Airflow, Temporal or Windmill for background jobs
Short answer: pick by how you want to write jobs. Apache Airflow runs scheduled data pipelines written as Python DAGs. Temporal runs long-lived, retry-safe workflows written as code in an SDK, and you deploy your own workers. Windmill turns scripts into flows, schedules, webhooks and internal apps with a built-in editor. Each has a Tracklayer template for Railway, and all three need Postgres.
Deploy Windmill on RailwayWhen you need a workflow engine
A cron job is enough until jobs depend on each other, need retries, or must be visible to someone other than the author. A workflow engine adds a scheduler, run history, retries and a UI. If you build with AI tools, it is the part that keeps generated scripts running unattended instead of living on a laptop.
Compare the three templates
| Airflow 3 | Temporal | Windmill | |
|---|---|---|---|
| You write | Python DAGs in a git repo | Workflows and activities in Go, Java, TypeScript, Python, .NET, PHP or Ruby | Python, TypeScript, Go, Bash or SQL scripts in the web editor |
| Pinned version | Airflow 3.3.2 | Server 1.32.1, UI 2.55.0 | 1.830.0 |
| Services | API server, scheduler, DAG processor, triggerer, Celery worker, init job, Postgres, Redis, Bucket | Server, UI, login gateway, namespace job, Postgres | Proxy, server, 3 workers, native worker, editor services, bootstrap job, Postgres |
| Your own code runs | In the Airflow workers | In worker services you deploy yourself | In the Windmill workers |
| Login | Admin user, generated password | Basic-auth gateway in front of the UI | Admin user, password rotated on first boot |
| Licence | Apache-2.0 | MIT | AGPLv3 or Apache-2.0 per file, plus Community Edition terms |
| Estimated idle cost | about $26 per month | about $10 per month | about $17 per month |
Cost figures are estimates from each template’s own documentation, not measurements. There is no ranking page for workflow engines on this site yet.
Deploy Windmill, step by step
Windmill is the shortest path from a script to a running job, so the guide uses it as the main example.
- Deploy the template. It creates the proxy, server, a default worker pool, a native worker, the editor services and Postgres.
- Wait for the bootstrap job. Its logs show that the admin password was rotated. The job stopping afterwards is expected.
- Sign in. Open the Proxy service’s public URL as
admin@windmill.devwithWINDMILL_ADMIN_PASSWORDfrom the Admin Bootstrap service. - Replace the default admin. In Instance settings, confirm the Base URL, add your own user as superadmin, sign in with it, then change or disable the default account.
- Scale workers. Each default worker replica runs one job at a time. Change the replica count on the Windmill Worker service.
- Plan upgrades. All four Windmill services share one release tag. Back up the Postgres volume, set the same new tag everywhere, and test in a separate environment first, because migrations cannot be rolled back.
Deploy Airflow or Temporal
Airflow. Wait for airflow-init, then sign in at the airflow-apiserver domain. The two example DAGs start paused. To run your own, set AIRFLOW_DAGS_GIT_URL, AIRFLOW_DAGS_GIT_REF and AIRFLOW_DAGS_GIT_SUBDIR on airflow-apiserver. Railway volumes cannot be shared between services, which is why DAGs come from git. For a private repository, add AIRFLOW_CONN_GIT_DAGS with an access token to every Airflow service. Scale with airflow-worker replicas or AIRFLOW__CELERY__WORKER_CONCURRENCY.
Temporal. Sign in at the temporal-ui-gateway domain. Deploy your worker service in the same project and set TEMPORAL_ADDRESS=${{temporal.TEMPORAL_ADDRESS}} and TEMPORAL_NAMESPACE=default. The template creates the default namespace. The history shard count is 512 and cannot change after the first start.
Costs
Railway bills usage, so cost follows RAM and CPU. The templates estimate it at Railway’s RAM and CPU rates:
- Airflow: idle about $26 per month, a small data team about $60 to $75, and each busy extra worker replica about $15 to $30 more.
- Temporal: idle about $10 per month, about $25 to $40 at tens of workflows per second with 7-day retention. Your own worker services are extra.
- Windmill: idle about $17 per month, about $66 for a small team with steady jobs. The Windmill README says its figures are not measured on Railway.
All three list the Hobby plan as the minimum. Airflow’s own table says the Trial plan is not enough. LLM or API costs from your jobs are billed by those providers.
Limits and licence notes
- Windmill: Railway has no privileged containers, so nsjail sandboxing is unavailable. Treat everyone who can write scripts as trusted. Enterprise-only parts, Docker-in-Docker and SMTP email triggers are not included.
- Temporal: the frontend has no authentication by default, so keep it private. Upgrade one minor version at a time.
- Airflow: it is the heaviest of the three to run, with eight services and about 2.2 GB of idle RAM in the template’s own estimate.
- Licences: see the FAQ for the Windmill Community Edition terms.
These are community templates, not affiliated with the Apache Software Foundation, Temporal Technologies or Windmill Labs.
Other options
For a lighter, visual option, see the n8n queue-mode guide. For AI agent builders, see Langflow and Flowise. More guides are listed at /guides/.
FAQ
How do I log in to each engine after deploying?
Airflow: wait for airflow-init to finish, then sign in at the airflow-apiserver domain with AIRFLOW_ADMIN_USERNAME (default admin) and the generated AIRFLOW_ADMIN_PASSWORD. Temporal: open the temporal-ui-gateway domain and sign in with BASIC_AUTH_USERNAME (default admin) and the generated BASIC_AUTH_PASSWORD. Windmill: after the Admin Bootstrap job logs that the password was rotated, sign in at the Proxy URL as admin@windmill.dev with WINDMILL_ADMIN_PASSWORD from the Admin Bootstrap variables, then add your own superadmin user.
Are my jobs kept when a service redeploys?
Workflow state lives in Postgres in all three templates. Airflow also keeps task logs in a Railway Bucket and loads DAGs from a git repository, so the Airflow services hold no local state. Temporal keeps workflow history in Postgres, with a default retention of 7 days for the default namespace. Windmill keeps its job queue and history in Postgres on a volume; its workers cache dependencies on local disk, so the first run of a script with new packages after a redeploy downloads them again.
What licences apply?
Airflow is Apache-2.0. Temporal and its Web UI are MIT. Windmill's source is licensed per file under AGPLv3 or Apache-2.0, with proprietary Enterprise code. The Windmill Community Edition images used by the template carry Windmill Labs' own terms: free use and redistribution as is, while selling, reselling, offering as a managed service, modifying or wrapping them requires an agreement with Windmill Labs. The template deploys the images unmodified. This is not legal advice.
Can I expose Temporal to workers outside Railway?
Only with care. The Temporal frontend has no authentication by default. The template keeps it on the private network. If you add a TCP proxy on port 7233, the template documents doing so only together with mTLS, because anyone who finds the address could start, signal or terminate workflows.