Quiet Failures, Loud Consequences: Streaming ML Drift Detection in Practice

A machine learning model in production is like a ship sailing blind, everything looks fine until it slams into a reef. And by then, it's too late.

This phenomenon, known as concept and model drift, is especially dangerous in real-time systems where decisions happen in milliseconds and rollback is usually not an option.

If not detected early, drift doesn’t just break your models — it misprices loans, misses fraud, and even risks lives.

This talk distills cutting-edge research and real production lessons into practical tools that can be apply today, even if the models are already in the wild. Based on ongoing PhD research and real-world implementations, we’ll walk through the following real live questions:

- How drift manifests in event-driven ML systems — and why traditional batch monitoring fails.

- Common algorithms for drift detection (i.e. DDM, EDDM, ADWIN, Page-Hinkley) — and how to benchmark them in streaming environments.

- An architecture for integrating drift-aware intelligence into Flink pipelines, with hooks for alerting, model retraining, or failover strategies.

- Lessons from production use cases, including trade-offs in detection latency, false positives, and system overhead.

Whether you're deploying ML models into dynamic data streams or just planning your streaming AI strategy, you'll leave with a blueprint for building drift-resilient ML pipelines — plus hands-on knowledge to detect, benchmark, and respond to drift before it becomes failure.