[YUNIKORN-2791] [Umbrella] Tracking non-Yunikorn allocations in the core

• Type: New Feature
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 1.7.0
• Component/s: core - scheduler, scheduler-interface, shim - kubernetes
• Labels: release-notes
• Target Version: 1.7.0
• Assignee: Peter Bacsko
• Reporter: Peter Bacsko
• Created: 07/Aug/24 19:45
• Updated: 31/Oct/24 14:51
• Resolved: 29/Oct/24 12:54

Description

Currently, we don't know what non-YK pods are assigned to a particular node in the core. We only track the total amount of allocations as occupiedResources object inside the objects.Node type. If the tracking somehow becomes out of sync with the actual cluster state, it's very difficult to know what went wrong, because these allocations are not shown in the state dump.

In order to enhance supportability, we want to track all non-YK pods per node on the core side.

Issue Links

• is related to: YUNIKORN-2866 [UMBRELLA] Support InPlacePodVerticalScaling (phase 2)
• is related to: YUNIKORN-2365 [UMBRELLA] Support InPlacePodVerticalScaling (phase 1)
• relates to: YUNIKORN-2958 Display foreign allocations on the web UI
• relates to: YUNIKORN-1765 [Umbrella] Support K8s vertical pod auto scaling

Sub-Tasks

• Create design doc: Resolved
• [core] Add non-YuniKorn allocation tracking logic: Resolved
• [SI] Add new constants for non-YK allocation tracking: Resolved
• [shim] Update core and SI dependency after YUNIKORN-2832: Resolved
• Health checker reports foreign allocation as orphan: Resolved
• [shim] Add non-YuniKorn allocation tracking logic: Resolved
• [SI] Remove occupiedResources from NodeInfo: Resolved
• [shim] Remove occupiedResource handling logic: Resolved
• [core] Remove occupiedResource handling logic: Resolved
• Expose foreign allocations tags on the REST interface: Resolved
• Create foreign pod e2e tests: Resolved
• Add MockScheduler test which verifies foreign pod tracking: Resolved