Save your ML model from degradation with data logging
summary
WhyLogs is an open-source data-logging library that profiles 100% of inputs to catch drift and prevent model degradation, trading samples for statistical summaries; a benchmark shows sampling produces much higher median error while profiling often yields near-zero error. Profiles compress multi‑GB datasets down to ~14 MB, offer a one‑line pandas API, handle images and text, are mergeable for distributed and streaming systems, and integrate with Spark and MLflow.
AI-written summary · about
chapters (12)
- 0:00Introduction and overview
- 2:29ML workflow in production
- 4:34Why model monitoring matters
- 8:05Data monitoring vs model monitoring
- 10:13Logging for ML monitoring
- 13:03Sampling vs profiling data
- 16:31Introducing Y Logs
- 19:00Profiling vs sampling benchmark
- 21:47Lightweight data logging
- 23:11Mergeability in logging
- 24:34Y Logs integration and swag
- 26:42Q&A and closing
related talks
18mAI Observability: How To Fix Issues With Your ML Model
10mML-based Log Analysis for Faster Debugging
33mOperationalizing Model Lifecycle with MLFlow
43mMachine Learning Engineering principles with Python and MLFlow
01h 21mBuilding a ML Model Observability pipeline
16mMLOps — Understanding and Fighting Model Degradation in Production
31mRunning Machine Learning Analytics On Traces
26mTrust Fall: Three Hidden Gems in MLFlow