
Newsgroups gmane.comp.db.postgresql.novice
Message-ID <[email protected]>
How do I get out of this list?

Avec mes meilleures salutations / Bescht Gréiss / Mit freundlichen Grüssen /  Kind regards

Laurent Asorne
Agence Immobilière Abby TOUSSAINT 
32 81 73 41
[email protected]
www.confiance.lu

> On 12 Feb 2019, at 17:39, Stephen Froehlich <[email protected]> wrote:
> 
> Hi Adam,
> 
> Sorry this got sent to my Spam folder and I just happened to check it this morning.
> 
> Start with this slide deck ... https://www.slideshare.net/SeanChittenden/postgresql-zfs-best-practices
> 
> Sean advocates small block sizes and turning off all error checking in the database, but I don't expect you to re-init your database to make the latter happen.  For a typical database workload with lots of small writes, having "recordsize=16K" for the zfs volume in question might be quite helpful.  For my application, I have found that it doesn't help much, nor does it hurt other than it hinders compression some.
> 
> The big helper you can do now is that you can safely set "full_page_writes = off" & "wal_compression = off" for the WAL (usually about a 30-50% bump in write speed) as ZFS's copy-on-write scheme makes corruption impossible, and lz4 is better than what Postgres uses internally.
> 
> The other big helper is having a nice, fast ZIL SLOG.
> 
> --Stephen
> 
> 
> 
> -----Original Message-----
> From: Adam Jensen <[email protected]> 
> Sent: Friday, January 25, 2019 11:18 AM
> To: Stephen Froehlich <[email protected]>; [email protected]
> Subject: Re: WAL on zfs Settings
> 
> Hi, Stephen.
> 
> I also use PostgreSQL with a ZFS filesystem (FreeBSD). Will you describe your ZFS setup and the relevant parts of your postgresql.conf?
> 
> 
>> On 1/4/19 8:24 AM, Stephen Froehlich wrote:
>> I am rebuilding my database server, and one of the novice mistakes I 
>> made the first time around is not creating distinct zfs filesystems 
>> for different tablespaces and of course the WAL. The database is for 
>> analysis work, which often means large data load-ins.  (Backups are 
>> purely occasional pg_dumps.)
>> 
>>  
>> 
>> My question is what are the optimal zfs settings for the WAL 
>> filesystem (block size, who does compression, etc.)?  The underlying 
>> hardware will be 3 NVME SSDs in a raidz configuration.  (I do already 
>> know that I can turn off full_page_writes for WAL on zfs.)
>> 
>>  
>> 
>> Thanks,
>> 
>> Stephen
>> 
>>  
>> 
>> ----------------------------------------------------------------------
>> --
>> 
>> Stephen Froehlich
>> Sr. Strategist, *Cable*Labs^® ^
>> 
>> 
>> [email protected] <mailto:[email protected]>
>> 
>> Tel: +1 (303) 661-3708
>> 
>>  
>> 
> 
>