Skip to content
BigFleet Providers

BigFleet Providers

Out-of-tree capacity providers for BigFleet — one shared library, one conformance harness, one place to read.
Provider author guide GitHub BigFleet

What this is

BigFleet is a fleet-level infrastructure autoscaler. It decides which machines should serve which Kubernetes clusters and provisions, reclaims, and rebalances them through pluggable CapacityProvider backends. A provider is the component that actually creates, configures, drains, and deletes machines on a specific substrate — AWS, GCP, libvirt, bare metal, and so on.

BigFleet ships zero real providers in its own repo, on purpose. Kubernetes spent years undoing in-tree CCM/CSI providers; BigFleet does not repeat that. Every real provider lives here, in bigfleet-providers — a repository kept deliberately separate from the main BigFleet repo.

Why one repository

Rather than one repo per provider, every provider lives together in a mono-repo so they share:

  • one correctness-critical library — providerkit — that gets fencing, idempotency, async dispatch, shard_metadata, and the machine field shape right once, so each provider only writes substrate-specific logic;
  • one conformance harness — pointed at the canonical acceptance suite in the BigFleet repo, so “BigFleet-compatible” means a passing run, not a promise;
  • one CI pipeline, and one place to read.

Each provider is still an independently buildable binary that can ship on its own cadence.

The contract

A provider is a gRPC server implementing bigfleet.v1alpha1.CapacityProvider; the BigFleet shard is the client that dials it. The contract is six RPCs — no Watch; reconciliation is List + Get:

RPCLifecycle
CreateSpeculative → Creating → Idle
ConfigureIdle → Configuring → Configured
DrainConfigured → Draining → Idle
DeleteIdle → Deleting → Speculative
Get / Listread inventory

Every provider owes the same cross-cutting obligations — async dispatch, idempotency, transition timeouts, fencing, the shard_metadata lifecycle, and the machine field shape. providerkit implements all of them so a provider author cannot get them subtly wrong. The authoritative wire contract and author guide live in the BigFleet repo and are consumed here from the Go module, never vendored — so the contract can’t drift.

Providers

ProviderCapacity typesStatus
_templateon-demand + spot (example)copy-me skeleton — passes conformance against an in-memory backend

Real providers (AWS, GCP, libvirt, …) are added by copying _template; this table grows as they land.

Build a provider

  1. Read the provider author guide — the spine.
  2. cp -r providers/_template providers/<name> and implement the substrate Backend.
  3. Declare capacity_type, price_per_hour, and interruption_probability honestly (a real probability for spot).
  4. Get make conformance-<name> green.

The step-by-step recipe is in CONTRIBUTING.md.

Where to go next