/n8n-self-hosting

Queue mode

Executions are pulled off a Redis queue by a pool of worker processes, so work runs in parallel and scales horizontally. Template: assets/docker-compose.queue.yml.

Architecture

The docs mark SQLite as not recommended for queue mode; this skill treats Postgres as mandatory — don't deploy queue mode without it. init-data.sh creates the non-root DB user n8n connects as, separate from the Postgres superuser. Official guide: https://docs.n8n.io/deploy/host-n8n/configure-n8n/scaling/enable-queue-mode.

The settings that make it queue mode

Set on main and every worker (the x-n8n anchor applies them to both):

• EXECUTIONS_MODE=queue
• QUEUE_BULL_REDIS_HOST=redis, QUEUE_BULL_REDIS_PORT=6379 (Redis auth/TLS/cluster vars exist for external Redis; QUEUE_BULL_REDIS_USERNAME needs Redis ≥ 6 — full list: https://docs.n8n.io/deploy/host-n8n/configure-n8n/basic-configuration/use-environment-variables/queue-mode)
• QUEUE_HEALTH_CHECK_ACTIVE=true (workers expose /healthz for probes — off by default)
• DB_TYPE=postgresdb + the DB_POSTGRESDB_* connection vars
• N8N_ENCRYPTION_KEY — identical everywhere. Workers decrypt credentials to run nodes; a mismatched key means workers can't decrypt and executions fail. The anchor sets it once from .env; never override it per-service.
• OFFLOAD_MANUAL_EXECUTIONS_TO_WORKERS=true — even "Test workflow" runs go to workers.

The main additionally gets the public-URL vars (N8N_HOST, WEBHOOK_URL, N8N_EDITOR_BASE_URL, N8N_PROTOCOL=https, N8N_PROXY_HOPS=1, N8N_SECURE_COOKIE=true) — workers don't serve the UI so they don't need them.

Env parity: the rule for every var you add later

Those public-URL vars are the only legitimate difference between the main and a worker. Everything else is behavioural — which database, which queue, which encryption key, which optional modules, how binary data is stored — and a worker that disagrees with the main about behaviour is a worker that executes your workflows wrongly.

That is why the template declares the common environment once, in the x-n8n-env anchor, and has the main merge it (<<: *n8n-env) before adding its public-URL vars. Add new behavioural vars to the anchor, not to the main's own environment: block.

This failure mode is nastier than it sounds, because the main is what you look at. The editor UI, the REST API and every node's dropdown are served by the main, so a feature enabled only there looks configured — it's the execution, on a worker, that fails. And with OFFLOAD_MANUAL_EXECUTIONS_TO_WORKERS=true even a manual "Test workflow" run lands on a worker, so there is no local sanity check that would catch it.

Check parity whenever you change the environment:

diff <(docker compose exec -T n8n env | sort) \
     <(docker compose exec -T --index 1 n8n-worker env | sort)

Only the public-URL/proxy vars above (plus per-container noise like HOSTNAME) should differ. Anything else in that diff is a bug you just introduced.

Optional modules

Some n8n features ship as backend modules that are off unless you list them in N8N_ENABLED_MODULES (comma-separated). The Agents feature is the current example: N8N_ENABLED_MODULES=agents. The template carries it commented out in the x-n8n-env anchor — uncomment it there and it reaches main and workers together.

Enabling a module on the main only produces a characteristic failure: the module's REST routes answer (so the UI populates), but the workers never registered the module's database entities, so any node that touches it dies at runtime with a TypeORM error like EntityMetadataNotFoundError: No metadata for "Agent" was found. If you see an error of that shape after enabling a feature, check env parity first — it is almost always this.

On Agents specifically, be straight with the user. The docs say "Queue mode isn't supported for agents yet, and connecting channels (such as Telegram) can fail" and recommend regular mode. That sentence is broader than what the code actually gates: the leader/multi-main checks apply to agent schedules and chat channels, while the Message an Agent node just runs the agent inline in whichever process executes the workflow — so it does work on a worker once the module is enabled there. Treat that as unsupported territory rather than a supported configuration: it can break on any upgrade, and the agent knowledge base additionally needs the sandbox vars (N8N_AGENTS_AI_SANDBOX_*, DAYTONA_*) on the workers too. If the user's main use of n8n is agents, single/regular mode is the honest recommendation.

Instance-wide OAuth apps (credential overwrites)

If you register a shared OAuth app with CREDENTIALS_OVERWRITE_*, only the main serves the registration endpoint, but CREDENTIALS_OVERWRITE_PERSISTENCE=true belongs on the workers too: a save is broadcast over pubsub so running workers reload immediately, yet a worker that cold-starts without the flag never reads the stored row and silently loses the overwrite. That is the classic parity failure from the section above — fine in the editor, broken in execution. See CREDENTIAL_OVERWRITES.md.

Scaling the workers

• Each worker runs worker --concurrency=5 (5 simultaneous executions per worker; the flag's upstream default is 10 — the template pins 5 as a safer floor for modest boxes; raise for many light executions, lower for heavy ones).
• More throughput = more workers. The template sets deploy.replicas: 2, which Docker Compose v2 honors under docker compose up (here it is not Swarm-only). To change the count, either edit deploy.replicas and re-run docker compose up -d, or override at launch with docker compose up -d --scale n8n-worker=N — a --scale value supersedes replicas (passing both at once just prints a harmless conflict warning).
• Rough capacity ≈ replicas × concurrency simultaneous executions, bounded by CPU/RAM and DB_POSTGRESDB_POOL_SIZE (default 2 per process — raise it if many workers exhaust the pool).
• Optionally cap load with N8N_CONCURRENCY_PRODUCTION_LIMIT (default -1 = off). Caveats: it counts only production executions (webhook/trigger) — manual, sub-workflow, error, and CLI runs bypass it — and in queue mode a value other than -1 supersedes the worker --concurrency flag. Details: https://docs.n8n.io/deploy/host-n8n/configure-n8n/scaling/control-concurrency.

Binary data: database or external storage — NOT filesystem

• Queue mode does not support filesystem binary mode (the docs are explicit), even with a shared volume — workers wouldn't reliably resolve each other's files. The template sets N8N_DEFAULT_BINARY_DATA_MODE=database: binary data lives in Postgres, visible to main and every worker. Doc: https://docs.n8n.io/deploy/host-n8n/configure-n8n/scaling/handle-binary-data.
• database mode makes execution pruning load-bearing — big payloads now grow Postgres, so the template sets EXECUTIONS_DATA_PRUNE=true + age/count caps explicitly (binary data is pruned as part of execution-data pruning).
• External storage (S3 / Azure Blob) is Enterprise-licensed — n8n won't even start in s3 mode without a valid license. If licensed: N8N_DEFAULT_BINARY_DATA_MODE=s3 + N8N_EXTERNAL_STORAGE_S3_* vars, and an S3 lifecycle policy is mandatory (in s3 mode n8n delegates binary pruning to the bucket instead of doing it itself). Doc: https://docs.n8n.io/deploy/host-n8n/configure-n8n/scaling/use-external-storage.

Optional: dedicated webhook processors

For very webhook-heavy instances you can run n8n webhook processes (same env as a worker) and route /webhook/* + /webhook-waiting/* to them at the proxy, keeping the main process responsive. Set N8N_DISABLE_PRODUCTION_MAIN_PROCESS=true on the main so it stays out of the webhook pool. Most deployments don't need this — add it only when webhook intake is the bottleneck. Doc: https://docs.n8n.io/deploy/host-n8n/configure-n8n/scaling/enable-queue-mode.

Not in this template: multi-main (HA)

Running multiple main processes with leader election/failover exists but is Enterprise-only (N8N_MULTI_MAIN_SETUP_ENABLED=true on every main). On the community edition the main process stays a singleton; scale workers instead. If a user asks for "HA n8n" without an Enterprise license, that's the honest answer.

Memory

Queue mode wants more RAM than single: a practical floor is ~4 GB, with each worker wanting ~1–2 GB depending on workload — rules of thumb from real deployments; the docs don't publish sizing tables. For OOM crashes the docs' first advice is to redesign the workflow (batches, sub-workflows, less Code node), then raise the heap via NODE_OPTIONS=--max-old-space-size: https://docs.n8n.io/deploy/host-n8n/configure-n8n/scaling/fix-memory-issues. Confirm the box is sized before deploying.

Verify

docker compose ps     # postgres & redis healthy, then n8n (main) + workers Up
docker compose logs caddy | grep -i 'certificate obtained'
curl -fsS --retry 5 --retry-delay 10 https://<fqdn>/healthz
docker compose logs n8n-worker | grep -iE 'ready|listening|jobs'   # worker is up + listening

# main and workers must agree on everything except the public-URL vars
diff <(docker compose exec -T n8n env | sort) <(docker compose exec -T --index 1 n8n-worker env | sort)

A real test: run a workflow from the editor and confirm a worker logs that it executed it.