PP-507: Add support for filtering nodes as per the job request


Mike your understanding is correct. The initial implementation will not utilize Python interpreter. Maybe I did not put it correctly, scheduler running the interpreter as non-root was one of the ways I thought we could deal with it, we could also impersonate as the submitter (in a hook) and then run the expression in the hook. But these are just ideas.


Hi Arun, maybe my concerns in my items 2 and 3 are based on a misunderstanding. Can you give an example illustrating the caveat you describe with the following statement?

If node_filter is used in conjunction with node_group_key, node_filter will take precedence over node_group_key set on server or queue.

I read that as “this feature is incompatible with placement sets”, but perhaps I am wrong :slight_smile:

P.S., if it ends up mattering, I agree that “same” is better than “any”.