Docker Compose in production: the honest checklist

Published October 1, 2026

⏱️ 2 min read

Compose is not the problem

A well-configured docker-compose.yml runs small production workloads fine. What fails is the default config: no restart policy, unbounded logs, no probes. Run this checklist and Compose stops being your weak link.

The nine checks

1. restart: unless-stopped on every service. Without it, a host reboot leaves you dark.
2. Log caps:

   logging:
     driver: json-file
     options: {max-size: "10m", max-file: "3"}

   Unbounded json-files are the #1 disk-full cause we see.
3. Healthchecks: real healthchecks plus depends_on: condition: service_healthy, so the app does not boot before the DB accepts connections.
4. Memory limits per service (mem_limit) — the OOM killer otherwise shoots the whole host.
5. Pinned image tags, never latest. Reproducibility beats convenience.
6. Secrets out of the file: env_file gitignored, or Docker secrets.
7. Pull policy: pull_policy: always on deploy nodes so rollback-by-tag actually works.
8. Networks: internal network for DB traffic; only the proxy publishes ports.
9. A real reverse proxy (Caddy/Nginx) with auto-TLS in front, not raw app ports.

When to graduate

Compose stops being honest past one host, or when you need rolling deploys. Then move to Swarm-mode Compose files or Kubernetes — same YAML instincts, better control plane. Lab both cheaply on hourly cloud boxes: Vultr or DigitalOcean. (Partner links.)