A model can degrade in production even when nothing about the model changes. The issue is often the data.

Upstream changes like renamed columns, shifted formats, or unexpected nulls don’t always break pipelines. The data still looks valid, but it’s no longer what the model expects. Predictions drift, and the problem often goes unnoticed until business metrics drop.

In practice, data issues cause more production incidents than model issues.

This chapter focuses on how to catch these problems early, from basic checks like schema validation to statistical tests that detect subtle distribution shifts.