Distribution Shift is the discrepancy between the data distribution during training and the distribution encountered during deployment — when $P_{test}(X, Y) eq P_{train}(X, Y)$, model performance degrades, sometimes catastrophically, because the model has not learned to handle the new data characteristics.

Types of Distribution Shift

• Covariate Shift: $P(X)$ changes but $P(Y|X)$ stays the same — input distribution changes.
• Label Shift: $P(Y)$ changes but $P(X|Y)$ stays the same — class proportions change.
• Concept Drift: $P(Y|X)$ changes — the relationship between inputs and outputs changes over time.
• Domain Shift: The data comes from a different domain (different fab, sensor, process recipe).

Why It Matters

• Silent Degradation: Models fail silently under distribution shift — accuracy drops without obvious errors.
• Semiconductor: Process drift, tool degradation, new products all cause distribution shift — models must handle it.
• Monitoring: Continuous monitoring for distribution shift is essential in production ML systems.

Distribution Shift is the world changed but the model didn't — performance degradation when deployment data differs from training data.