Deployment Guide

CLI Automation

• Wrap training commands in a scheduler (cron, Airflow, GitHub Actions) using the same flags you run locally:

  python main.py --config config.yaml --subset-size 75000 --skip-evaluation
  python main.py --skip-training --evaluate
  

• Export PYTHONPATH=$(pwd) before invoking scheduled tasks to ensure src/ modules resolve correctly.
• Archive the generated run directory under models/run-*/ alongside the HTML evaluation report for traceability.

Promoting Models

• Successful runs copy metrics and artifacts into models/best_model/ unless you disable promotion.
• Use --force-update-best in automation only when you have upstream regression guards.
• Record the promoted commit hash, config, and metrics in your release notes (consider storing in logs/promotion_history.json).

Serving the API

• Run the packaged server directly:

  python api_server.py --port 8000
  

  or with Uvicorn for live reload in development:

  uvicorn api.main:app --host 0.0.0.0 --port 8000 --reload
  

• Containerise the service by copying requirements.txt, api/, src/, config.yaml, and models/best_model/. Expose port 8000 and mount logs/ for rotating access logs.

Monitoring & Logging

• Access CLI and trainer logs under logs/; rotate or export them to your observability stack.
• API responses include confidence scores—aggregate them to monitor prediction drift. Low average confidence suggests retraining or adjusting the subset size.
• Watch the data/review/unknown_*.csv folder for codes outside the known mapping; feed them back into the correction workflow.

Continuous Documentation

• Rebuild and publish the MkDocs site whenever documentation changes:

  pip install -r requirements-docs.txt
  mkdocs build
  

• Commit the contents of the site/ folder to a gh-pages branch or enable GitHub Pages with an Actions workflow similar to the ISCO pipeline repository.

Security Considerations

• Treat config.yaml as non-secret but avoid embedding credentials; use environment variables for buckets or databases.
• Sanitise datasets before copying models across networks; BCEA codes may fall under regulatory review.
• Review dependencies via pip install -r requirements.txt --require-hashes when locking down production builds.