MLOps实践

Quick Reference

Critical Traps

Training-Serving Skew:

• - Preprocessing in notebook ≠ preprocessing in service → silent bugs
• Pandas in notebook → memory leaks in production (use native types)
• Feature store values at training time ≠ serving time without proper joins

GPU Memory:

• - requests.nvidia.com/gpu: 1 reserves ENTIRE GPU, not partial memory
• MIG/MPS sharing has real limitations (not plug-and-play)
• OOM on GPU kills pod with no useful logs

Model Versioning ≠ Code Versioning:

• - Model artifacts need separate versioning (MLflow, W&B, DVC)
• Training data version + preprocessing version + code version = reproducibility
• Rollback requires keeping old model versions deployable

Drift Detection Timing:

• - Retraining trigger isn't just "drift > threshold" → cost/benefit matters
• Delayed ground truth makes concept drift detection lag weeks
• Upstream data pipeline changes cause drift without model issues

Scope

This skill ONLY covers:

• - CI/CD pipelines for models
• Model serving and scaling
• Monitoring and drift detection
• Reproducibility practices
• GPU infrastructure patterns

Does NOT cover: ML algorithms, feature engineering, hyperparameter tuning.