Redis

pip install 'tinyjobs[redis]'

app = TinyJobs("redis://localhost:6379/0")
app = TinyJobs("redis://localhost:6379/0", prefix="billing")    # several apps on one server

Nothing else changes: the same tasks, workers, cron, hooks, stats and check. Every state change is one Lua script, so a claim is as atomic as on SQLite, and any number of workers on any number of machines can share the queue. With the default clock every machine keeps Redis's time, so their own clocks need not agree; a run_at or Retry(at=...) you pass is read on the local clock and moved onto Redis's. Reading that time takes a round trip, so on a host too loaded to answer within half a second a delayed job can start up to a quarter of a second early. A clock= you pass yourself is used as it is, and must then agree across machines. Leases hold even while a machine still runs on an old reading: the scripts never reap by a time ahead of Redis's and never stamp a lease from one behind it. Keep the Redis server and its replicas on NTP: after a failover to a node whose clock is ahead by more than a lease, every running job looks abandoned and runs again, or fails when it has no attempts left. When upgrading from 0.7.0, which used each machine's own clock, stop or upgrade every 0.7.0 process, workers and senders alike, before starting the new workers if their clocks drift from Redis's by more than a lease, and grant TIME to a Redis user restricted by ACLs. It passes the same contract suite as SQLite.

Turn persistence on. Redis keeps the queue in memory. Killed with SIGKILL mid-run, a Redis on its default snapshots lost the whole queue; with appendonly yes the same crash lost nothing, and the jobs that were running came back once their leases expired. A restart, graceful or not, is otherwise harmless: workers log the outage, back off, and carry on.

Worth knowing: a job costs about 1 KB of Redis memory, retention prunes it like on SQLite, and a job cannot join your database's transaction. All keys share one hash slot ({prefix}), so it runs on a single Redis or a primary with replicas, not spread across a cluster.

Each pool holds at most 100 connections (max_connections=). A burst beyond that waits for a free one rather than failing, and gives up with redis-py's ConnectionError after pool_timeout= seconds (20). The app has one pool, opened on first use; worker.run() and tinyjobs worker open one of their own for claims, outcomes and cron, plus one for each extra loop.