Awesome Search KG

Metarank - Personalized Ranking That Actually Reads Your Clicks

2 min read

Metarank - Personalized Ranking That Actually Reads Your Clicks

Source: https://www.codeline.co/thoughts/repo-review/2025/metarank-personalized-ranking-that-actually-reads-your-clicks Author: Florian Narr Published: 2025-12-15

A code-level “repo review” of Metarank, the open-source real-time personalization / Learning to Rank service written in Scala.

Summary

Metarank sits between a search engine and its users: feed it events (clicks, purchases, impressions) and it reranks results or recommendations on the fly, learning per session. It occupies a rare middle ground between massive internal ranking platforms and thin single-model wrappers — shipping 20+ built-in feature extractors (CTR, user agent, referer, interaction counts, time decay) and handling the whole feedback loop (event ingestion → model training → serving) as a single stateless service runnable in Docker. YAML-driven config lets you reach a working LambdaMART ranker without writing Scala.

Architecture (from the code)

  • FeatureMapping — turns YAML feature/model definitions into instantiated feature extractors and predictors. Predictors include LambdaMARTPredictor, ShufflePredictor, TrendingPredictor, MFPredictor (collaborative filtering via ALS), and BertSemanticPredictor (neural search). Bridges Metarank features to the underlying ltrlib LTR library.
  • MetarankFlow — event-processing pipeline on fs2 streams + cats-effect. A TrainBuffer collects clickthrough data; ImpressionInject synthesizes negative signals (shown-but-not-clicked) — the feedback loop that makes the model learn.
  • FeatureValueFlow — Scaffeine cache with a custom ticker; per-feature refresh interval so high-frequency features (click counts) update every event while slow features skip recomputation.
  • RankApi — http4s POST /rank/{model}; ?explain=true exposes per-item feature values for debugging.

Rough edges (per author)

  • Custom feature extractors require extending BaseFeature in Scala + recompiling a JAR — no plugin/scripting layer.
  • Production state is Redis-only (Feature Store concern); in-memory store is demo-only.
  • Heavy JVM/Scala Docker image, non-instant startup; slowed development since ~September 2025.

Bottom line

The most complete open-source option for adding click-based personalization to an existing search system without building feature engineering, the feedback loop, and training from scratch — best for teams already comfortable with JVM services and Redis.

People

Graph View

Backlinks