BigFleet

What BigFleet is

BigFleet is a fleet-level infrastructure autoscaler. It receives capacity needs from many Kubernetes clusters and provisions or reclaims machines through pluggable, out-of-tree CapacityProvider backends. It’s the reference implementation of the design described in the BigFleet paper and the Fleet-Scale Kubernetes paper.

BigFleet is not a scheduler. It does not place pods, simulate kube-scheduler, manage cluster lifecycle, or run quota / admission. It sits one layer below the cluster autoscaler.

Where to go next

• Quickstart — bring up BigFleet on a kind cluster in five minutes.
• Concepts — Need, Profile, Penalty, Cost, the three Phases, victim score.
• Architecture — two-tier design, decision engine phases, static stability.
• API reference — CRDs and gRPC services.
• Operator guide — install, metrics, runbook.
• Scaling guide — sizing for 10K to 100M machines.
• Provider author guide — implementing a CapacityProvider.

Status

v1 feature-complete. Tested via race-detector unit tests, multi-cluster e2e on kind, deterministic simulator with golden traces, long-running soak, provider conformance suite, and Helm chart render smoke tests. See the implementation plan for milestone history.

Real provider implementations (AWS, GCP, Azure, bare-metal) live in separate repos by design — see the provider author guide.