Deploy n8n Queue Mode | Worker, Redis and Postgres, All Pinned

n8n in queue mode

n8n split across a main instance and a worker, with Redis as the queue and Postgres for state. Everything pinned.

What this fixes

The existing "n8n with workers" template has four problems, and they compound:

• No image tags at all. n8nio/n8n and bitnami/redis are written without a version, which means :latest. Two deploys a month apart are not the same software, and a redeploy can change n8n underneath a database it has migrated.
• Bitnami closed its public tag catalogue. That Redis image is no longer a dependable pull, and when it fails it does so quietly.
• Redis and Postgres are exposed through public TCP proxies. A queue backend and a database do not need to be reachable from the internet; here they talk over the private network only.
• Postgres is on :latest as well.

Roughly one deployment in six fails.

What this runs

n8nio/n8n:2.32.5 for both the main instance and the worker, redis:8.6.5-alpine, and postgres-ssl:18 — every one pinned.

Two things that took finding

• The n8n image is Alpine, and the private network is IPv6. Without ENABLE_ALPINE_PRIVATE_NETWORKING the app resolves nothing and exits with "Unable to connect to Redis" — with no hint that the cause is DNS.
• Redis has to bind both stacks. --bind :: alone is not enough, and leaving the default binds loopback, which is reachable from nowhere.

The worker's health check also runs on 5680, not 5679: n8n's own task broker already holds 5679 inside the worker container, and the collision kills it on startup.

Verified

Not by a status page. An account was created, a webhook workflow activated, the production URL called, and the execution came back success with mode webhook — which only happens if the main instance queued it to Redis and the worker picked it up.

Scaling

The worker runs ten executions at a time (--concurrency=10). For more throughput, raise replicas on the worker service; the main instance stays at one.

Binary data stays in the database, because a filesystem volume is not shared between main and worker. For large files, point n8n at S3 instead.