[Serving Catalog] Add HPA configurations

[raywainman]

One important (and non-trivial) aspect of running model servers today is to ensure they are able to scale horizontally in response to load. Today, traditional CPU/Memory-based autoscaling are not sufficient for this type of workload (see this doc).

In order to help Kubernetes users bootstrap their use of these workloads and be able to operate them at scale, we would like to add HPA configurations in the serving catalog such that all the heavy lifting around providing metrics, adapters and configurations is all working out of the box.

Proposal:

• Enrich current manifests in the serving catalog repository with HPA configurations.
• This will be composed of several pieces:
  • Local Prometheus that is scraping metrics from the model servers directly. Or this could be a Cloud vendor-specific product that provides the same functionality (for example, GKE could use Google Managed Prometheus).
  • Prometheus Adapter making the workload metrics collected by the Prometheus implementation available to HPA.
  • HPA configuration using one of the model server's metrics (eg. queue length).
• Leverage Kustomize's patching functionality to allow the user to choose what "flavour" of HPA they would like, to start we could provide two different types:
  • Latency target (corresponding to A1).
  • Optimal throughput (corresponding to A2).

References:

• KV-Aware LLM Autoscaling and Evaluation Metrics
• Kubernetes LLM Auto Scaling Examples

[raywainman]

/assign raywainman

[jjk-g]

From an implementation perspective, hpa configs as kustomize components allows re-use across layers.

Similar to how accelerator configs are used

[k8s-triage-robot]

The Kubernetes project currently lacks enough contributors to adequately respond to all issues.

This bot triages un-triaged issues according to the following rules:

• After 90d of inactivity, lifecycle/stale is applied
• After 30d of inactivity since lifecycle/stale was applied, lifecycle/rotten is applied
• After 30d of inactivity since lifecycle/rotten was applied, the issue is closed

You can:

• Mark this issue as fresh with /remove-lifecycle stale
• Close this issue with /close
• Offer to help out with Issue Triage

Please send feedback to sig-contributor-experience at kubernetes/community.

/lifecycle stale