fix: chart 0.7.5 — raise KEDA interceptor cold-start timeouts (waitTimeout 180s, responseHeaderTimeout 30s)

Chart.yaml

@@ -23,14 +23,14 @@ description: |
       Git).
 
 type: application
-version: 0.7.7
+version: 0.7.5
-appVersion: "0.7.7"
+appVersion: "0.7.5"
 
 # All 6 subcharts now resolve from registry.odoosky.cloud (mirrored
 # 2026-05-08, KEDA stack added 2026-05-09). Mirror-first discipline
-# + China-region readiness: a Jetstack / Traefik / Longhorn /
+# for regional / air-gapped readiness: a Jetstack / Traefik /
-# external-secrets-io / KEDA outage no longer blocks new tenant
+# Longhorn / external-secrets-io / KEDA outage no longer blocks new
-# cluster bootstrap.
+# tenant cluster bootstrap.
 #
 # Original upstream sources (for re-sync if a chart bumps):
 #   cert-manager     → https://charts.jetstack.io

values.yaml

@@ -148,15 +148,6 @@ traefik:
   # every TCP-80 request bounce to TCP-443 with a 301 — applies
   # uniformly to all IngressRoutes on this cluster, no per-instance
   # Middleware or duplicate IngressRoute needed.
-  # KEDA HTTP add-on routing — Studio chart's IngressRoute lives in
-  # the per-instance tenants namespace but its backend Service lives
-  # in odoosky-system (where KEDA was installed by this chart in 0.7.4).
-  # Without allowCrossNamespace=true, Traefik silently returns 404 on
-  # the cross-ns reference. Enabling here unblocks every Studio across
-  # every tenant on every cluster — single platform-wide setting.
-  providers:
-    kubernetesCRD:
-      allowCrossNamespace: true
   ports:
     web:
       redirectTo:
@@ -335,35 +326,6 @@ external-secrets:
 # Subchart values pass through under the dep name (`keda:`) below.
 keda:
   enabled: true
-  # crds.install — explicit true. KEDA's chart default is true, but
-  # somewhere in the parent-subchart values merge our keda subchart
-  # was resolving crds.install=false on a fresh render — leaving CRDs
-  # absent. Without ScaledObject CRD installed, the operator pod
-  # crashloops at startup ("failed to wait for scaledobject caches
-  # to sync") and Argo's apply of the keda-add-ons-http interceptor's
-  # ScaledObject fails ("no matches for kind"). The whole platform
-  # sync stalls there, the wildcard cert is never issued, and Tower's
-  # UI flips to "Failed - TLS cert renewal failing -739746d left"
-  # (Tower computes notAfter - now() against Go's zero-time when the
-  # Secret is missing, hence the bizarre negative-day display).
-  # Repro: havari-server03 onboarding 2026-05-10. Manual unblock was
-  # `kubectl apply --server-side --force-conflicts -f keda-crds.yaml`.
-  # This explicit true codifies that step into the chart so any
-  # newly-onboarded server gets CRDs on first sync.
-  # KEDA puts CRDs in templates/crds/ (not chart/crds/), so they
-  # need to flow through helm template — which they do once this is
-  # set. ArgoCD's App-level ServerSideApply=true sync option SHOULD
-  # cover the 262 KiB annotation overflow these CRDs would hit with
-  # client-side apply (scaledjobs + scaledobjects each blow the limit
-  # because their schemas are huge), but ArgoCD's App-level setting
-  # falls back to client-side for resources that already exist with
-  # a different field-manager. Belt-and-suspenders: also annotate
-  # each CRD per-resource so the SSA path is taken regardless of
-  # who owned the field-manager first.
-  crds:
-    install: true
-    additionalAnnotations:
-      argocd.argoproj.io/sync-options: ServerSideApply=true
   # operator + adapter + webhook — keep CPU/RAM modest. KEDA polls
   # event sources every pollingInterval (default 30s); on a cluster
   # with no ScaledObjects it does no work.
@@ -409,29 +371,6 @@ kedaHttpAddon:
 # buffers each cold-start request until the target pod is Ready.
 # The scaler is the control loop watching HTTPScaledObject status.
 keda-add-ons-http:
-  # kube-rbac-proxy sidecar — upstream HTTP add-on 0.8.0 references
-  # gcr.io/kubebuilder/kube-rbac-proxy:v0.13.0 which was retired from
-  # gcr.io. Override to a current image mirrored to our registry
-  # (multi-arch preserved via crane copy from quay.io/brancz upstream).
-  # Without this override the controller-manager pod ImagePullBackOffs
-  # forever and HTTPScaledObjects never reconcile.
-  images:
-    kubeRbacProxy:
-      name: registry.odoosky.cloud/odoosky/docker-mirror/kube-rbac-proxy
-      tag: v0.18.0
-  # imagePullSecrets — required so the kube-rbac-proxy sidecar
-  # (mirrored at registry.odoosky.cloud) can pull. The operator
-  # container itself pulls from ghcr.io which needs no auth, but
-  # k8s applies imagePullSecrets per-pod (covers all containers).
-  operator:
-    imagePullSecrets:
-      - name: docker-mirror-pull
-  # Same pattern as keda.crds.install above — explicit true so the
-  # HTTPScaledObject CRD lands on every fresh server. Without it, the
-  # interceptor never gets its watch table synced (logs: "table has
-  # not synced") and Studio cold-starts hang at "Connecting…".
-  crds:
-    install: true
   interceptor:
     replicas:
       # Scale the interceptor itself with HPA on its own metrics —
@@ -441,6 +380,26 @@ keda-add-ons-http:
       # handles bursts above that.
       min: 1
       max: 3
+      # waitTimeout (→ KEDA_CONDITION_WAIT_TIMEOUT) — how long the
+      # interceptor holds a request waiting for the target Deployment
+      # to scale 0→1 and become Ready. Upstream default is 20s; an AI
+      # Studio cold-start is image pull (~33s for the ~900 MB
+      # opencode-odoo image) + app boot (~30-60s) ≈ 60-90s, so 20s
+      # 502s the operator's first click ("context deadline exceeded
+      # while waiting for workload reach > 0 replicas"). 180s lets the
+      # first click WAIT through the cold-start instead of erroring;
+      # a warm pod is unaffected (this only caps the 0→1 wait).
+      # NOTE: the studio-template-v3 chart's per-HSO `spec.timeouts`
+      # block does NOT work — HTTPScaledObject CRD 0.8.0 has no such
+      # field; interceptor timeouts are GLOBAL, set here.
+      waitTimeout: "180s"
+    # responseHeaderTimeout (→ KEDA_RESPONSE_HEADER_TIMEOUT) — how
+    # long the interceptor waits for response headers after forwarding
+    # to a warm backend. Upstream default 500ms is far too tight for a
+    # just-woken Studio's first route load (OpenHands boots its
+    # browser env lazily); 30s is generous headroom without masking a
+    # genuinely hung backend.
+    responseHeaderTimeout: "30s"
     resources:
       requests:
         cpu: 50m