zfs scrub enable by default

[Karl Denninger]

On 8/4/2020 13:25, Warner Losh wrote:
> On Tue, Aug 4, 2020 at 10:55 AM Steve Wills <swills at freebsd.org> wrote:
>
>> Hi,
>>
>> On 8/4/20 11:54 AM, Matthew Ahrens wrote:
>>> This question was raised elsewhere, and I agree with this reply from
>>> George Wilson, my colleague and an expert in the i/o subsystems of ZFS
>>> as well as having lots of experience with customers:
>>>
>>>      Having scrubs enabled by default is a great idea but at Sun (and
>>>      Delphix too) we found that the impact was often too much for some
>>>      workloads/customers. This is the challenge we faced and why there
>>>      was never a policy to enable it everywhere. We did explore ideas to
>>>      make the impact less and to be able to always scrub. Some of those
>>>      ideas included periodic or continuous scrubs where the impact could
>>>      be reduced by only scrubbing portions of the pool at a time, at a
>>>      reduced i/o rate. At Delphix, we have investigated similar concepts
>>>      and one of our interns prototyped one of the ideas.Much has changed
>>>      since the early scrub days and revisiting some of the earlier ideas
>>>      and investigating new ones is probably a good topic for the
>>>      community. I do think that just enabling scrub by default without
>>>      further enhancements would still be too impactful for some customers
>>>      but the concept definitely has merit.
>>>
>> Thanks for that! Very informative. I thought the Fishworks storage
>> appliances had it on by default, but maybe I'm mistaken or maybe it
>> changed over time.
>>
>> I wonder what "some" means, that is, is it 80% of people? 50%? 20? And
>> what percent would mean "too many" to have it on and expect them to tune
>> it if needed. I suppose there's no way to know.
>>
>> There are definitely some interesting ideas for how to limit the impact
>> of scrub, but those would definitely have to be built and proven, of
>> course.
>>
> Yea, without numbers, it's unclear what to do with this advice since it
> says both "do it" and "don't do it" depending on how you read it. Like
> Steve said, if "some" is 80% it's a clear case for not enabling by default.
> If it is 5% or 10%, then the case is clear to enable it by default...
>
> I'm in the 'enable by default' camp *NOW* and keep a close eye out for the
> next six months. If there's only a couple of issues, leave it for the
> release. If there's all kinds of issues, then turn it back off.
>
> Better scrubbing is always possible, depending on the workload. We have a
> much better scrubber than before, and I think we should at least try it by
> default absent data indicating a big issue.
>
> Warner
> _______________________________________________

It hammers the living daylights out of performance on a RaidZx 
filesystem. On a mirror set not nearly so much.

If you have an actual workload on that RaidZ system while it's running 
you are very likely to see material and objectionable performance 
impact.  If you can start it when there's no material load and it 
finishes before there is some you won't notice it, but if not, well.....

-- 
Karl Denninger
karl at denninger.net <mailto:karl at denninger.net>
/The Market Ticker/
/[S/MIME encrypted email preferred]/
-------------- next part --------------
A non-text attachment was scrubbed...
Name: smime.p7s
Type: application/pkcs7-signature
Size: 4897 bytes
Desc: S/MIME Cryptographic Signature
URL: <http://lists.freebsd.org/pipermail/freebsd-fs/attachments/20200804/0df091fc/attachment.bin>