jvm.chisel.db.cluster

Multi-query admission control for the chip: the control plane that turns the fixed pool of reusable operator blocks into a shared, multi-tenant resource.

Where jvm.chisel.db.schedule answers “can this plan run on this inventory, and where?”, this namespace answers the scheduler’s real question — “which of the currently free blocks does this query get, and what does everyone else keep?”.

The model is the essay’s “allocate CPU cores, not synthesise hardware” loop:

query arrives ─▶ admit (check free units, pin placement, record cost) query finishes ─▶ release (units return to the free pool) next query ─▶ admit (reuses the same physical blocks, no recompile)

State is a plain persistent map — cluster, admit, release are pure functions, so a rejected admission allocates nothing (atomicity is structural) and an admissions queue could be threaded through an agent or atom without changes here.

Placement within a query is first-fit: each node draws the lowest-indexed free unit of its kind. Preemption, defragmentation and runtime rebalancing are deliberately deferred — the seams (:cost-fn on admit, the :allocations map being plain data) are where they plug in later.

admit

(admit cl query-id plan)(admit cl query-id plan {:keys [cost-fn]})

Try to admit plan as query-id onto the cluster. Pure: returns {:ok? true :cluster cl’ :admission {…}} with the updated state and the admission record ({:query :placement :units :cost}), or {:ok? false :reason …} with the cluster untouched. Refuses when: * query-id is already admitted, * the plan uses an operator the chip has no blocks for, * free units of any kind cannot cover the plan’s demand. Unit choice is first-fit (lowest free index per kind), excluding units pinned by queries already admitted. opts may carry :cost-fn, the same learned-cost-model seam as schedule/schedule.

view source