chore(cd): deploy pipeline per ADR-0110 round 29 (expand, deploy 100 %, smoke, auto-rollback)

Ouverte
  • Quelque chose ne va pas ?
#49 0 commentaires 0 réactions 0 personnes assignées Voir sur GitHub

Les mainteneurs répondent en général sous 1 jour

Personne n'a encore pris cette issue.

Évaluation

Difficulté
5/5
Temps estimé
Plus d'une semaine
Accessibilité débutants
25/100
Type d'issue
Fonctionnalité
Clarté
Plutôt claire
Activité
Active
Stack technique
github-actions, rust
Domaine
ci-cd, devops

Piste de recherche

Start with docs/src/architecture/cicd.md, ADR-0110, and the existing .github/workflows/ci.yml; inspect scripts/stage2-contract.sh for the deployed Worker checks and docs/src/architecture/infra.md for the migration access path. Then implement and document the ordered deploy, smoke, rollback, and asset-retention steps, including the staging failure controls and approval gate for Durable Object migrations.

Rédigé par le modèle d'indexation à partir du texte de l'issue.

Description

area:infra prio:p0 review:fable-2026-10-19

Context

Nothing deploys today. docs/src/architecture/cicd.md:80 reads "Any deploy. There is no workflow for it.", and :96-97 says nothing implements ADR-0110. The repository's workflows are ci.yml, contract-cell.yml and flake-hunt.yml. ADR-0110 (accepted in round 24) prescribes this sequence:

  • expand migration;
  • upload at 0 %;
  • smoke and an oracle subset through the version-override header;
  • 10 % for 10 minutes;
  • 100 %, with absolute-count gates and an automatic wrangler rollback (adrs/cd/0110-deploy-pipeline.md:38, :79; Q-0110.1, QUESTIONS.md:234).

The only script that tests a deployed Worker is scripts/stage2-contract.sh, which needs the test routes and the admin token (:36-39). Migrations are the migrate binary (docs/src/architecture/infra.md:128-129). Today the only way into the CNPG clusters is kubectl exec on the primary (research/0010 §2, lines 107-112).

Problem

  • FR-12: traffic is about 5 M requests a month (research/0004 §9, line 317), so 10 % for 10 minutes is roughly 100 requests, most of them the owner's. The ADR itself says percentage slices are meaningless at ≤50 users (0110-deploy-pipeline.md:79). The 10 % stage observes nothing the smoke suite does not, and the version-override plumbing costs build and maintenance.
  • Two constraints remain. A rollback does not revert connected resources, so migrations stay expand/contract (:15-17, :25). A deploy cannot be rolled back across a Durable Object lifecycle change (:18-19).
  • A third comes from ADR-0011 (round 26). The SPA fallback answers a missing hashed asset with 200 text/html, so deploys must keep old hashes until clients age out (adrs/ui/0011-static-asset-hosting.md:76). A 100 % deploy drops the previous version's assets.

Proposal

Add .github/workflows/deploy.yml, triggered on push to main and by hand for each environment (dev, staging and prod, per Q-0092.2), with credentials per ADR-0101.

  1. Expand migration. Q-0110.2 says CI runs it, but nothing gives CI a route to thrust-pg today, so this issue chooses one: a runner with cluster access, or a tunnel route with an Access service token. If #37 picks the Durable Object cell, this step becomes the object's own migration.
  2. Deploy. Record the current version id, then wrangler deploy at 100 %.
  3. Smoke. Run a new pass against the live hostname that needs no test routes: /healthz first; then, once #44 exists, sign in with a dedicated smoke account and read a seeded project. Vacuity guard: /healthz reports the deployed version id (from the version metadata binding, or a var set at deploy), and the smoke checks it equals the new one.
  4. Roll back. If the smoke fails, run wrangler rollback <previous id> and fail the job. If the release adds a [[migrations]] tag, skip the automatic rollback and require manual approval (the DO-lifecycle rule).
  5. Assets. Once #44 adds static assets, include the previous release's hashed files in each upload, or answer 404 for a missing /assets/* file.

Acceptance criteria

  • deploy.yml exists and runs the steps in order on push to main.
  • Control: on staging, a release with a smoke check that fails on purpose ends with the previous version active again (wrangler deployments list) and the job red.
  • Control: the smoke pass fails when pointed at a Worker that still serves the previous version id.
  • A release that adds a DO migration stops for approval and does not roll back automatically.
  • After a deploy, an asset URL from the previous release still returns its JavaScript, not HTML.
  • docs/src/architecture/cicd.md describes the pipeline, and ADR-0110 (round-29 note) and Q-0110.1 match it.

References

  • Review: .agents/plans/contrail/reviews/2026-10-19-fable-architecture-review.md (FR-12)
  • ADR-0110 (:13-20, :25, :38, :79; round-29 note), ADR-0101, ADR-0011 (:76), ADR-0090, ADR-0093, ADR-0130
  • Q-0110.1, Q-0110.2, Q-0092.2
  • docs/src/architecture/cicd.md:73-80, :94-97; docs/src/architecture/infra.md:128-129; scripts/stage2-contract.sh:36-39
  • research/0004-rust-on-workers-binding-matrix.md §9 (traffic model); research/0010-thrust-postgres-infra-plan.md §2 (database access path)

Blocked by: #43

Finding: FR-12

Langage dominant
Rust
Étoiles
0
Forks
0
Merge moyen
1 h 35 min
PR mergées (30 j)
67

Préparer son environnement

Par où commencer

  1. Lisez l'issue en entier, puis le guide de contribution du projet.
  2. Signalez en commentaire que vous la prenez — cela évite que deux personnes fassent le même travail.
  3. Forkez le dépôt et travaillez sur une branche.
  4. Ouvrez une pull request qui référence le numéro de l'issue.

Autres issues de aRustyDev/thrust

Toutes les issues de aRustyDev/thrust

Issues similaires

Plus d'issues Rust

Recevez les nouvelles issues par e-mail

Un résumé court des issues GitHub adaptées aux débutants.