← AI Terminology
Data Drift
Data drift is a change in the input distribution seen by a deployed model compared to training or earlier production, which can degrade performance without code changes.
A core MLOps monitoring concern.
A core MLOps monitoring concern.
Why It Matters in AI
User language, product mix, and seasons shift. Models silently fail if drift is ignored. Monitoring features/embeddings and triggering retrain or RAG updates is production ML hygiene.
Key Points
| Aspect | Description |
|---|---|
| LLM | New slang, tools, policies; prompt/RAG refresh |
| Types | Covariate shift; related to concept drift |
| Detect | PSI, KS tests, embedding distance, perf proxies |
| Related | Concept drift, MLOps monitoring |
| Tooling | Evidently, WhyLabs, custom dashboards |
| Response | Retrain, adapt, fallback, collect labels |
Simple Analogy
A navigation model trained on summer roads struggling when winter snow changes how everything looks — the world moved.
Common Usage Examples
- Weekly PSI on key features
- Embedding cluster shifts in tickets
- Alert when retrieval hit rates drop
- Schedule refreshes for drifting domains
Summary
In short: Data drift is input distribution change in production — silent model decay that monitoring and refresh cycles must catch.