Business Context

Autodesk’s internal data platform supports analytics and ML workloads for products such as Fusion 360. The data engineering team wants a model that predicts whether a scheduled pipeline run will fail before execution completes, so operators can prioritize intervention and reduce downstream SLA misses.

Dataset

You are given historical metadata for 420,000 pipeline runs collected from Autodesk Data Platform orchestration logs over 18 months. Each row represents one scheduled job run, with the target indicating whether the run failed within the execution window.

• Target: failed_run (1 if the pipeline run failed, 0 otherwise)
• Class balance: roughly 8.7% failures and 91.3% successful runs
• Missing data: ~12% missing in runtime features for new or rarely scheduled jobs; ~6% missing in data quality metrics due to disabled checks

Success Criteria

A good solution should identify likely failures early enough to support alert routing and triage. Aim for PR-AUC above 0.45, recall above 75% at precision above 35%, and clear feature-level explanations that a data engineering team can act on.

Constraints

• Nightly batch scoring on ~25,000 upcoming runs
• Inference latency should remain under 10 minutes for the full batch
• Model output must be interpretable enough for pipeline owners
• Retraining should be feasible on a monthly cadence without specialized GPU infrastructure

Deliverables

1. Build a binary classification model for failed_run.
2. Explain preprocessing choices for mixed feature types and missing data.
3. Design features that capture instability, drift, and dependency risk.
4. Evaluate the model using metrics appropriate for imbalanced classification.
5. Describe how you would deploy and monitor the model in production on Autodesk’s data platform.