Awesome Search KG

Optimizing Vespa Latency with Match-Features at Vinted

2 min read

Optimizing Vespa Latency with Match-Features at Vinted

Dainius Jocas, Vinted Engineering, 2025-11-06.

How Vinted used Vespa’s match-features to both serve ranking features in-engine (instead of round-tripping to Redis) and, more surprisingly, to cut query latency by eliminating the document-summary fetch.

What match-features are

Rank features attached to each hit during the matching phase, returned without a separate document-summary fetch. Introduced in Vespa 2021. They carry floating-point numbers and tensors (not strings/booleans natively); string labels are encoded as mapped tensors via tensorFromLabels(attribute(my_feature)). Declared in a rank-profile:

match-features {
    my_feature
}

Use 1 — Vespa as the feature store

Vinted encodes additional ranking signals that don’t live in the document itself — cross-user interactions, category statistics, contextual signals — and computes/returns them via match-features. This makes Vespa both the storage and computation layer, removing the network round-trip to Redis for online features. (Compare Feature Store: this is in-engine feature serving rather than an external store.)

Use 2 — Latency optimization (the surprising win)

The deeper result: during heavy data redistribution (500M+ updates/hour across 50+ schemas), the document-summary .fill() step — which fetches summaries after matching — became unreliable, with latency spikes >100 ms when documents moved between content nodes mid-query.

By selecting only match-features from a schema (Vespa 8.596.7+), Vinted eliminated the summary fetch entirely, collapsing query execution from a two-phase scatter-gather into a single round-trip.

Results

  • P99 latency: ~9 ms → ~3 ms
  • Redistribution-related latency spikes vanished
  • Mean latency stabilized at ~430 µs at 7.5k RPS per container

Why It Matters

A concrete, non-obvious lesson: the document-summary fetch is a hidden second phase, and pushing the values you actually need into match-features can both replace an external feature store and remove a whole network phase from the query path. Directly relevant to the feature-logging step of Vespa Learning to Rank — the same mechanism that dumps training features also serves them online cheaply.

Vespa Learning to Rank · Vespa · Feature Store · Reranking · Retrieval Pipeline

People

  • Dainius Jocas — Vinted; author (also co-authored the recommendation-retrieval post)

Graph View

Backlinks