Stop Failures Before They Happen: Gradient Boosting
summary
Predicts SLA breaches after the first task using a lightweight 11-feature vector and runs inference inside Prefect UI to attach a risk score that can abort or reroute runs. Trains on ~5,000 synthetic runs, benchmarks models, and reports CatBoost + SHAP as the best combo with ~0.87 ROC AUC and per-run drivers. Demo shows automatic stop of high-risk runs and logs top features; rollout guidance covers shadow mode, code-version features, drift monitoring and safe gating.
AI-written summary · about
chapters (10)
- 0:00The Pain: Long Workflows That Fail Late
- 0:53Why Prefect Is the Perfect Place for Early Signals
- 2:19Defining Failure as an SLA Breach (and What We Can Do About It)
- 3:23Case Study: Portfolio Risk Flow & Early Feature Signals
- 5:54Modeling Approach: CatBoost, Baselines, and Evaluation
- 8:04Explaining Risk: SHAP Drivers Developers Can Act On
- 9:16Live Demo in Prefect UI: Low-Risk Run vs High-Risk Run
- 12:44From Demo to Production: Code Changes, Safe Rollout, and Drift Monitoring
- 14:22Beyond SLAs: Rerouting, Smarter Retries, and the Big Takeaway
- 15:44Wrap-Up and Thanks
related talks
01h 28mMastering gradient boosting with CatBoost
01h 27mCatBoost - The New Generation of Gradient Boosting
26mIntroducing ModelOps: How to deploy and manage every type of model
35mSHAPtivating Insights: unravelling blackbox AI models
39mMachine Learning in Financial Credit Risk Assessment
19mModel Performance Monitoring - A Practitioner's Perspective
29mOpeningTheBlackBox
37mCreating correct and capable classifiers