Feature Store

Definition

A feature store is the system that persists and serves the ranking features an LTR model consumes — BM25 components, popularity/CTR counters, item metadata (price, category), user/profile signals — keeping training-time and serving-time feature values consistent.

Why it matters for ranking

In a search system, ranking features are either:

• Stored in the index — fast at query time but requires frequent full reindexing when features change, or
• Served from an external feature store — decoupled from the index, updatable in real time, but adds a serving dependency and network hop.

This tradeoff is central to multi-stage / secondary re-ranking: an external re-ranker like Metarank keeps aggregated user/item feature state in Redis, pulled in a single batched request per re-rank (a key driver of re-ranking latency).

Related Concepts

• Learning to Rank · LambdaMART · Reranking · Retrieval Pipeline

• Click Signals · Personalization — common feature sources

• LTR Feature Engineering — which features to store and serve; the “what” to the Feature Store’s “how”

Articles

• Learn-to-Rank with OpenSearch and Metarank — feature store vs. in-index features tradeoff
• Metarank - Personalized Ranking That Actually Reads Your Clicks
• Part 1 - Learning to Rank for E-Commerce Search at Otto — Solr feature store