per queue workload time to live?

[richardhenwood]

Does queue offer a per queue workload time to live?

I am interested in this because I would like to implement a 'debug' queue. Here is an example of queue design that includes a debug/devel queue from NASA HECC. Here is a page focused on their devel queue.

A debug queue is somewhere to verify a job runs and small scale before queuing up for a full run. A key property of a debug queue is that it kills jobs after a short running time. This keeps resources free for the next debug job. Is it possible to apply TTL on jobs entering a queue with Kueue currently?

[kannon92]

The one enforcement we have for ttl on jobs is https://kueue.sigs.k8s.io/docs/concepts/workload/#maximum-execution-time.

This would be on the workload level though.

[richardhenwood]

Thanks for this. My reading of this is:

• maximumExecutionTimeSeconds is a TTL (in seconds) that is applied to a workload.
• it is the duration of time the workload is allowed to spend in Admitted.
• Admitted (in a clusterqueue) is the state the workload is in when Pods exist and are doing useful work.
• If I have a way to force a defined maximumExecutionTimeSeconds I can get the behavior I am looking for.

If you have any corrections to these thoughts, I would appreciate it.

[kannon92]

I think you could do this via a webhook or a Mutating admission policy. Set that field if someone specifies a certain local queue.

[richardhenwood]

For watchers, this is related to this discussion:
#6587