The scaling fix: replicas that actually run — no more silent collapse to one container, no more 502 backends.
Fixed
- Scaling a docker service past one replica no longer produces bad gateways. Two stacked causes, both fixed: replica clones inherited the primary's
-phost port publishing, so on published-port services every replica died on "port is already allocated" and scaling silently collapsed to one container; and the Traefik config rendered the DESIRED replica count (services.replicas), so between saving a higher count and the next deploy — or whenever a replica failed to start — the proxy listed backends that did not exist and round-robined a share of every request into 502s. Replicas no longer inherit the host port (public traffic reaches them over the service bridge via Traefik), and the proxy now renders the replica count the last deploy actually achieved (newservices.runtime_replicas, migration 0063, written by the deploy pipeline). After upgrading, redeploy replica services once to record their running count.