Tag

machine learning

2 cards

Cards

Data drift shown as two shifted distributions and concept drift as a correlation that falls apart between 2024 and 2025, with shadow mode, canary, full rollout and rollback as deployment stages, and accuracy decaying from 94% to 71%

Data PlatformDay 89Oct 29, 2026

The Model Didn't Break. The World Did.

system-design data-platform model-drift

The same feature defined twice — once by a data scientist against the warehouse and once by an engineer against the operational database — producing training/serving skew, with a feature store defining it once for both offline and online reads

Data PlatformDay 87Oct 27, 2026

The Model Was Trained On A Different World

system-design data-platform feature-store