Documentation overview · Architecture guide
This proposed sequence follows the reviewed Fleet designs. Milestones are unimplemented. Owned-metal qualification and rented/managed adoption can progress independently while sharing allocation and installation contracts.
The Machine lifecycle enhancement defines the five public resources, private reservation/operation records, Milo tenancy boundaries, and recovery acceptance criteria. Its functional target uses a persistent simulated provider; real hardware qualification remains in stage 3 below.
Record decisions for API scope and field ownership, project identity and IAM, durable reservation/operation storage, artifact retention, provider fact schemas, and the initial implementation toolchain and module path. Introduce code and generation tooling after those choices are explicit.
Acceptance: demonstrate a recoverable exclusive reservation and quota protocol, including two projects racing for one machine, policy withdrawal at commit, retries, and project/object recreation. No hardware effects are needed.
Implement physical plans, reservations, trusted registration, and evidence evaluation. Import existing inventory, provider state, and active allocations in observation mode. Expose contradictions and unknown facts.
Acceptance: conflicting rack/port claims and wrong-host registration are rejected; import does not alter power, boot, networking, or an active reservation.
Integrate one execution adapter on disposable lab capacity and independent management infrastructure. Connect class publication, shared allocation, networking, OS installation, console/power, and release.
Acceptance: a supported non-Talos customer OS completes claim, installation, network handoff, operation, reinstall, and release without Cluster/NodeClaim or a mandatory ongoing guest agent. Two consumer projects share provider inventory without double allocation. Lost responses, stale executors, access revocation, and incomplete cleaning cannot cause duplicate effects or premature reuse.
Normalize servers.com/NetActuate facts, retain existing identities and reservations, and compare candidate configuration with current outputs. Introduce managed recipes, membership, consumer pool gates, and observed capabilities. Transfer each generated output to one Fleet-controlled writer only after parity.
Acceptance: adoption causes no reinstall or silent rename; replacement uses new claims with proven predecessor fencing; paused-rollout repair preserves each node's admitted revision; procurement follows operator policy.
Observe existing Private DNS deployments. Introduce a resolver NodeService as a third producer behind a new frontend in one staging cell, then enable readiness gating and migrate the existing frontends. Coordinate runtime class installation through OS configuration before retiring the old per-host wiring.
Acceptance: new nodes wait for every gated slot, a single failed slot degrades without evicting tenants, rolls preserve a serving slot, drains preserve service until tenants leave, and replacement never activates unfenced shared addresses. Before expanding to another region, prove independent broker storage and writers allow it to publish and serve during another region's outage.
Qualify bounded rollouts, provider failure and takeover, backup restoration, physical maintenance, and another execution integration. Add deployment/release automation and operational guidance for the qualified implementation.
Acceptance: recovery cannot replay stale authority, discard cleanup obligations, or allocate quarantined capacity. Wider rollout follows demonstrated lifecycle and failure handling rather than controller process health alone.