Which rollout strategy works on a single host with Docker Compose and a reverse proxy?

Open question

The rollout strategies of orchestrators assume multiple instances, a controller and controllable traffic splitting. A single host running Docker Compose offers little of this: docker compose up -d usually replaces the running container, and between the old one stopping and the new one becoming ready, the proxy responds with an error. The Compose specification does define update_config with order: start-first (new task first, brief overlap) and a rollback_config in its deploy section; whether and how a given Compose setup without an orchestrator implements them is part of the question. Reverse proxies such as Traefik offer weighted distribution (weighted round robin) between services, which could be used to approximate a canary. Specifically:

• Which approximation have small teams actually run for months: two named containers (blue/green) with a switched proxy rule, a second container with a healthcheck and start-first behavior, weighted distribution in the proxy, or simply a short outage with a maintenance page?
• What broke in the process: database migrations that both versions cannot tolerate; volumes that only one container may hold; long-lived connections (WebSockets, uploads) that drop during the switch; proxy configuration maintained both in a file and in container labels at the same time?
• How was the switch back tested: as a rehearsal before it was needed, or only when it was needed?
• At what number of services, deployments per week or users did teams find the move to an orchestrator worthwhile, and what held them back beforehand?
• Are there measurements of the actual outage duration with a plain up -d plus healthcheck compared with a blue-green approximation?

What a useful answer contains

Host and service size, proxy and Compose version, the exact approximation with a configuration excerpt, time in operation, failures that occurred and their cause, measured or estimated outage duration (stating what was measured), and whether the solution survived a change of staff. Proposals without operating experience are welcome if they are labeled as such.