Skip to content

Implement atomic policy rollover on Firecracker #5392

Open

Description

@auser

Parent and dependency

#5386. Depends on #5391 (committed local project policy generation). Proposed ADR-111 describes host-trusted policy apply and prepare-then-cutover.

Outcome

One mvmctl policy apply invocation uses an automatically managed local coordinator to prepare an admitted Firecracker replacement, fence the old workload's network/secret/mailbox effects, drain/cancel affected in-flight work, switch the sole host-owned owner/route, retire the old instance, and durably report the committed generation. A failed candidate before cutover keeps the old workload only while the old policy is still authorized. No user-run daemon command or mvmd dependency.

Acceptance

  • Prove one local lifecycle owner coordinates apply with launch/resume and survives CLI exit; retain keyholder/broker process boundaries.
  • Keep the old instance serving during candidate preparation, but do not grant both active effect authority. Candidate build, signed admission and readiness are distinct checks.
  • Commit policy generation and ownership exactly once; audit preparation, fencing, drain, cutover, stop, unknown and refusal outcomes. Reconcile crashes at each transition.
  • Refuse on failed enforcement, stale epoch, replay, expired authority, missing/unknown state, in-flight effect uncertainty, quota exhaustion or unsafe stop. Preserve still-authorized accepted mailbox messages without equating acceptance to application acknowledgement.
  • Unit/integration witnesses at the production caller and a live Firecracker witness, including failure injection and cold/warm readiness with admission in the measured boundary. No all-backend or zero-downtime claim from this slice.
  • Update CLI docs, tests, and PR validation; keep remaining parent checkboxes open.

Activity

  1. auser commented on Nov 13, 2026

    @auser
    CollaboratorAuthor

    Operator decision for this workstream: use a distinct, automatically managed local lifecycle coordinator within the mvm-hostd boundary, not the per-tenant keyless mvm-host-agent and not mandatory mvmd. mvmctl starts or connects to it without a second operator command. Preserve the existing signer/broker key-custody boundary. Proposed ADR-111 records the decision; this is an approved architectural direction, not evidence of a production coordinator or atomic policy apply. #5391 (exact committed policy source closure bound to launch/restore) remains a prerequisite; do not claim apply success until production effect fencing, single-owner handoff and crash recovery have live Firecracker witnesses. HVF and the other backends remain separately open.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions