On Tue, Jan 21, 2025 at 7:12 AM Anthony D'Atri <[email protected]> wrote:
> > On Jan 21, 2025, at 7:59 AM, Kasper Rasmussen 
> > <[email protected]> wrote:
> >
> > 1 - Why do this result in such a high - objects degraded - percentage?
>
> I suspect that’s a function of the new topology having changed the mappings 
> of multiple OSDs for given PGs.  It’s subtle, but when you move hosts into 
> rack CRUSH buckets, that’s a different set of inputs into the CRUSH hash 
> function, so the mappings that come out are different, even though you 
> haven’t changed the rules and would think that hosts are hosts.

Also, in the process of moving the hosts one by one, each step creates
a new topology which can change the ordering of hosts, incrementally
putting things out of whack.

> > 2 - Why do PGs get undersized?
>
> That often means that CRUSH can’t find a complete set of placements.  In your 
> situation maybe those would resolve themselves when you unleash the recovery 
> hounds.

We started noticing this kind of issue around pacific, but haven't
fully tracked down what broke yet.
See https://tracker.ceph.com/issues/56046 for similar.

Undersized or degraded should only happen -- by design -- if objects
were modified while the PG did not have 3 OSDs up and acting.
Kaspar: I assume the cluster was idle during your tests?
Also -- can you reproduce it without norecover/nobackfill set ?

Could you simplify your reproducer down to:

> HEALTH_OK
> ceph osd crush move ksr-ceph-osd1 rack=rack1
> ceph pg ls undersized / degraded # get a pgid of a degraded PG
> ceph pg $pgid query

Cheers, dan


-- 
Dan van der Ster
CTO @ CLYSO
Try our Ceph Analyzer -- https://analyzer.clyso.com/
https://clyso.com | [email protected]
_______________________________________________
ceph-users mailing list -- [email protected]
To unsubscribe send an email to [email protected]

Reply via email to