Spot Capacity Orchestrator  ·  View 29 of 32  ·  6 · Operations

Roadmap

An eight-week proof of the three things the design cannot argue its way past, a single-region MVP, then breadth.

Editable source SVG draw.io All views
Proof · 8 weeks Notice relay p99 ≤ 1 s 5,000 subscribers 800-node drain evict inside budget RGW at 10 GB/s 800 final writes MVP · one region Contracts · 5 tiers admission refuses Pools · caps · exclusion Agent · Kubernetes · Slurm Checkpoint client Spark · PyTorch Floor · capped fallback Daily 1% drill moved up from phase 2 Phase 2 Learned correlation Advisory drain per contract Warm headroom Three regions · quotas Commitment-aware fallback Poison-unit quarantine Phase 3 Generic checkpoint CRIU, where it fits Predictive scoring Contract recommendations Contract-priced chargeback Roadmap — Prove the Notice Path, Then One Region, Then Breadth A failed proof item changes the design before the MVP starts: a slow relay moves notice delivery onto the hypervisor, a slow store moves final writes to requeue. v 1.0 · owner Compute Platform Architecture · date 2026-09

Proof first

  • Notice delivery at 5,000 subscribers inside 1 s, an 800-node drain inside the budget on real Kubernetes and Slurm clusters, and 10 GB/s of final checkpoints into RGW. Each of them is a physical limit the architecture depends on, not a feature.

Departures from the requirement

  • The daily reclamation drill moves from Phase 2 into the MVP. It uses the provider's own reclaim API and costs about 1% of reclaimable node-hours in extra churn, and it is the only way to show fail-static and drain behaviour are still true on an ordinary Tuesday.
  • Everything else follows the requirement's phasing.

What a failed proof changes

  • A relay too slow at fleet scale moves notice delivery to a per-hypervisor channel. A checkpoint store that cannot absorb the burst changes tier-1 contracts from final-write to more frequent periodic checkpoints and requeue.