Streaming & Online Machine Learning
Machine learning on unbounded data — the one-pass streaming model, sublinear sketches (reservoir sampling, Bloom filters, Flajolet-Martin, DGIM), online learning and regret, concept drift detection, and production streaming architectures.
6 lessons
01From Batch to Streaming
The data-stream model — unbounded input, one pass, bounded memory — and why event time, not processing time, is the axis that matters.
Streaming Algorithms & Sketches
Sublinear-memory summaries of a stream — reservoir sampling, Bloom filters, Flajolet-Martin distinct counts, and DGIM for counting ones in a sliding window.
Online Learning & Regret
Learning one example at a time — the online convex optimization setup, regret, online gradient descent and its O(√T) bound, FTRL, and online-to-batch conversion.
Concept Drift & Adaptation
When the stream's distribution changes — types of drift, detectors like ADWIN and DDM, adaptive windowing, and incremental learners such as Hoeffding trees.
Streaming ML in Production
Wiring it together — log-based ingestion, stream processors, online/offline feature parity, feature freshness, streaming inference, delayed labels, and the monitor-adapt loop.
Streaming & Online ML Quiz
Test your understanding of the streaming model, sketches, online learning and regret, concept drift detection, and production streaming architecture.