One process, no Redis

The situation

An admin panel needs background work: CSV imports, exports, scheduled reports, outbound email, automation runs, LLM calls. The industry answer is a queue service, a worker process, and a scheduler — three things to deploy, three things to monitor, and a Redis instance to lose.

Adminium’s deployment target is one container, one VPS, or a desktop app on a laptop. Every piece of infrastructure it requires is a cost paid by every person who self-hosts, forever.

The decision

One Node process, and nothing beside it.

Concern	How it works instead
Queue	adminium_jobs, a table. Workers claim a row with a portable UPDATE guard, so claiming is safe without a lock service.
Retries and dead-lettering	Columns on the same row: attempt count, exponential backoff, a terminal status.
Scheduler	croner inside the process. Each named schedule enqueues a job; a guard skips the tick while the last job it enqueued is still pending or running, and optional jitter de-synchronises fleets.
Rate limiting	A fixed-window counter in a Map, plugged into @fastify/rate-limit as a custom store. Keys embed the bucket name, so buckets share one map without colliding.
Realtime	Server-sent events and a WebSocket from the same server.
Cache	The database, or a per-process memo. There is no cache tier.

What it costs

This is the honest part, and it is not in the marketing copy.

• Two replicas double-enqueue schedules. The no-overlap guard remembers the last job this process enqueued. It is in memory, not in the database, so a second replica’s scheduler does not see it. Jitter spreads the ticks out; it does not deduplicate them.
• Rate limits multiply by the replica count. The counter is per process, so N replicas grant N times the allowance.
• Extra workers are safe. The claim guard is the one piece that is correct across processes.

So the supported shape is one process. Running more is possible, and the two caveats above are what you are signing up for.

What it means for a contributor

• New background work is a job kind, registered in apps/server/src/jobs/. It is never a new process, a setInterval in a route file, or a detached child.
• Do not reach for a cache that is not the database. If a value is worth keeping between requests, it is worth a table or a column.
• Do not add a dependency on a broker. If a feature seems to need one, the feature needs a different shape.

The self-hosting pages say the same thing to operators: Overview & requirements opens with it.