fix(apps): add startupProbe to homepage, headlamp, actual-budget

k8s/bases/apps/actual-budget/helm-release.yaml

@@ -49,6 +49,23 @@ spec:
                       matchLabels:
                         app.kubernetes.io/name: actualbudget
                         app.kubernetes.io/instance: actual-budget
+              # Chart hardcodes startupProbe absence; values override the
+              # liveness/readiness blocks below but not startupProbe. Gate
+              # liveness/readiness on the container actually serving HTTP.
+              # initialDelaySeconds skips past the ~10s cold-start window so
+              # the first probe lands on a serving container — zero failure
+              # events during the merge-queue's 90s steady-state Warning
+              # check.
+              - op: add
+                path: /spec/template/spec/containers/0/startupProbe
+                value:
+                  httpGet:
+                    path: /
+                    port: http
+                  initialDelaySeconds: 20
+                  periodSeconds: 5
+                  timeoutSeconds: 3
+                  failureThreshold: 12  # 60s grace beyond initial delay
   # https://github.com/community-charts/helm-charts/blob/main/charts/actualbudget/values.yaml
   values:
     replicaCount: ${actual_budget_replicas:=1}

k8s/bases/apps/headlamp/helm-release.yaml

@@ -71,6 +71,26 @@ spec:
                 value:
                   name: tmp-dir
                   mountPath: /tmp
+              # Chart hardcodes liveness/readiness with K8s defaults
+              # (timeoutSeconds: 1, failureThreshold: 3, periodSeconds: 10).
+              # Headlamp is KEDA-scaled to 0 in prod; every cold start logs
+              # 1-3 Unhealthy probe warnings while the Go binary initialises.
+              # Add a startupProbe so liveness/readiness are gated until the
+              # main container is actually serving.
+              # initialDelaySeconds skips past the cold-start window so the
+              # first probe lands on a serving container — zero failure
+              # events during the merge-queue's 90s steady-state Warning
+              # check.
+              - op: add
+                path: /spec/template/spec/containers/0/startupProbe
+                value:
+                  httpGet:
+                    path: /
+                    port: http
+                  initialDelaySeconds: 20
+                  periodSeconds: 5
+                  timeoutSeconds: 3
+                  failureThreshold: 12  # 60s grace beyond initial delay
           - target:
               kind: Deployment
               name: headlamp

k8s/bases/apps/homepage/helm-release.yaml

@@ -52,6 +52,28 @@ spec:
                       matchLabels:
                         app.kubernetes.io/name: homepage
                         app.kubernetes.io/instance: homepage
+              # Chart hardcodes liveness/readiness with K8s defaults
+              # (timeoutSeconds: 1, failureThreshold: 3, periodSeconds: 10).
+              # Homepage takes ~13s to start serving on a fresh pod, so each
+              # rollout produces 3 Unhealthy probe warnings per pod and leaves
+              # only ~17s of headroom before the liveness restart fires. Add
+              # a startupProbe to gate liveness/readiness during initial boot.
+              #
+              # initialDelaySeconds is past the observed ~13s cold start so
+              # the first probe lands on a serving container — zero failure
+              # events during a normal rollout (the merge-queue's 90s
+              # steady-state Warning check would otherwise count probe
+              # failures fired in that window).
+              - op: add
+                path: /spec/template/spec/containers/0/startupProbe
+                value:
+                  httpGet:
+                    path: /
+                    port: http
+                  initialDelaySeconds: 20
+                  periodSeconds: 5
+                  timeoutSeconds: 3
+                  failureThreshold: 12  # 60s grace beyond initial delay
           # Flicker fix: with replicaCount=2 and the chart's default ClusterIP
           # service, the Kubernetes widget's poll for node/cluster CPU/memory
           # round-robins between two pods that each maintain their own