[Adding Sebastian, dev@ceph.io]

Some things to improve with the OSD create path!

On Mon, 20 Jan 2020, Yaarit Hatuka wrote:
> Here are a few Insights from this debugging process - I hope I got it right:
>
> 1. Adding the device with "/dev/disk/by-id/...." did not work for me, it
> failed in pybind/mgr/cephadm/module.py at:
> https://github.com/ceph/ceph/blob/master/src/pybind/mgr/cephadm/module.py#L1241
> "if len(list(set(devices) & set(osd['devices']))) == 0"
> because osd['devices'] has the devices listed as "/dev/sdX", but
> set(devices) has them by their dev-id.... (which is the syntax specified as
> the example in the docs, which I followed).
> It took me a couple of days to debug this :-)
>
> 2. I think that cephadm should be more verbose by default. When creating
> OSD it only writes "Created osd(s) on host 'mira027.front.sepia.ceph.com'"
> (even in case creation failed...). It will help if it outputs the different
> stages so that the user can see where it stopped in case of error.
>
> 3. ceph status shows that the OSD was added even if the orchestrator failed
> to add it (but it's marked down and out).

IIUC this is ceph-volumes failure path not cleaning up?  Is this the
failure you saw when you passed the /dev/disk/by-id device path?

It seems like ceph-volume completed successfully all this time, but since I always passed /dev/disk/by-id and not /dev/sdX to 'ceph orchestrator osd create', this intersection was always empty:
set(devices) & set(osd['devices']) [1]
The other part of the condition was also true, so the 'continue' happened all the time.
Therefore the orchestrator does not even try to:
self._create_daemon('osd',...) [2]

Not sure why the OSD count is incremented, though.

> 4. I couldn't find the logs that cephadm produces.
> I searched for them on both the source (mira010) and the target (mira027)
> machines in /var/log/ceph/<fsid>/* and couldn't find any print from either
> the cephadm mgr module nor the cephadm script. I also looked at /var/log/*.
> Where are they hiding?

The ceph-volume.log is the one to look at.

I looked at ceph-volume.log, but couldn't find any orchestrator / cephadm module log messages there... like this one:

> 5. After ceph-volume creates its LVs, the host's
> lvdisplay/vgdisplay/pvdisplay showed nothing. I had to run "pvscan --cache"
> on the host in order for those commands to output the current state. This
> may confuse the user.
>
> 6. I think it's also a good idea to have another cephadm feature "cephadm
> shell --host=<host>" to open a cephadm shell on a remote host. I wanted to
> run "ceph-volume lvm zap" on one of the remote hosts and to do that I sshed
> over, copied the cephadm script and ran "cephadm shell". it would be cool
> if we could do that from the original machine.

The cephadm script doesn't know how to ssh.  We could probably teach it,
though, for something like this... but it might be simpler for the
user to just 'scp cephadm $host:', as that's basically what cephadm would
do to "install" itself remotely?

sage