We have done lots of preemption hardening work in 1.6.0. Fair sharing was also fixed in 1.6.0 I do not think we looked at the impact on the examples. This behaviour change is most likely expected. Please file a jira to follow up on this as part of the phase work [1].
Wilfred [1] https://issues.apache.org/jira/browse/YUNIKORN-2872 On Mon, 16 Sept 2024 at 04:24, Yu-Lin Chen <[email protected]> wrote: > +0 (non-binding) > > I found that the simple preemption example have different result comparing > to the previous release. > > - > > https://github.com/apache/yunikorn-k8shim/tree/master/deployments/examples/preemption > > Previous, if I followed the steps, there should have 2 pods running from > "normal-priority-job-with-9999-priority" job before the cleanup step. > However, I only got 1 "normal-priority-job-with-9999-priority" pod in > 1.6.0-rc2. I have no clue why we have this behavior change. > [image: image.png] > > I'm sorry for not conducting a comprehensive investigation, but today is > the last day of voting. > Therefore, I would like to confirm if this is also an unexpected behavior > for others? > > > > Other items I've checked: > > - Verified signatures and checksums > - Built on Ubuntu 23.04(amd64) with go1.22.2 linux/amd64, deploy on > Kind 1.29.1 > - E2E tests passed in a new created Kind Cluster > - Check Restful APIs > - Validated changes in > - [YUNIKORN-2631] Support canonical labels for queue/applicationId in > Admission Controller > - [YUNIKORN-2504] Support canonical labels and align metadata > retrieving order in shim > - [YUNIKORN-2810] Throw a warning if a pod has inconsistent metadata > > > Yu-Lin Chen. > Best Regars. >
