Enabling GWInference in helm chart breaks HPA

[ryan-dyer-sp]

Describe the bug
Enabling the gwAPIInferenceExtension in the helm chart breaks HPA. Likely need another schema object added to the nginxproxy.spec.kubernetes.deployment in order to customize the resourcing of the endpoint-picker container.

To Reproduce
Steps to reproduce the behavior:

1. in a helm deployment set .gwAPIInferenceExtension.enabled=true This in turn creates an additional container as part of the nginx gateways which does not have a resource request set. This prevents any existing HPA functionality from working as the pod no longer has all containers with defined requests.

Expected behavior
HPA continues to work when enabling inference.

Your environment

• Version of the NGINX Gateway Fabric - 2.2.0

[nginx-bot]

Hi @ryan-dyer-sp! Welcome to the project! 🎉

Thanks for opening this issue!
Be sure to check out our Contributing Guidelines and the Issue Lifecycle while you wait for someone on the team to take a look at this.

[ryan-dyer-sp]

@sjberman This PR appears to set the container resources to the same as what is defined by the nginx container. IDK anything about that container, but does its resource usage scale linearly with the nginx container? If not, then this could still pose a problem for HPA as if for example, a user has nginx set to an entire core, then a normal 70-80% desired utilization would need to be bumped down to 35-40 just to get it to scale (assuming endpoint picker does little).

Happy to open a ticket to have this enhanced. Or should the HPA switch to container based scaling instead of pod based?

[sjberman]

@ryan-dyer-sp The endpoint picker does do very little (essentially passes a request through). The intention of the PR was to just fix the HPA bug for our patch release coming soon. Based on your comment it probably does require an enhancement to allow for specific settings for that container. I didn't realize that the HPA splits the utilization evenly between the containers.

per-container metrics for the nginx container would be best if possible, and we could also open a ticket for configuring the resources for the endpoint picker container. Do you think the bug fix and per-container metrics are enough to accomplish what you need for now? Or is setting the resources on the other container required now?