System Design

Two-Stage ML Recommendation Pipeline

System Design diagram · made in Skemio

System architecture diagram illustrating candidate retrieval reducing 10 million items to 500, followed by deep ranking and re-ranking producing top 20 items.
System architecture diagram illustrating candidate retrieval reducing 10 million items to 500, followed by deep ranking and re-ranking producing top 20 items.
Open it in SkemioMake your own

What it shows

  • Candidate retrieval: Approximate nearest neighbors drops 10M to 500 items.
  • Deep neural ranking: Heavy multi-task models rank 500 items in 30ms.
  • Business re-ranking: Deduplication and freshness rules yield final 20 items.

The takeaway

Recommendation architectures scale by combining fast candidate retrieval with precision deep ranking.

More system design diagrams