Deploy and Host SQLMesh transformations on Railway

Reviewed SQLMesh plans with private state and warehouse Postgres.

Template release v1.0.4; see VERSION. This release must be qualified independently. A source release does not prove marketplace publication; the maintainer's marketplace registry records final status and deploy links.

About Hosting SQLMesh transformations

SQLMesh incrementally transforms SQL data with tested plans and blocking audits. This evaluation stack deploys three private services: an exiting hourly runner and independent PostgreSQL state and warehouse databases. Each database has its own 5000 MB durable volume. Models, tests and audits are baked into the image.

Why Deploy SQLMesh transformations on Railway

Keep SQLMesh snapshots and intervals independent from the warehouse, replace ephemeral runners without losing progress, and review a bounded plan before promotion. Cron 0 * * * * (UTC) uses Start Command ./start.sh and restart NEVER; jobs execute sqlmesh run prod and exit, never applying model changes. There is no continuously running runner, HTTP healthcheck, public UI or database proxy. The fresh default intentionally refuses uninitialized prod until an explicit reviewed plan is applied inside the private network. README.md documents bounded maintenance access and restoration of the exact cron settings. Build-only SUCCESS and database readiness are not completed native SQL jobs.

Common Use Cases

Evaluate incremental SQL transformations with deterministic order/revenue fixtures.
Practice tested backfills and deliberately blocking data-quality audits.
Learn durable state, disposable runners and coordinated two-database recovery.

Dependencies for SQLMesh transformations

Deployment Dependencies

Locked SQLMesh/Python dependencies and runtime pins in Dockerfile and uv.lock.
Two digest-pinned PostgreSQL services with independent secrets and volumes; compose.yaml and .railway/railway.ts define the exact images.
Public standalone tech-progress/sqlmesh-transformations, main / release-v1, root /; verify the release's immutable tag and source authorization.
Operator access for reviewed plan initialization; a separate raw-data producer for real workloads.

This is not a high-availability or production-scale warehouse. Historical v1.0.3 private operator/application, blocking audit, replacement and paired restore passed. Its default job was manually restarted; historical hourly scheduler firing was not observed. Every release requires current native job and scheduler evidence under PUBLISHING.md. The MIT recipe license covers newly authored source only; upstream/runtime duties and notices remain separate. Standard deletion plus verified zero compute and disclosed retention is the cleanup boundary, without physical-erasure or billing-zero claims.