Background

By default Serving sets only the cpu requests for queue proxy, which is set to 25m.

There are options to change that for all services by setting/enabling in config-deployment:

1. Sets the queue proxy's memory request.
2. If omitted, no value is specified and the system default is used.
   queue-sidecar-memory-request: "400Mi"

1. Sets the queue proxy's memory limit.
2. If omitted, no value is specified and the system default is used.
   queue-sidecar-memory-limit: "800Mi"
   Also you can set that per service via queue.sidecar.serving.knative.dev/resourcePercentage annotation but that 

does not work as expected: https://github.com/knative/serving/issues/7349#issuecomment-602125160.

The full algo of how resources are adjusted is here: https://github.com/openshift-knative/serving/blob/release-v1.7/pkg/reconciler/revision/resources/queue.go#L93

RHODS competitor Arrikto has already tested defaults for ML workloads:

https://docs.arrikto.com/user/serving/performance/configure.html#queue-proxy-sidecar

See long thread here: https://redhat-internal.slack.com/archives/CD87JDUB0/p1678287529899759

Not having defaults stops users from using resource quotas (customers have asked about it).

As a side note SM sets the following by default for injected side-cars:

                        "resources": {
                            "limits": {
                                "cpu": "2",
                                "memory": "1Gi"
                            },
                            "requests": {
                                "cpu": "10m",
                                "memory": "128Mi"
                            }
                        }, 

User Story

As a user I would like to be able to configure kservice resources including queue proxy

and understand how they relate.

AIs:

• restart discussion upstream if we need to set defaults for common use cases
• document upstream/downstream how resources are configured