Deployments Guideline
Tech Stack
- Workflows: Upstash Workflows + QStash
- Cache: Upstash Redis
- Platform: Vercel
Non-Negotiables
- Rollback plan must exist and be exercisable
- Dead-letter handling must exist and be operable (visible, replayable)
- Side-effects (email, billing, ledger) must be idempotent or safely re-entrant
- Drift alerts must have remediation playbooks
Context
Deployments handles post-deployment operations — rollback, feature flags, runbooks, incident response. Pre-production CI/CD lives in delivery.
When something goes wrong, can an operator fix it without deploying code?
Driving Questions
- What happens when a job fails permanently?
- How would an operator know something is stuck?
- Can failed workflows be safely replayed without duplicating side-effects?
- What's the rollback plan if a deploy breaks something critical?
- What runbooks exist, and what runbooks should exist but don't?