- Shipped
- November 6, 2026 at 5:30 PM UTC
- Author
- Kamo
- Commit
- 8c852fa
Node closed an idle keep-alive connection after 5 s; Traefik keeps its idle connections to a pod for 90 s and reuses them. A request sent as Node closed the socket was lost and Traefik answered its own 11-byte 'Bad Gateway' within milliseconds. Go replays such a request only for GET/HEAD, and the retry middleware stops once the request headers are written, so the 502s were all POST /api/generate-otk, all with RetryAttempts 0, before and after the retry route was added (2026-11-01 15:03-16:55 UTC, 2026-11-01 23:55, 2026-11-02 15:38). keepAliveTimeout is now 95 s (KEEP_ALIVE_TIMEOUT overrides), above Traefik's idle timeout, so the proxy always closes first.