LinkedIn Professional Network  ·  View 16 of 30  ·  5 · Runtime

Feed Ranking Pipeline

Candidates, filter, score, policy, serve: the ranking is kept separate from feed storage, as the requirement asks.

Editable source SVG draw.io All views
Candidates Inbox fan-out on write FollowFeed first-pass ranked Recommended Sponsored Filter Privacy + blocks Moderation state Seen dedupe Score Features Venice · Feathr Second-pass ranker Quasar model Chronological degraded mode Policy Diversity rules Ad slotting Serve Decorate URNs Feed page p95 ≤ 300 ms ranker down Feed Ranking Pipeline Data store Application we own External / third party Security / platform Decision point failure / alternate Ranking reads candidates and features; it never writes to feed storage, so either can change alone. v 1.0 · owner Feed Relevance · date 2026-09

Decisions

  • Privacy, block and moderation filters run before ranking, so the model can never rank what the member may not see
  • Features come from Venice through Feathr and are defined once, for training and for serving
  • Diversity and ad positions are a rules layer after the model, so a policy change does not need a retrained model

Degraded modes

  • Ranker down: a chronological network feed plus trending, and the page still renders
  • Recommendations slow: network content only
  • Features stale: the last good Venice version is served

Risks

  • Ranking for engagement can reward outrage. The policy layer and moderation state are the counterweight
  • Offline and online features can drift apart; health assurance watches for it (view 25)
◀  Post Distribution — Hybrid Fan-out All views Connection Request — Idempotent Flow  ▶
Solution Architecture v1.0 · LinkedIn's own stack in LinkedIn-run colos · Enterprise Architecture · 2026-0916-feed-ranking-pipeline