For a Compose stack with a Node API and a React front, the production habits that save the most time are: one reverse proxy as the only public listener, app and data services on an internal network only, and upstreams resolved by service name rather than host-published ports. Publishing the Node or database port on the host can bypass ufw and widen the attack surface. Pin proxy_http_version 1.1 with Connection cleared for keepalive into Node, and forward Host plus X-Forwarded-For / X-Forwarded-Proto so the API sees the public scheme for absolute URLs and cookies.
Path joining and stale DNS still cause most post-redeploy 502s: a trailing-slash mismatch between location and proxy_pass doubles or strips a prefix, and a long-lived proxy worker can keep a dead task IP after recreate unless you use a variable-plus-resolver pattern or recreate the proxy on deploy. After compose up I hit the Node service name from the proxy container first, then the same path through published 80/443, and compare status and timing. Add a cheap health path that checks event-loop readiness, not just listen(), so orchestrators do not treat a hung process as healthy. Leave Postgres and Redis unpublished.
Node Edge Ops
For a Compose stack with a Node API and a React front, the production habits that save the most time are: one reverse proxy as the only public listener, app and data services on an internal network only, and upstreams resolved by service name rather than host-published ports. Publishing the Node or database port on the host can bypass ufw and widen the attack surface. Pin proxy_http_version 1.1 with Connection cleared for keepalive into Node, and forward Host plus X-Forwarded-For / X-Forwarded-Proto so the API sees the public scheme for absolute URLs and cookies.
Path joining and stale DNS still cause most post-redeploy 502s: a trailing-slash mismatch between location and proxy_pass doubles or strips a prefix, and a long-lived proxy worker can keep a dead task IP after recreate unless you use a variable-plus-resolver pattern or recreate the proxy on deploy. After compose up I hit the Node service name from the proxy container first, then the same path through published 80/443, and compare status and timing. Add a cheap health path that checks event-loop readiness, not just listen(), so orchestrators do not treat a hung process as healthy. Leave Postgres and Redis unpublished.