Applying vertical-pod-autoscaler to the Tekton TaskRun pods.

[MIna-Maher]

Hey Team, I'm trying to use vertical-pod-autoscaler along with tekton TR through /scale subresource to the TaskRun CR but when applying the configured TR CR , the tekton validating/mutating webhook rejects the configure as replicas is not part of the taskRun CRD.

Could you please guide me how we could apply vertical-pod-autoscaler to the TaskRun pods ?
If this is not feasible at the moment, is there any other options We can leverage to add the vertical scaling option to the tekton pods ?

Thanks Team

[afrittoli]

Hello @MIna-Maher - thanks for creating this issue.

While Tekton relies on Pods for the implementation of its TaskRuns, the workload are more similar in nature to jobs, which are started, execute a chunk of work and end. Something like a vertical autoscaler works better for a stateless services, which can have multiple replicas serving requests in parallel and which can be killed and recreated without consequences.

[MIna-Maher]

Thnaks @afrittoli , For your response. As I understand using VPA along with Tekton is not valid option?
If yes, Is there any other options we could leverage to apply vertically scaling feature to Tekton pods?

[dippynark]

We are looking for the same functionality. VPA supports Jobs and CronJobs so this does feel like a reasonable ask: https://github.com/kubernetes/autoscaler/blob/8f0c84a96483eb787cfb76af228fc9b212da069f/vertical-pod-autoscaler/pkg/target/fetcher.go#L156-L159

However, the main technical blocker is that VPA looks for the topmost controller of a Pod:

• https://github.com/kubernetes/autoscaler/blob/8f0c84a96483eb787cfb76af228fc9b212da069f/vertical-pod-autoscaler/pkg/utils/vpa/api.go#L203-L204
• https://github.com/kubernetes/autoscaler/blob/8f0c84a96483eb787cfb76af228fc9b212da069f/vertical-pod-autoscaler/pkg/admission-controller/resource/vpa/matcher.go#L55

But for Pods created by Tekton that would be a PipelineRun, which typically have randomly generated names and so cannot be configured in the targetRef of a VPA.

Also, for tools that integrate with Tekton (e.g. Lighthouse) they typically have their own topmost controller (e.g. LighthouseJobs) so solving this issue within the Tekton project wouldn't help with these higher level tools.

So it seems that the design of VPA and Tekton are incompatible since the more generic label selector functionality of VPA has been removed for good reason: kubernetes/autoscaler#6493 (comment)