fs: Mantle: A Programmable Metadata Load Balancer

[michaelsevilla]

• Design/Motivation presented at August CDM
• Testing: mds: add tests for mantle (programmable balancer) ceph-qa-suite#1145
• Relies on the Lua source used in Lua object class support #7338 (that PR is also pending discussion on ceph-devel)

[jcsp]

I'm not sure we want this: iirc the --debug-mds flag isn't a prefixish thing, so this change will silence all the balancer and migrator logs for people who currently set --debug-mds

[jcsp]

This looks to be exactly what was discussed at the last CDM, so all of my comments are just detail things. At a high level: 👍

Does #7338 need poking along somehow?

[michaelsevilla]

I already poked #7338 and @noahdesu is going to take care of it.

Dumb question: is it better to rebase (keep history clean but lose your comments) or just add more commits?

[jcsp]

You can either add more commits (and rebase later before merge) or just do a straight history-squashing rebase.

Usually we leave comments on the "files changed" view (not individual commits) so that they survive a rebase, but I mistakenly put some on commits here -- my fault if they get lost in a rebase.

[gregsfortytwo]

Luckily Github finally fixed that nonsense, at least as long as you navigate from the PR view into the commit (not sure if you just go straight to a commit). You can tell it's working because all the comments are on the diff, not on a commit. ;)

Personally I prefer commits prefixed with "SQUASH" that are intended to be rebase, as it makes life easier on the reviewer when there are follow-up comments, but that may just be me.

[gregsfortytwo]

I see we did get some updated commits at some point that at least invalidated most comments. @jcsp, want to check this again? Be nice to get it in for @batrick's scaling tests.

[jcsp]

This is good to go now from my point of view (apologies @michaelsevilla for delay). Please could you do a quick rebase as something else has touched CMakeLists.txt in the meantime