Awesome Search KG

Netflix

2 min read

Netflix

Global streaming platform with 260M+ subscribers. Search powers both external content discovery and internal asset management.

Search Context

Netflix’s content catalog is a federated graph — different teams own different entity types (titles, talent, studios, rights). Making this graph searchable is an unusually complex distributed systems problem.

Key Engineering Work

  • Content catalog represented as distributed graph (movies, shows, people, studios)
  • Entity embeddings capture semantic relationships from graph neighborhood
  • Federated query routing: sub-queries sent to appropriate partition owners; merged at orchestration layer
  • Enables similarity search across entity types (“movies similar to X”)

Internal: AMP Elasticsearch Indexing

  • AMP (Asset Management Platform) indexes media assets with high-frequency metadata updates
  • Alias-based blue-green indexing: build index_v2 → validate → atomically switch alias → delete index_v1
  • Zero-downtime reindexing via alias decoupling
  • Mapping design: field type choices (keyword vs. text, nested vs. flattened) affecting performance

Articles

Query Success Measurement: Click Residual

Walter Underwood developed the click residual metric (actual minus expected clicks per query) at Netflix. It flagged “batman” as an underperformer because The Dark Knight doesn’t contain the word, which an alias fixed. It also flagged “cj7”, a film not yet released in the US, which led the metadata team to add a catalogue entry. See Click Residual - A Query Success Metric.

Influence on Ranking Experimentation

Outside its own search stack, Netflix’s 2017 engineering write-up on interleaving for personalization algorithms is a frequently cited reference point for Interleaving as an online evaluation method. Booking.com cites those reported results as the reason it investigated interleaving for search ranking — see Beyond Algorithms - Ranking at Scale at Booking.com.

Graph View

Backlinks