From Data Silos to Feature Serving: How We Unified Fragmented Data for Training, Inference, and Reuse

Leena Bhandari, Andrea Cicalese, Durai Ramasubramanian10 Nov 2026

Problem

ML features are scattered across operational databases, warehouse tables, event streams and per-team pipelines. Each new model starts with finding the signals and writing one-off extraction code, and the next team repeats the work. Copying everything into one central store would only add another silo to keep in sync.

Example from the article: the feature deal_amount_changes_30d (how many times a deal's amount changed in 30 days) must be available for both training and real-time inference.

Solution in brief

A registry records where each feature comes from. Features are ingested into the store or read in place from the CRM, a team's service or a team's Iceberg table. Online values live in HBase, history lives in Iceberg as wide point-in-time rows, and a loader syncs the two.

Questions this case answers
  • How to give many teams one retrieval interface for features that live in different systems, without duplicating data or keeping copies in sync
  • How to store feature history so that a training row sees only values that existed at its timestamp, even when events arrive late
  • How to let many teams discover and reuse the same ML features instead of rebuilding them for every model?
  • How to serve features from operational systems without copying them into yet another store?

Tools and methods

  • Feature registry (MySQL/Vitess) — catalog of definitions, online and offline source, description, tags and owners for each feature.
  • Online store (Apache HBase) — latest value per feature for low-latency inference.
  • Offline store (Apache Iceberg on Amazon S3) — full feature history for training and batch work.
  • Apache Spark — streaming job that stages online updates; batch libraries for offline inference and backfills.
  • OpenMetadata — data catalog where features appear with definitions, ownership and lineage.
  • Point-in-time rows — one wide row per entity and moment, so training needs no joins across feature tables.
Continue reading — it's free

Google or a one-time email code. No password.