The cluster is a copy loop over the boxes you already rent

Embarrassingly parallel work, 40,000 documents to embed or 300 hyperparameter points to score, usually gets a managed cluster: a head node, an autoscaler, and a bill in node-hours that runs while the workers wait for you.

On Hoody the cluster is a loop. You tune one worker container, snapshot it as worker-v3, and call POST /api/v1/containers/{id}/copy forty times, cycling target_server_id across the three servers you rent. Each copy starts on its own and pulls its next item with POST /api/v1/sqlite/kv/queue/pop?db=jobs against one coordinator's SQLite URL. There's no head node, because the queue has an address.

When the work ends, the forty copies sit idle, and idle copies cost the flat-rate boxes nothing extra: they're copy-on-write clones. Next week you copy again.

There's no autoscaler. You don't scale. You copy, and the bill is three servers.

You don't scale. You copy, and the bill is three servers.