Hi,
I came to the same conclusion after doing various tests with rooms and
failure domains. I agree with Maged and suggest to use size=4,
min_size=2 for replicated pools. It's more overhead but you can
survive the loss of one room and even one more OSD (of the affected
PG) without losing data. You'll also have the certainty that there are
always two replicas per room, no guessing or hoping which room is more
likely to fail.
If the overhead is too high could EC be an option for your setup?
Regards,
Eugen
Zitat von "ST Wong (ITSC)" <ST@xxxxxxxxxxxxxxxx>:
Hi all,
Tested 4 cases. Case 1-3 are as expected, while for case 4,
rebuild didn’t take place on surviving room as Gregory mentioned.
Repeated case 4 several times on both rooms got same result. We’re
running mimic 13.2.2.
E.g.
Room1
Host 1 osd: 2,5
Host 2 osd: 1,3
Room 2 <-- failed room
Host 3 osd: 0,4
Host 4 osd: 6,7
Before:
5.62 0 0 0 0 0
0 0 0 active+clean 2019-02-12 04:47:28.183375
0'0 3643:2299 [0,7,5] 0 [0,7,5] 0
0'0 2019-02-12 04:47:28.183218 0'0 2019-02-11
01:20:51.276922 0
After:
5.62 0 0 0 0 0
0 0 0 undersized+peered 2019-02-12
09:10:59.101096 0'0 3647:2284 [5] 5
[5] 5 0'0 2019-02-12 04:47:28.183218
0'0 2019-02-11 01:20:51.276922 0
Fyi. Sorry for the belated report.
Thanks a lot.
/st
From: Gregory Farnum <gfarnum@xxxxxxxxxx>
Sent: Monday, November 26, 2018 9:27 PM
To: ST Wong (ITSC) <ST@xxxxxxxxxxxxxxxx>
Cc: ceph-users@xxxxxxxxxxxxxx
Subject: Re: will crush rule be used during object
relocation in OSD failure ?
On Fri, Nov 23, 2018 at 11:01 AM ST Wong (ITSC)
<ST@xxxxxxxxxxxxxxxx<mailto:ST@xxxxxxxxxxxxxxxx>> wrote:
Hi all,
We've 8 osd hosts, 4 in room 1 and 4 in room2.
A pool with size = 3 using following crush map is created, to cater
for room failure.
rule multiroom {
id 0
type replicated
min_size 2
max_size 4
step take default
step choose firstn 2 type room
step chooseleaf firstn 2 type host
step emit
}
We're expecting:
1.for each object, there are always 2 replicas in one room and 1
replica in other room making size=3. But we can't control which
room has 1 or 2 replicas.
Right.
2.in<http://2.in> case an osd host fails, ceph will assign remaining
osds to the same PG to hold replicas on the failed osd host.
Selection is based on crush rule of the pool, thus maintaining the
same failure domain - won't make all replicas in the same room.
Yes, if a host fails the copies it held will be replaced by new
copies in the same room.
3.in<http://3.in> case of entire room with 1 replica fails, the pool
will remain degraded but won't do any replica relocation.
Right.
4. in case of entire room with 2 replicas fails, ceph will make use
of osds in the surviving room and making 2 replicas. Pool will not
be writeable before all objects are made 2 copies (unless we make
pool size=4?). Then when recovery is complete, pool will remain in
degraded state until the failed room recover.
Hmm, I'm actually not sure if this will work out — because CRUSH is
hierarchical, it will keep trying to select hosts from the dead room
and will fill out the location vector's first two spots with -1. It
could be that Ceph will skip all those "nonexistent" entries and
just pick the two copies from slots 3 and 4, but it might not. You
should test this carefully and report back!
-Greg
Is our understanding correct? Thanks a lot.
Will do some simulation later to verify.
Regards,
/stwong
_______________________________________________
ceph-users mailing list
ceph-users@xxxxxxxxxxxxxx<mailto:ceph-users@xxxxxxxxxxxxxx>
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
_______________________________________________
ceph-users mailing list
ceph-users@xxxxxxxxxxxxxx
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
