Panel 02 - Data Quality & Feature Stores
summary
Panel lays out concrete checks and tools: three failure modes—training data quality, serving data quality, and online-offline consistency—with solutions like differential data validation and AWS Deequ for column statistics, KLL sketches for streaming quantiles, KS distance to detect prediction-distribution shifts, real-time streaming pipelines that compute distances and fire alerts, and embedding + clustering (UMAP) to visualize unstructured feature drift. Examples include side‑by‑side logged vs backfill comparisons and production alerts that pinpoint changed feature cardinality or cadence so teams can triage rapidly.
AI-written summary · about
chapters (9)
- 0:00Introduction to panel
- 1:26Defining data quality in feature stores
- 7:48Airbnb’s data quality approach
- 12:28Training data quality checks
- 15:59Arize’s ML observability platform
- 19:10Feature store vs general data quality
- 21:39Future of data quality in feature stores
- 25:08Unstructured data quality challenges
- 33:21Closing remarks and Q&A
related talks
21mLarge-scale Data Quality Verification (...)
05mData Quality and Validation in ML Pipelines
18mData quality: the key to long term happiness
37mData Validation at Scale: Managing Data Quality in Complex Data Pipelines
44mPreventing, Diagnosing & Curing Bad Data
28mHands-on Workshop: How To Perform Data Quality At Scale
53mData Mesh in Practice - Assuring Data Quality at Scale
30mBetter data quality with ML & visualization