Spot Capacity Orchestrator · View 16 of 32 · 4 · Data
Decisions
- Eligibility first: a pool the workload cannot use (architecture, accelerator, local NVMe, 100 GbE) is never ranked for it, however cheap.
- Depth before price. The score weights forward capacity and observed interruption above rate, because on premises the cheapest pool is cheap precisely because guaranteed tenants are about to want it.
- A pool whose reclaimable rate reaches 85% of the guaranteed rate stops being a discount pool. Running there buys interruption risk for almost nothing.
The on-premises advantage
- Reclamation here is caused by guaranteed demand on shared host aggregates, and the provider can publish both the aggregate membership and the reservation calendar. Correlation groups start from that known topology instead of being guessed, and the forward signal is a calendar, not a probability.
Retention
- Rankings and their inputs: 90 days at full fidelity, 13 months aggregated. Pool behaviour is seasonal, with quarter-end and year-end demand from guaranteed tenants, so a year is the minimum useful window.
