Kubernetes ingress and AWS service specifics
Ingress returns 404 for a path that should work
Check the host matches the request's Host header exactly; path and pathType (Prefix vs Exact — /api Exact won't match /api/orders); ingressClassName (wrong class → no controller serves it → default backend 404); the backend Service name and port; rewrite annotations (nginx.ingress.kubernetes.io/rewrite-target) changing the path the app receives; and whether another Ingress claims the same host/path.
S3 returns 403 for a static site
- The bucket policy / CloudFront Origin Access Control (OAC) doesn't allow the reader.
- The object doesn't exist and the requester lacks
s3:ListBucket— S3 returns 403 instead of 404 to avoid revealing whether objects exist. - Block Public Access enabled while trying to use a public website endpoint.
- KMS-encrypted objects without
kms:Decryptfor the reader.
CloudFront 403/404 on SPA routes
Single-page apps (React/Angular) serve routes like /dashboard from index.html. S3 has no /dashboard object → 403/404. Fix: CloudFront custom error responses mapping 403 and 404 to /index.html with response code 200 (or a CloudFront Function for URL rewrites).
API Gateway quirks
{"message":"Missing Authentication Token"}403 — often simply a wrong path or method, not auth.- 429 — stage/method throttling or usage plan limits.
- 502 — Lambda integration returned a malformed response (wrong proxy response shape) or errored.
- 504 — integration timeout (29 s default for REST APIs).
- 401 only through the gateway, not direct — authorizer configuration, or the
Authorizationheader isn't forwarded/mapped.
Health checks and redirects
If the app redirects HTTP to HTTPS (or / to /login), an ALB/ingress health check expecting 200 marks every target unhealthy → 503s for users. Point health checks at a dedicated path that returns 200 (e.g. /healthz), or configure the success codes (200-399).
Liveness vs readiness endpoints
- Liveness should return 200 as long as the process is alive and not deadlocked — don't fail it because a dependency is down (that just restarts healthy pods in a loop).
- Readiness should fail when the pod can't serve traffic (dependency it truly needs unavailable, warming up), removing it from load balancing without restarting it.