0.7 release readiness: gates, milestones, and evidence tracker

[hardbyte]

Purpose

This is the single release-readiness tracker for 0.7. It records the gates above individual PRs, the evidence that satisfies them, and explicit deferrals. The 0.7 roadmap remains planning context; an issue appearing in the milestone does not by itself make that issue a release blocker.

Current candidate — 2026-07-14

• main: 1e519307 (feat(queue-storage): staged ring-ledger authority and mixed-version upgrade #425 merge).
• Merged release stack: fail-safe migration handling and rollout preflights (fix(migrations): fail safe on newer schemas, pre-flight live-runtime check, migrate progress (#392) #426), staged ring-cursor authority (feat(queue-storage): staged ring-ledger authority and mixed-version upgrade #425), idle rotation skip (perf(queue-storage): skip ring rotation when idle (#371) #409), compact deadline claims (perf(queue-storage): compact deadline claim batches (#246) #410), append-only ring rotations (perf(queue-storage): ring cursors via append-only rotation ledgers (#371) #415), tracing (Tracing #110), health/readiness (Worker health & readiness endpoints (+ awa health probe) #368), runtime overrides (Hot-reloadable per-queue runtime overrides — Tier 1: dispatch knobs (#385) #396), and the canonical-engine deprecation gate (Deprecate the canonical engine: 0.7 migrate gate requires finalized clusters; remove paths in 0.8 #370).
• Required N-1 stepping-stone: v0.6.2, published for Rust, Python, Docker, and standalone binaries.
• Current main CI: 29222160609, green.
• Latest nightly on the merged stack: 29239037678, green.
• Release PR chore: cut 0.7.0-alpha.1 #421 is stale and conflicts with main; its changelog still describes the superseded lockstep migration.
• ADR-041 PR docs(adr): define rolling upgrade policy and migration checklist #429 was merged into feat(queue-storage): staged ring-ledger authority and mixed-version upgrade #425's feature branch after feat(queue-storage): staged ring-ledger authority and mixed-version upgrade #425 reached main. Its commit is not on main and must be forwarded before release.

0.7.0-alpha.1 gates

Alpha status does not relax migration correctness: the alpha writes the real v041-v043 schema and must pass the same mixed-version proof as a stable release.

Upgrade contract

• Publish 0.6.2 with the newer-schema fail-safe and v043 compatibility guard.
• Merge fix(migrations): fail safe on newer schemas, pre-flight live-runtime check, migrate progress (#392) #426 and feat(queue-storage): staged ring-ledger authority and mixed-version upgrade #425.
• Land ADR-041 and its migration-authoring checklist on main (docs(adr): define rolling upgrade policy and migration checklist #429 follow-up).
• Add Automated adversarial upgrade rehearsal (mixed-version, live traffic) in CI #427's automated adversarial rehearsal as a required pre-tag workflow, using the released 0.6.2 artifact and the current build.
• Add a pinned 0.6.2 forward-compatibility leg to scripts/compat-matrix.sh. The supported N-1 contract must be exercised directly; the existing 0.6.0 leg may remain as an additional data-plane regression test.
• Prove migrate-first, binary-first, and overlapping migration/binary rollout orderings.
• Prove concurrent 0.6.2/0.7 claim and completion behavior under live traffic.
• During the all-0.6.2-on-v043 interval, prove crash and heartbeat rescue continue, then prove deadline-overrun batch claims are rescued exactly once after the first 0.7 maintenance runtime starts.
• Force a final 0.6.2 column-only rotation immediately before flip; reconcile and verify all three ledgers under the flip locks.
• Restart the released 0.6.2 artifact after flip and prove the database fence rejects it before rotation or pruning.
• Reconcile every accepted job to exactly one designed terminal outcome, including retry, DLQ, cron, scheduled, and hard-killed work.

Performance contract

• Run the exact release candidate through the 800 jobs/s reference gate and record latency, depth, and completion rate.
• Run the W=64/128/256 saturation sweep and report release-candidate numbers without combining results from earlier feature branches.
• Repeat the 60-minute pinned-MVCC reference soak in ledger authority, followed by the recovery window; require bounded depth during the pin and full reclaim after release.
• Record the RFC: v0.7 storage engine — segment storage + cursor allocator #295 Gate A decision. Current evidence supports incremental storage evolution; the full segment-engine redesign is not implied by the landed ring/receipt work.

Release mechanics

• Rebase chore: cut 0.7.0-alpha.1 #421 after every release blocker above lands.
• Rewrite its changelog for the staged rolling upgrade, fix(migrations): fail safe on newer schemas, pre-flight live-runtime check, migrate progress (#392) #426/feat(queue-storage): staged ring-ledger authority and mixed-version upgrade #425, and the published 0.6.2 prerequisite; keep Unreleased first and use the actual release date.
• Run full CI on the final chore: cut 0.7.0-alpha.1 #421 head, including migration shards, Python, E2E, chaos, telemetry, CLI wheel, and package version consistency.
• Smoke the built Rust crates, Python wheels, standalone binaries, and Docker image before tagging.
• Tag and publish v0.7.0-alpha.1; verify every public registry independently.

Stable 0.7 gates

Correctness and public surfaces

• Fix awa.jobs compat view: pathologically slow under un-pruned backlog and blind to receipt-plane running claims #422, or explicitly remove awa.jobs from the supported operator/read surface. It currently misses receipt-plane running claims and can time out under a large retained backlog.
• Audit the Rust public API against 0.6.2 and establish the post-0.7 patch gate (CI: enforce the stability policy's Rust-API row with cargo-semver-checks #402).
• Make Python exports/stubs agree under CI (Python DX: complete type stubs with a CI typing gate + error-message audit #378).
• Complete or explicitly narrow the 0.7 promises around the SQL producer contract and HTTP response types (Stabilize the SQL enqueue contract as a public producer API #342, awa-ui response-type discipline: handlers stop returning JobRow (the hygiene half of #143) #403).
• Keep the full TLC suite and all expected counterexample/witness configurations green on the release candidate.

Stability and operations

• Accumulate 14 consecutive green nightlies after the final migration and compatibility-harness shape lands.
• Treat Nightly flakes: receipt-plane rotation floor and mixed-fleet quiet-window assertions lack runner-contention margin #399/CI flake: Python suite fixture pool-timeouts in late test_start_config tests (~15 min into the run) #420 as release blockers only if they recur; every new failure gets an issue and root-cause evidence.
• Rehearse alpha.1 on a real finalized 0.6.2 cluster through migration, mixed fleet, flip, sustained traffic, and rollback within flip-aware builds.
• Publish a concise 0.6-to-0.7 operator procedure and verify every command against the released artifacts.
• Produce stable release notes with exact benchmark provenance, known limitations, and package verification.

Explicit non-blockers and deferrals

• Deep fragmented backlog wedges drain at ~40% of ceiling with no recovery (single-generation claim filter) #418 is a pre-existing extreme fragmented-backlog throughput ceiling, not a regression introduced by the 0.7 stack. Keep it as follow-up unless release-candidate evidence worsens the 0.6.2 behavior.
• E9: planner & storage micro-tuning matrix (plan_cache_mode, extended statistics, group commit, BRIN spike) #419's experiment matrix is complete; route its deployment recommendations to operations documentation and close the experiment issue.
• awa queue await-drain: block until in-flight work is zero (upgrade windows, deploys) #424 (queue await-drain) is useful deployment ergonomics, not required for correctness because the existing status/finalize surfaces expose the gates.
• The full RFC: v0.7 storage engine — segment storage + cursor allocator #295 segment-engine redesign is deferred unless a separately recorded Gate A decision brings a bounded subset into scope.
• Unstarted product expansion does not gate 0.7: dependencies (Job dependencies: run job B only after job A completes #14), serverless tick (Serverless-friendly dispatch: tick() endpoint for zero-infrastructure deployments #118), per-key controls (Per-key execution control: concurrency limits and fairness within a queue #340), producer backpressure (Backpressure policy: queue depth limits and producer-side flow control #341), callback deployable (Callback ingress as a deployable: awa callbacks serve + promote ADR-027 to Accepted #372), Helm (Kubernetes deployment artifacts: first-party Helm chart (or kustomize base) #344), attempt timeline (Admin UI: per-attempt timeline view with trace link-out #375), and the versioned docs site (Versioned documentation site #377).
• Maintenance-role/scheduling work (Add a maintenance-only runtime role #282/maintenance: split rescue + promote into a separate tokio task (v0.7) #303), additional operations tooling (awa doctor: one-command operational diagnostics #373/Connection-pooler compatibility: NOTIFY self-test, polling fallback, pgbouncer CI leg, docs matrix #374/Grafana alert pack: codify the operational watch-list as importable alerts #376/Operations handbook: consolidate the production-running docs into one narrative #379), and further runtime-override tiers (Hot reload for worker runtime parameters #385) ship only if independently ready; they do not delay the tag.

Evidence log

Date	Candidate	Evidence	Result
2026-07-13	1e519307	main CI 29222160609	Green after #425 merge
2026-07-13	1e519307	nightly 29239037678	Green: chaos, compatibility, failover, and models
2026-07-13	v0.6.2	release	Rust, Python, Docker, and 12 GitHub assets published
2026-07-12/13	0.6 transition candidates	#427 comments	Live canonical-to-queue-storage rehearsals passed; these are not mixed 0.6.2/0.7 evidence

Add evidence here only when it names the exact SHA/artifact, workload, run link, and pass/fail boundary.

Release order

1. Forward ADR-041/docs(adr): define rolling upgrade policy and migration checklist #429 to main and refresh this tracker.
2. Build and pass Automated adversarial upgrade rehearsal (mixed-version, live traffic) in CI #427 plus the pinned 0.6.2 static compatibility leg.
3. Run the final performance gates and record RFC: v0.7 storage engine — segment storage + cursor allocator #295's decision.
4. Rebase and rewrite chore: cut 0.7.0-alpha.1 #421; run full CI and package smoke tests.
5. Publish alpha.1 and dogfood it through a real upgrade.
6. Close the stable-only correctness/public-surface gates and complete the nightly window before v0.7.0.

[hardbyte]

2026-07-12 release review: not ready for alpha.1

The merged #409/#410/#415 runtime changes look coherent and the measured p99 / receipt-working-set improvements are supported. The release train still has blocking upgrade and validation work:

• fix(migrations): fail safe on newer schemas, pre-flight live-runtime check, migrate progress (#392) #426 is green but still needs its dedicated Python live-runtime exception, plus a 0.6.x backport/step-stone for current_version's legacy normalization destructively rewrites schema_version for schemas newer than the binary #392.
• feat(queue-storage): staged ring-ledger authority and mixed-version upgrade #425 must reconcile column cursors into all three ledgers inside the flip transaction and replace the ineffective current_slot = -1 fence with a real old-binary startup/heartbeat rejection. Its draft CI currently skips the integration suites.
• Automated adversarial upgrade rehearsal (mixed-version, live traffic) in CI #427 is still required: the manual rehearsal was stop-the-world after a hard kill, not a live mixed-version rehearsal.
• Nightly chaos/benchmark failure (2026-07-10) #412 remains open on current main: two fresh nightlies fail Rust chaos and the pinned compatibility matrix; the latest also fails failover smoke.
• Migration runner: contain mid-run failures (savepoint-per-migration or single transaction) #423 was closed as obsolete because fix(migrations): transaction-scoped advisory lock for cancellation-safe, atomic upgrades #408 already made the current migration run atomic.
• chore: cut 0.7.0-alpha.1 #421 must remain last and be rebased after the above; its current release notes still describe the superseded lockstep v042 design.

Benchmark wording also needs source qualification: the checked-in final-stack gate is ~10.4k/s versus ~10.6k/s old with ~30% lower p99; the ~12.9k/+15% result is the workload-specific #409 branch cell, not the checked-in final integration pair.

[hardbyte]

Release-train update: #426 is green and mergeable at f90b08b5; the required 0.6.2 stepping-stone is #428 at 70da3140; #425 is at b9d5d237 with flip-time reconciliation, a database-enforced post-flip fence, corrected durable upgrade wording, and a passing pinned-artifact compatibility matrix. #425 remains draft until #426 and #428 land and it is rebased on their final heads.

[hardbyte]

Tracker reset on 2026-07-14. The issue body is now the authoritative release state: #425/#426 and v0.6.2 are complete; #427 is the primary remaining upgrade gate and is now in the v0.7.0 milestone. Earlier comments are retained as historical review context, not current status.

[hardbyte]

2026-07-15 release-gate update:

• test(failover): count pending terminal rollup deltas #432 merged as a26e45cf; the failover smoke race was a test reader omitting pending terminal-rollup deltas. Its complete nightly and full-CI rerun are green, and Nightly chaos/benchmark failure (2026-07-10) #412 is closed.
• Draft test(upgrade): rehearse released 0.6.3 against current #433 (f9df8b0a) starts Automated adversarial upgrade rehearsal (mixed-version, live traffic) in CI #427 with the pinned 0.6.2 static contract, released-0.6.3/current migrate-first traffic, all-three-ledger flip reconciliation, old-artifact fence, hard-killed maintenance-leader deadline rescue, and exact terminal reconciliation.
• The first Linux rehearsal found real migration/live-worker deadlocks. test(upgrade): rehearse released 0.6.3 against current #433 now retries only SQLSTATE 40P01 at the complete atomic transaction boundary and includes a deterministic regression that proves PostgreSQL detects the lock cycle and the retry reaches v043.
• The next Automated adversarial upgrade rehearsal (mixed-version, live traffic) in CI #427 ordering cells require a real 0.7.0-alpha.1 build. main still reports 0.6.1, so current source cannot honestly prove binary-first capability signaling or auto-flip. The version bump must precede those cells; changelog/release-note finalization can still remain last.
• v0.6.3 is the latest released N-1 artifact used dynamically; v0.6.2 remains the documented minimum stepping-stone and static compatibility floor.

Final-head #433 full CI is running. Its nightly is queued behind a superseded workflow that GitHub has not yet terminated despite cancellation requests.