A multi-service Cloudflare Tunnel stays boring when every public hostname maps to a stable Compose service name and internal port on the same user-defined network as cloudflared — never a host-published localhost that disappears the moment the connector itself runs in a container. Keep ingress rules hostname-shaped and ordered most-specific first, end with a catch-all that returns 404 so unknown names fail closed, and put Zero Trust Access in front of admin UIs before the app’s own login ever renders. Path-based splits under one hostname need the upstream to understand those prefixes; otherwise you get a silent 200 HTML shell that still cannot load its API.
After a config reload I check active connectors and tunnel availability first, curl each upstream by service name from inside the cloudflared network, then hit the same hostname through the public edge. A 502 with a healthy connector is almost always wrong port, wrong network membership, or a leftover ingress pointing at a renamed service. When you add a second connector for HA, every member must be able to reach every ingress backend it might receive — otherwise failover just moves the miss to a box that cannot see that service.