Nope, umpap is currently impossible on this clusters 😬
due client lib (guys works on update now).
ID CLASS WEIGHT REWEIGHT SIZE RAW USE DATA OMAP META AVAIL
%USE VAR PGS STATUS TYPE NAME
-166 10.94385 - 11 TiB 382 GiB 317 GiB 64 KiB 66 GiB 11 TiB
3.42 1.00 - host meta115
768 nvme 0.91199 1.00000 932 GiB 36 GiB 30 GiB 8 KiB 6.0 GiB 896 GiB
3.85 1.13 1 up osd.768
769 nvme 0.91199 1.00000 932 GiB 22 GiB 18 GiB 4 KiB 4.0 GiB 909 GiB
2.41 0.71 0 up osd.769
770 nvme 0.91199 1.00000 932 GiB 38 GiB 31 GiB 8 KiB 6.3 GiB 894 GiB
4.04 1.18 2 up osd.770
771 nvme 0.91199 1.00000 932 GiB 22 GiB 18 GiB 0 B 3.9 GiB 910 GiB
2.33 0.68 0 up osd.771
772 nvme 0.91199 1.00000 932 GiB 37 GiB 30 GiB 4 KiB 6.1 GiB 895 GiB
3.93 1.15 2 up osd.772
773 nvme 0.91199 1.00000 932 GiB 34 GiB 28 GiB 4 KiB 6.0 GiB 898 GiB
3.65 1.07 1 up osd.773
774 nvme 0.91199 1.00000 932 GiB 32 GiB 26 GiB 8 KiB 5.4 GiB 900 GiB
3.43 1.00 1 up osd.774
775 nvme 0.91199 1.00000 932 GiB 36 GiB 30 GiB 4 KiB 6.1 GiB 895 GiB
3.91 1.14 2 up osd.775
776 nvme 0.91199 1.00000 932 GiB 36 GiB 30 GiB 4 KiB 6.4 GiB 895 GiB
3.90 1.14 1 up osd.776
777 nvme 0.91199 1.00000 932 GiB 36 GiB 30 GiB 8 KiB 6.1 GiB 895 GiB
3.89 1.14 2 up osd.777
778 nvme 0.91199 1.00000 932 GiB 32 GiB 27 GiB 8 KiB 5.5 GiB 899 GiB
3.48 1.02 1 up osd.778
779 nvme 0.91199 1.00000 932 GiB 21 GiB 17 GiB 4 KiB 3.7 GiB 911 GiB
2.23 0.65 0 up osd.779
TOTAL 11 TiB 382 GiB 317 GiB 65 KiB 66 GiB 11 TiB 3.42
MIN/MAX VAR: 0.65/1.18 STDDEV: 0.66
Second PG landed... don't see any huge spikes on ceph_osd_op_latency metric.
k
> On 21 Apr 2021, at 17:12, Dan van der Ster <[email protected]> wrote:
>
> Yes, with the fixes in 14.2.19 PG removal is really much much much
> better than before.
>
> But on some clusters (in particular with rocksdb on the hdd) there is
> still a rare osd flap at the end of the PG removal -- indicated by the
> logs I shared earlier.
> Our workaround to prevent that new flap is to increase
> osd_heartbeat_grace (e.g. to 45).
>
> With 3.5M objects in a PG, I suggest that you try moving one PG with
> upmap and watch how it goes (especially at the end).
_______________________________________________
ceph-users mailing list -- [email protected]
To unsubscribe send an email to [email protected]
