rail-knative/docs/substrate-runbook.md

Knative substrate runbook: declared CPU requests and HPA ceilings

The Knative Serving and Kourier v1.22.0 install on railiance01 is performed by railiance-cluster/install/knative/install.sh from checksum-pinned upstream assets. rail-knative declares the CPU requests that install must carry, in substrate/v1.22.0/cpu-requests.patch.yaml (strategic-merge patches), the HPA ceilings in substrate/v1.22.0/hpa-caps.patch.yaml, and substrate/v1.22.0/kustomization.yaml (overlay for the staged core.yaml and kourier.yaml; crds.yaml stays a separate first apply).

Deployment	Container	Upstream v1.22.0	2026-09-21	Declared (WP-0003)
knative-serving/activator	activator	300m	50m	10m
kourier-system/3scale-kourier-gateway	kourier-gateway	200m	50m	10m
knative-serving/net-kourier-controller	controller	200m	30m	10m
knative-serving/controller	controller	100m	30m	10m
knative-serving/webhook	webhook	100m	30m	10m
knative-serving/autoscaler	autoscaler	100m	30m	10m

HPA	Upstream min/max	Declared min/max (WP-0003)
knative-serving/activator	1/20	1/3
knative-serving/webhook	1/5	1/2
kourier-system/3scale-kourier-gateway	1/10	1/3

Only CPU requests and HPA maxReplicas differ from upstream. Memory requests, all limits, minReplicas and the HPA CPU targets are upstream's. The RAIL-KNATIVE-WP-0003 column (60m total, 7d p95 at most 7.2m per workload) is declared but not live until the railiance-cluster overlays carry it and the installer re-runs; see that workplan. The 2026-09-21 values were set live as ADMINISTER @ realm:kubernetes/railiance01, activation=APPROVED by the founder, because the node had 100% of its allocatable CPU requested and backups could not be scheduled. Record: the-custodian/docs/kubernetes-change-gate-decision.md.

The owner installer now renders these requests into the manifests before apply, using separate Serving and Kourier overlays. Its verifier checks all six values. Implementation and read-only live diff evidence are recorded in RAIL-BS-WP-0015, implemented by railiance-cluster commit 3a5432270e275e978d6c8a99529fa7f8be6eef57.

A plain re-apply of unpatched upstream manifests restores the upstream column. Every apply or upgrade must preserve these patches, and an upgrade to a new version needs a new substrate/<version>/ with container names re-checked against that release.

Lessons

A request cut cannot roll out on a node whose requests are exhausted. A rolling update creates the new pod before the old one stops. When no CPU is left to request, the new pod stays Pending, so the old pod never stops, even though the change would free capacity. On 2026-09-21 all six rollouts deadlocked until 50m was freed elsewhere. Free capacity before changing a Deployment's requests; then the first rollout frees more for the next.

The HPAs scale on percent of request. The HPAs on the activator, the Kourier gateway and the webhook target 100% CPU utilisation of the request. Lowering a request makes the same load read as a higher percentage, so the HPA scales out sooner (at about 50m or 30m of use, not 300m or 100m). At 2026-09-21 use they read about 2-10%. If the request is raised or lowered again, check the HPA targets in the same change. At 10m requests (WP-0003) the 2026-10-21 use reads 10-40%, and the declared maxReplicas bound any scale-out on the single node to 3+2+3 pods.

Read-only drift check

ssh railiance01 'kubectl get deploy -n knative-serving -o jsonpath="{range .items[*]}{.metadata.name} {.spec.template.spec.containers[0].resources.requests.cpu}{\"\n\"}{end}"; kubectl get deploy -n kourier-system -o jsonpath="{range .items[*]}{.metadata.name} {.spec.template.spec.containers[0].resources.requests.cpu}{\"\n\"}{end}"; kubectl get hpa -n knative-serving; kubectl get hpa -n kourier-system'