Eric Liang

@scv119 want to shepherd this one?

> In the current runtime env plugin mechanism design, a plugin must implement a whole runtime env type (e.g., pip, conda, etc). Cgroup is different here. Because it is more...

I'd like to see the discussion outcome then. If @edoakes and @architkulkarni think it's not a good fit for runtime env plugin, then I'm fine with that decision.

I made some edits to https://docs.google.com/document/d/1x1JAHg7c0ewcOYwhhclbuW0B0UC7l92WFkF4Su0T-dk/edit# that I think will enable cgroups to be supported (namely, adding priority, create_for_worker, delete_for_worker). Cgroups might be implementable out of the box as a...

+1 On Thu, Apr 14, 2022, 9:27 PM Hao Chen ***@***.***> wrote: > I think we are on the same page that resource control is better > implemented as a...

> IMO, # 1 is more important, if we don't support this, users will have to manually reload data from the checkpoint. Our idea to fix this issue is that,...

> I assume your purpose is to save object store memory by pinning less objects, right? But this seems to complicate the design. My preference is that checkpoints are only...

> But one thing I am confused by in the "multiple owner" design is how the usual protocols will change after recovery. Normally, the advantage of having a single owner...

I'm trying to think through the multiple owner collision issue. Is it actually complex? If a process arbitrarily picked say, the first owner that registered for an object, it seems...

Great analysis! My two takeaways: - Option 2 is perhaps much more complicated than may appear (as @stephanie-wang has mentioned earlier). - Option 3 might be doable with named actors....