[YUNIKORN-1889] Remove instanceType from scheduling cycle

• Type: Bug
• Status: Closed
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: 1.3.0
• Fix Version/s: 1.4.0
• Component/s: core - scheduler
• Labels: newbie pull-request-available
• Target Version: 1.4.0
• Assignee: PoAn Yang
• Reporter: Wilfred Spiegelenburg
• Created: 31/Jul/23 07:07
• Updated: 20/Nov/23 16:42
• Resolved: 07/Aug/23 06:16

Description

There is no need for the node instanceType in the scheduling cycle or in the preemptor at all.

The only case in which it is used in the preemptor and that is to create a reserved allocation. If, and when, a reserved allocation is created in the TryPreemption  we return a non nil allocation. That is the only time.
That non nil allocation in the end gets processed at the Partition level. That processing gets finalised in partition.allocate(). We have a hold of the node itself there, we should add the instance type at that point to the allocation. That same partition.allocate() is called for all post processing of all allocations. That means we can remove the instance type from all scheduling calls and set it in one location.

It will indirectly also allow us to fix a bug that occurs when allocating a reservation on a non-reserved node. That case changes the node that the allocation is linked to and could thus change the node type.

Issue Links

• is a parent of: YUNIKORN-1905 instance type cleanup
• is caused by: YUNIKORN-1385 To provide visibility to application's aggregated resource consumption
• links to: GitHub Pull Request #605