Description
Scope
Make production launch latency acceptance one unambiguous contract before optimizing it. Governing program #5415 and architecture #5413. No warm or production speed claim follows merely from this benchmark contract.
Acceptance
- Reconcile ADR-001's cold <300 ms / warm <50 ms including admission with
public/src/content/docs/reference/performance.md's prepared-artifact verification and dispatch windows. Pin the invocation-to-guest-readiness start/end points and required verification, admission and per-sample refusal checks; do not move an irreversible effect's check past its last undo point to shorten a report. - Define separate acquisition, first-use publication/assembly, archive verification, prepared cold, warm claim and replacement evidence for the x86_64 Linux/KVM/Firecracker first target. A warm miss is cold, not a warm sample. Keep the existing at-least-20-measured-sample/two-warm-up report rule or obtain an explicit contract change.
- Make reports reject missing/mislabelled phases, stale/unsupported backend evidence and a claim that bypassed current authorization. Add a hermetic reporter/refusal test and record the named host and exact candidate revision for live witnesses.
Context and validation
Read ADR-001 check-time law, public/src/content/docs/reference/performance.md and draft ADR-059 on PR #5326. Mirror the existing xtask performance report/claim witness gates; do not infer production latency from local mvmctl Dev runs or a direct-VMM probe. Validate reporter tests with cargo test -p xtask, cargo run -p xtask -- check-all, and exact-head CI. Live measurements belong to the final evidence ticket, not this contract-only ticket.
Depends on: owner timing-contract review. Independent of: the first schema and warm API reviews. All release-speed claims depend on this decision.
Activity
auser commentedon Nov 13, 2026
Owner/performance review packet — production clock
This is a proposed measurement contract, not a latency witness or a change to today's requirements. ADR-001 requires admission before boot/effect; the project requires cold <300 ms and warm <50 ms including admission. The current performance reference separately reports bundle verification outside the prepared-cold dispatch window, while also requiring startup <300 ms from command invocation. Those clocks must be reconciled explicitly for a composable production bundle rather than silently borrowing the favorable one.
Recommended production lanes
source_acquisition,first_use_publication,bundle_verify,prepared_cold,warm_claim, andsafe_replacementas separate phases; do not average a miss into a prepared hit. Publish both full user/API-to-guest time and host-authoritative-request-to-ready time, naming the start timestamp and selected backend.mvmdlaunch request at the host through fresh bundle/input/revocation verification, signed-plan and host-policy admission, actual VMM start or claimed child, to guest-ready/allowed first effect. Report coordinator/network overhead separately, not as proof of the host gate. No check may be performed after its last undo point or counted only on an unrelated preflight.Decision required
Should fresh full-bundle verification be inside the gated production invocation window, or separately pre-published as
bundle_verify_mswith a new freshness/lease proof that still checks current authority during the gate? Recommend the former for the first production claim; the latter is only acceptable after an explicit security/performance ADR and a live proof that no revocation or changed-byte window opens. This may expose an inability to meet the budget; do not redefine the clock to hide it. No production speed checkbox should close on a localmvmctlDev run or a direct-VMM probe.auser commentedon Nov 13, 2026
Owner approved the recommended strict production-clock direction in #5413: fresh full-bundle verification and admission stay inside the invocation-to-readiness evidence, while acquisition and first-use publication are reported separately. The performance reference/ADR must still be reconciled and reviewed; no speed witness or checkbox is complete.
auser commentedon Nov 13, 2026
PR #5436 proposes the authenticated production invocation-to-readiness measurement boundary and clarifies why the existing prepared-artifact dispatch lane does not prove it. This is docs-only; production instrumentation, backend-specific live evidence, strict cold/warm acceptance and failure-path witnesses remain open. No acceptance checkbox is complete yet.
auser commentedon Nov 13, 2026
PR #5436 merged the planned production measurement boundary into the build contract and public performance reference. The current
prepared_cold_artifactdispatch budget remains unchanged and is explicitly not a production-composition speed witness. The new reporter/refusal gate, backend-specific live samples and exact cold <300 ms / warm <50 ms evidence (including admission) are still required; keep all acceptance boxes open until those are demonstrated.