📄 中文摘要
在机器学习的部署中，训练-服务偏差是一种常见的失败模式。数据科学团队在Python中构建模型，并在Jupyter Notebook中实现99%的AUC。随后，他们将逻辑交给后端工程团队，后者在生产API中用Java或Go重新实现特征计算。然而，模型在生产中失败，原因在于逻辑中的微小、不可见的差异。例如，在计算“平均交易价值”时，Python逻辑通过向前填充处理空值，而Java逻辑则将空值视为0。这种差异导致模型在生产环境中的表现与预期不符。
训练-服务偏差：无声的模型杀手
出处: Training-Serving Skew. The Silent Model Killer
发布: 2026年2月23日
📄 English Summary
Training-Serving Skew. The Silent Model Killer
Training-Serving Skew is a common failure mode observed in machine learning deployments. The Data Science team builds a model in Python, achieving 99% AUC in a Jupyter notebook. They then hand off the logic to the Backend Engineering team, who re-implement the feature calculations in Java or Go for the production API. However, the model fails in production due to small, invisible differences in logic. For instance, when calculating 'Average Transaction Value', the Python logic handles null values by forward-filling, while the Java logic treats them as 0. Such discrepancies lead to unexpected model performance in production environments.