[SPARK-35956] Support auto-assigning labels to less important pods (e.g. decommissioning pods)

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 3.2.0
• Fix Version/s: 3.3.0
• Component/s: Kubernetes
• Target Version/s: 3.3.0
• Assignee: Holden Karau
• Reporter: Holden Karau
• Created: 30/Jun/21 18:01
• Updated: 08/Mar/22 11:46
• Resolved: 23/Jul/21 23:26

Description

To allow for folks to use pod disruption budgets or replicasets we should indicate which pods Spark cares about "the least", those would be pods that are otherwise exiting soon.

 

With PDBs the user would create a PDB representing the label of decommissioning executors and this could have a higher number of unavailable than the PDB for the "regular" execs. For people using replicasets in 1.21 we could also set a label of "controller.kubernetes.io/pod-deletion-cost" (see https://github.com/kubernetes/kubernetes/pull/99163 ) to hint to the controller that a pod is less important to us.

Issue Links

• links to: [Github] Pull Request #33270 (holdenk)
• links to: [Github] Pull Request #33270 (holdenk)
• links to: [Github] Pull Request #35767 (dongjoon-hyun)