Aller au contenu principal

Kubernetes Generated-Site Memory Safety Implementation Plan

For Codex: Use superpowers:test-driven-development for each behavior change and superpowers:verification-before-completion before claiming completion.

Goal: Make future generated-site Deployments advertise realistic memory, roll a single replica without a temporary duplicate, probe a lightweight health endpoint, and spread replicas across nodes.

Scope: Renderer and Kubernetes plugin settings only. Existing Work rows, generated repositories, live Deployments, routes, clusters, and node sizing are explicitly out of scope for this code change.

Design: The renderer defaults memory requests to a 512Mi admission floor, the smallest request above the measured roughly 468–499 MiB low-water working sets. This floor is not a sufficiency claim for sites measured in the 0.4–0.8 GiB range; those need higher measured per-Work requests. Platform-managed sources normalize legacy explicit requests below 512Mi during deploy only after validating the effective request/limit pair; custom clusters retain their values and the full Kubernetes quantity syntax. A settings hook rejects malformed managed quantities, request-over-limit, and new managed request/limit values below the floor. Single-replica Deployments use a surge-free 0/1 rollout, explicitly accepting a brief outage while the replacement becomes Ready; startup keeps / to warm the catalogue, readiness/liveness use /api/health, and a soft hostname topology spread preference applies to the existing pod selector.

Task 1: Lock renderer behavior with failing tests​

Files:

  • Modify: packages/plugins/k8s/src/__tests__/manifest.renderer.spec.ts
  1. Add tests for the 512Mi default and explicit override.
  2. Add tests for single-replica maxSurge: 0 / maxUnavailable: 1, its intentional brief-outage tradeoff, and multi-replica 1/0.
  3. Add tests that startup keeps / for catalogue warm-up while readiness and liveness use /api/health.
  4. Add a test for a soft hostname topology spread preference using the Deployment selector, avoiding tainted-control-plane hard-constraint deadlocks.
  5. Run pnpm test -- src/__tests__/manifest.renderer.spec.ts and confirm the new assertions fail for the expected missing behavior.

Task 2: Implement renderer safety defaults​

Files:

  • Modify: packages/plugins/k8s/src/manifest.renderer.ts
  1. Render 512Mi when no memory request is supplied.
  2. Render a surge-free strategy for one replica while retaining ordinary rolling availability for multiple replicas.
  3. Preserve the startup probe on / and point readiness/liveness at /api/health.
  4. Add a hostname topology spread preference with maxSkew: 1, ScheduleAnyway, and the existing selector.
  5. Re-run the renderer test and confirm it passes.

Task 3: Lock managed-cluster settings behavior with failing tests​

Files:

  • Modify: packages/plugins/k8s/src/__tests__/k8s.plugin.spec.ts
  1. Assert schema defaults without restricting Kubernetes' DecimalSI, BinarySI, or exponent syntax.
  2. Assert validation accepts 512Mi, 0.5Gi, and exponent forms on managed sources and preserves valid custom-cluster quantities.
  3. Assert validation rejects malformed managed quantities, request greater than limit, and managed request/limit values below 512Mi.
  4. Assert deploy normalizes a legacy managed 256Mi override to 512Mi, fails before any API write when the effective limit is too low, and preserves a custom-cluster 256Mi override.
  5. Run pnpm test -- src/__tests__/k8s.plugin.spec.ts and confirm the new assertions fail for the expected missing behavior.

Task 4: Implement settings validation and legacy normalization​

Files:

  • Modify: packages/plugins/k8s/src/k8s.plugin.ts
  1. Add exact Kubernetes DecimalSI, BinarySI, and DecimalExponent quantity parsing for managed-floor comparisons.
  2. Add schema defaults without a syntax-narrowing pattern.
  3. Add validateSettings() for managed quantity syntax, request/limit ordering, and managed-source minimum request/limit values while leaving custom-cluster validation to its API server.
  4. Normalize legacy sub-minimum managed requests at deploy render time without changing stored Work settings, but fail safely before any API write if the effective limit cannot cover the request.
  5. Re-run the focused tests and confirm they pass.

Task 5: Verify the complete plugin change​

  1. Run pnpm test -- src/__tests__/manifest.renderer.spec.ts src/__tests__/k8s.plugin.spec.ts.
  2. Run the package typecheck/build commands exposed by packages/plugins/k8s/package.json.
  3. Run git diff --check and inspect the rendered manifest diff for one- and two-replica inputs.
  4. Record rollback as reverting the renderer/plugin commit; no live resource changes are part of this code commit.