Closed grondo closed 1 year ago
Probably just trying to keep the number of allowed cases to a comprehensible minimum, since the updates aren't signed.
Seems like it would be ok to update jobspec resources
since it's just an input to the scheduler, and it's the output of the scheduler (under instance owner control) that determines what happens at run time.
(Sorry now that seems like "duh..." but I needed to talk that through :-)
I can propose a change.
Edit: I think the shell and the IMP only access the original signed jobspec anyway?
In flux-framework/flux-coral2#92 the jobspec of a job with a DWS spec is modified in place by the dws jobtap plugin.
It would be better to use an RFC 21
jobspec-update
event to make this modification, but currently RFC 21 forbids updating any keys that do not start withattributes.
.I'm not sure the rationale behind limiting
jobspec-update
events to only theattributes
section of jobspec. Perhaps we can expand this to allow updates toresources.
as well since there is this use case necessary in flux-coral2?