2017-06-29 5:14 GMT+00:00 Tim Serong <tserong@xxxxxxxx>: > On 06/29/2017 05:04 AM, Sage Weil wrote: >> There is some goofy behavior with the systemd unit files that tries to >> make a new ceph-mgr daemon automagically appear next to the mons during a >> cluster upgrade: >> >> - The ceph-mon@.service triggers ceph-mgr@ to start >> - The ceph-mgr@.service tries to create a ceph-mgr key using whatever >> mon key it can find >> >> This was intended to ease the upgrade path by auto-creating ceph-mgr >> daemons along the way. > > It's also helpful for new installs; it means you don't actually have to > explicitly configure mgr daemons in order to get a running cluster. You > just end up with as many mgrs as mons automatically. > > That said, it is somewhat obscure, and also can make it slightly > irritating if you want to run mgrs on separate nodes (then you have to > explicitly deploy mgr, and also either uninstall the ceph-mgr package > from the mon nodes, or at least `systemctl mask ceph-mgr@$(hostname)` on > those nodes. IMO this is a good point, please remove the interdependency between mon and mgr packages. >> Do we really want this? I was reminded this behavior is here because of >> #19994, which points out that this isn't implemented for upstart (still >> used by trusty 14.04). >> >> The jewel -> luminous upgrade process is currently documented here >> >> http://docs.ceph.com/docs/master/release-notes/#upgrade-from-jewel-or-kraken >> >> and basically has a step that says 'verify mgr daemons are running and, if >> not, create some'. Is that sufficient? > > Probably. Especially given that now ceph-mgr isn't optional (you'll end > up with "HEATH_ERR; no active mgr"), there's incentive for everyone to > configure some mgr daemons ;-) > > More seriously: the systemd magic was able to ensure mgr was running > with Kraken, prior to the various deployment tools growing support for > deploying mgr, and (possibly?) prior to mgr appearing in the docs. > > That magic still works fine and I personally don't mind keeping it, but > if it's going to result in inconsistent automatic behaviour across > distros, that's probably bad. Maybe then for Luminous it's best now to > drop the magic and require everyone to explicitly configure mgr, as for > all other daemons? +1 to removing magic, it is making things unnecessarily complicated for people who want to deploy their cluster with some kind of automation tool. I was happy that the ceph-create-keys service has finally been dropped, please don't let it return through the backdoor. -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
