wilfred-s commented on pull request #16: URL: https://github.com/apache/incubator-yunikorn-scheduler-interface/pull/16#issuecomment-629931534
I see a number of issues with this way of updating: 1) queue type created If queues are created them as an unmanaged queue the queue will be cleaned up if there are no apps submitted into the queue. If they are a managed queue then the queue config will be out of sync with the memory structure. A next reload of the queues config the update will mark the queue as removed and ready for cleanup. So we are back to the previous case. 2) queue and placement rule interaction With placement rules active the newly created or deleted queues could change the behaviour of the system without providing traceability. If a rule states "no create" and the queue is not in the system a placement rule would fail, or succeed, after an update. There is no way for an admin to trace this anywhere without going to logs as we do not keep the state anywhere. 3) there is no state kept Beside the above rule impact we have a generic state issue. This means that there is no way to recover an added queue. Queues for all other cases are either generated when the config is loaded or via placements rules. The newly added queues do not fit in either case which means they are not restored. 4) conflicting update For the update case if we set a new max on the queue and the queue is already over the max what is the expected behaviour. Are we cancelling the resources over the new limit or do we not take action. The action itself, updating the max resources for instance, might be idempotent the result might not be based on what we implement. Point 1 and 3 are show stoppers for me at this point. ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected]
