Re: Best filesystem for a high load db

On 11/26/2014 4:16 AM, Maila Fatticcioni wrote:
> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
>
> On 11/25/2014 05:54 PM, Bill Moran wrote:
>> On Tue, 25 Nov 2014 17:27:18 +0100 Christoph Berg <cb@df7cb.de>
>> wrote:
>>
>>> Re: Bill Moran 2014-11-25
>>> <20141125111630.d05d58a9eb083c7cf80ed9f8@potentialtech.com>
>>>> Anything with a journal is a performance problem. PostgreSQL
>>>> effectivly does its own journalling with the WAL logs. That's
>>>> not to say that there's no value to crash recovery to having a
>>>> journalling filesystem, but it's just to say that our
>>>> experience showed journaling filesystems to be slower. That
>>>> rules out ext4, unless you disable the journal. I seem to
>>>> remember ext4 with journalling disabled being one of the faster
>>>> filesystems, but I could be remembering wrong.
>>>
>>> If you are using a non-journalling FS, you'll be waiting for a
>>> full fsck after a system crash. Not sure that's an improvement.
>>
>> It's an improvement if: a) You're investing in high-quality
>> hardware, so the chance of a system crash is very low. b) The
>> database is replicated, so your plan in the event of a primary
>> crash is to fail over to the backup anyway.
>>
>> If both of those are in place (as they were at my previous job)
>> then the time it takes to fsck isn't an issue, and taking action
>> that causes the database to run faster when nothing is wrong can be
>> considered.
>>
>> Obviously, the OP needs to assess the specific needs of the product
>> in question. Your point is very valid, and I'm glad you brought it
>> up (as a lot of people forget about it) but sometimes it's not the
>> most important factor.
>>
>
> Thank you a lot to have shared with me your experiences.
> Indeed we will have two servers in cluster with high quality hardware
> so a fsck restore shouldn't be a big problem.
> I will analize the xfs option as well and then I will decide.
>
> Thank you again,
> Maila Fatticcioni
>

Also, if you do some timings, please share it with us, it'd be nice to
have some more data points.

-Andy