Skip to main content
OperationsEngineering
What this does: Defines how new model versions move safely from staging to production using a 2-week promotion cycle — ensuring zero-downtime updates and easy rollback if issues arise.
Our deployment strategy prioritizes safety and zero-downtime updates. We utilize a 2-week promotion cycle with distinct “Staging” and “Production” environments detached via Docker Compose configuration.

Environment Architecture

Promotion Lifecycle

Week 1: Staging (Integration)

  1. Deploy Candidate: Tag the release v3.0.0-rc1.
  2. Smoke Tests: Run critical path tests (pytest -m smoke).
  3. Performance Verification: Ensure inference latency < 200ms.

Week 2: Production (Go Live)

  1. Backup: Snapshot the current PostgreSQL metadata.
  2. Traffic Shift: In a Blue/Green setup, route traffic to the new containers.
  3. Monitoring: Watch DriftDetector logs for immediate anomalies.

Scripts & Automation

We use tailored scripts to automate these risky operations.

Rollback Procedure

If a critical failure is detected (e.g., Error Rate > 1%), execute an immediate rollback:
  1. Revert Image: Point docker-compose.prod.yml back to v2.x.x.
  2. Restore DB: If schema migration failed, restore from the pre-deploy backup.
  3. Restart: docker-compose up -d --force-recreate.