Deploy pipeline — Export Platform
| | | |--|--| | **Owner** | Platform SRE | | **Cadence** | Continuous deploy (merge to main → staging → prod) | | **Last updated** | 2026-03-10 |
Deploy pipeline — Export Platform
| Owner | Platform SRE |
| Cadence | Continuous deploy (merge to main → staging → prod) |
| Last updated | 2026-03-10 |
Pipeline stages
PR ──merge──→ main ──build──→ staging ──gate──→ production
│ │
image + migration slow-roll (10% → 50% → 100%)1. Build (GitHub Actions)
Triggered on merge to main touching services/export-* or infra/:
| Step | What it does |
|---|---|
| Test | Unit + integration + export smoke (samples a real export end-to-end in staging) |
| Build | Docker image tagged git-sha, pushed to ECR |
| Migrate | bin/rails db:migrate against staging with expand-contract pattern (see DB schema doc) |
| Deploy staging | Rolling update of staging ECS service |
2. Staging smoke
Post-deploy checks run for 5 minutes:
- Export API health endpoint returns 200
- Can enqueue and complete a small CSV export
- Queue depth stable (no runaway retries)
- Alert: none firing for export-critical SLOs
3. Production gate
A human (or automated changelog check) approves the production deploy:
| Check | Who / what |
|---|---|
| Changelog entry exists for user-visible changes | PR template enforces |
| Migration dry-run against prod clone | CI step (commented until DB migration exists) |
| Canary smoke passes | 5-min monitoring window on canary |
4. Production deploy
Slow rollout via ECS:
| Stage | % traffic | Hold | Rollback if |
|---|---|---|---|
| Canary | 10% | 5 min | Error rate > 0.5% or p95 latency > 2x baseline |
| Half | 50% | 10 min | Any SEV-2 alert firing |
| Full | 100% | — | — |
Rollback: git revert + regular pipeline. ECS retains the previous task definition for immediate manual rollback via aws ecs update-service.
Database migrations
We follow expand-contract for zero-downtime deploys:
| Phase | Example | Deploy stage |
|---|---|---|
| Expand | Add column export_jobs.schedule_id as nullable |
Deploy 1 (code reads both old and new) |
| Migrate | Backfill schedule_id for existing rows |
Deploy 1 + background job |
| Contract | Remove old column, make schedule_id NOT NULL |
Deploy 2 (code no longer references old) |