[YUNIKORN-1242] gang: delay in scheduling after using all placeholders

• Type: Bug
• Status: Closed
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: 0.10
• Fix Version/s: 1.1.0
• Component/s: core - common
• Labels: pull-request-available
• Target Version: 1.1.0
• Assignee: Wilfred Spiegelenburg
• Reporter: Wilfred Spiegelenburg
• Created: 20/Jun/22 07:00
• Updated: 13/Sep/22 14:31
• Resolved: 28/Jun/22 03:07

Description

The check in the normal allocation Application.tryAllocate() does not take into account that there can be multiple types of gang definitions. Which  triggers weird behaviour if you either request more instances of a gang member than there are defined or there is a mismatch between the placeholder and the real allocation size.

Requirement is to have at least 2 gang types defined with placeholders and 1 of those has all its placeholder exhausted and the other is not.

This causes the pod for the gang type that has exhausted its placeholders to be delayed for scheduling until the other gang type's placeholder time out.

 

Issue Links

• Blocked: YUNIKORN-1252 expose timed out placeholders count in application REST info
• is depended upon by: YUNIKORN-1241 remove RequiredNode from PlaceholderData
• Testing discovered: YUNIKORN-1246 release reservation of required node in tryAllocate
• links to: GitHub Pull Request #417