Classic recommendation cascade

Classic recommendation cascade A data-flow diagram generated by Archify. 01 / Sources 02 / Retrieval 03 / Pre-ranking 04 / Ranking 05 / Re-rank & blend Content corpus · millions · 01 / Sources · indexed offline Content corpus millions indexed offline User history · recent actions · 01 / Sources User history recent actions Retrieval channels · ANN · CF · graph · 02 / Retrieval · parallel Retrieval channels ANN · CF · graph parallel Pre-ranker · light model · 03 / Pre-ranking Pre-ranker light model Feature store · hydrated features · 03 / Pre-ranking Feature store hydrated features Heavy ranker · P(action) model · 04 / Ranking · largest cost Heavy ranker P(action) model largest cost Ads auction · runs in parallel · 04 / Ranking Ads auction runs in parallel Re-ranker · diversity · rules · 05 / Re-rank & blend Re-ranker diversity · rules Blender · Top N to the feed · 05 / Re-rank & blend Blender Top N to the feed millions → ~10⁴ user embedding ~10⁴ → ~10³ hydrate candidates ~10³ → ~10² scored ~10² won ads ordered posts Legend primary data async batch data store data flow

Why layers

  • • Each cheap layer exists so the expensive one sees fewer items
  • • Cost is set by what reaches the heavy ranker, not by what enters

Latency budget

  • • Server-side ranking gets ~300 ms at P99
  • • The heavy ranker owns the largest share of it

What stays off the critical path

  • • Ads auction runs beside the organic pipeline
  • • Corpus indexing and feature computation happen offline