Model Promotion Pipeline Example

A fraud detection pipeline demonstrating model promotion across branches.

Feature	Where	What it does
@project_schedule	TrainFlow	Cron schedule on main only; other branches deploy without a schedule
@project_trigger	EvaluateFlow	Automatically runs when TrainFlow publishes training_complete
promote_assets	CI workflow	Promotes assets from feature branch to main on PR merge

Pipeline

Feature branch:
  TrainFlow (manual trigger) → EvaluateFlow (pass/fail gate) → ReportFlow

On PR merge:
  CI promotes assets to main with @candidate → teardown feature branch

Main branch:
  TrainFlow (weekday 8 AM ET cron) → EvaluateFlow → ReportFlow

Main vs Feature Branches

On a feature branch:

• TrainFlow deploys but has no cron (triggered manually)
• EvaluateFlow auto-triggers after training, reports pass/fail
• Models stay on the feature branch until PR merge

On PR merge:

• CI runs promote_assets() to copy assets to main with @candidate alias
• CI tears down the feature branch (flows, assets, metadata)

On main:

• TrainFlow runs on a cron (0 8 * * 1-5 ET)
• EvaluateFlow auto-triggers after training
• ReportFlow prints the latest model

Local Development

# Train
python flows/train/flow.py run

# Evaluate (requires a registered model from TrainFlow)
python flows/evaluate/flow.py run

# Report
python flows/report/flow.py run

Deploying

Push to GitHub. The CI pipeline handles three lifecycle events:

1. Push — deploys flows to the platform (main gets cron, feature branches don't)
2. PR merge — promotes assets to main with @candidate, then tears down the feature branch
3. Branch delete — tears down branch resources