Skip to content
MultihullMultihull

Reliability principle

Multihull exists to keep models serving. Defaults spend money to buy redundancy.

Defaults

Section titled “Defaults”
reliability:
  minWarmProviders: 2
  overprovision: 1.4
  fallbackScaleToZero: false
  spot: false
  • At least two providers warm.
  • Fallbacks never scale to zero.
  • On-demand before spot.
  • Capacity over-provisioned by 1.4x, the Envoy priority-spillover factor.

Cost features are opt-in

Section titled “Cost features are opt-in”

Cheapest-GPU placement, spot, and scale-to-zero fallbacks exist but must be enabled, and none can lower redundancy below the configured floor.

Fail loudly before deploy, quietly during

Section titled “Fail loudly before deploy, quietly during”

hull doctor and hull plan surface credential, quota and GPU-availability problems up front. At runtime, failover is invisible to callers.

Boring infrastructure

Section titled “Boring infrastructure”

Files in git, SQLite or object storage for state, gRPC or a JSON file between control plane and router. No message bus, no database cluster required.