golang-production

Go in production

Predictable process under SIGTERM, observable and with limits. Language: golang. HTTP: golang-apis. General: production-readiness. Rust mirror: rust-axum-production.

Boot and config

• Env (or file) once in main. Typed struct; fail if missing DATABASE_URL / secret.
• Do not read env in the handler.
• log/slog JSON in prod; level by env (LOG_LEVEL).

APP_ENV=production
LISTEN_ADDR=0.0.0.0:8080

Process and shutdown

• http.Server own (ReadHeaderTimeout, IdleTimeout). Not just r.Run().
• SIGINT/SIGTERM: Shutdown(ctx) with timeout (e.g.: 30s). Close listeners, then DB pool.
• Background goroutines with errgroup + ctx canceled on shutdown (golang-goroutines).
• ListenAndServe in goroutine; the main waits for the signal.

Health

Endpoint	Meaning
/live	Process up
/ready	Critical dependencies (DB ping)

• Liveness should not kill the pod just because the DB fluctuated for 100ms — follow the orchestrator.
• No auth; no internals in the body.

Timeouts, limits, resilience

• Timeouts in every HTTP/DB client (context.WithTimeout).
• Body limit in Gin (MaxBytesReader / middleware).
• Retries only on transitive errors idempotent, backoff + jitter.
• Pool: SetMaxOpenConns aligned with instances × DB max_connections.
• http.Server.ErrorLog / slog; not log.Fatal in the handler.

Observability

• slog with request_id, method, path, status, latency. No PII/secrets.
• Metrics (Prometheus) if the team already uses it: latency, 5xx, in-flight.
• OpenTelemetry if the repo already has it — do not add a new stack in a feature.
• net/http/pprof on port internal or protected — never open on the internet.

HTTP security

• Headers: security project middleware (nosniff, frame, HSTS behind TLS).
• Explicit CORS. Trusted proxy (SetTrustedProxies) if there is a load balancer.
• Panic recover in Gin → 500 + log; do not bring down the process on the request.

Build and runtime

• CGO_ENABLED=0 when possible. Image: skill docker.
• GOMAXPROCS aligned with cgroup (automatic in recent Go / automaxprocs if already used).
• Graceful in Kubernetes: preStop + terminationGracePeriodSeconds > Shutdown timeout.

Anti-patterns

• gin.Default() with colored logger in prod without slog
• Without Shutdown
• /debug/pprof public
• Retry on non-idempotent POST
• Health that queries 12 dependencies and becomes an eternal false negative

Integration with other skills

• Gin: golang-apis
• SQL: golang-sqlc
• gRPC server: grpc (same shutdown/health)
• Docker: docker
• Go-live: production-readiness

Completion criteria

• http.Server + shutdown with timeout
• /live and /ready correct
• Timeouts in clients; body limits
• Structured logs without secrets
• pprof not public