LinkedIn Professional Network  ·  View 25 of 30  ·  6 · Operations

Relevance ML Lifecycle

How feed, search and jobs ranking improve: a loop that closes through what members actually do.

Editable source SVG draw.io All views
Track impressions · clicks Features Frame · Feathr Train Spark · TonY Validate offline metrics Deploy Quasar · model repo Ramp T-REX A/B Observe health assurance Pro-ML one ML platform labelled events training sets candidate passes gate live at 1% guardrails hold drift signals Relevance ML Lifecycle Queue / topic Data store Application we own Decision point Security / platform v 1.0 · owner AI Platform · date 2026-09

Decisions

  • One platform, Pro-ML, for every ranking team: features defined once in Frame and Feathr, trained on Spark and TonY, executed by Quasar, ramped by T-REX (LinkedIn, 2019)
  • The loop closes through tracking: today's impressions and clicks are the next model's labels
  • Health assurance watches feature drift and the gap between offline and online behaviour

Open-source parts

  • TonY (TensorFlow on YARN, 2018) and Feathr (feature store, 2022) are open source
  • DARWIN is LinkedIn's JupyterHub-based data science workbench (2022)

Risks

  • A model trained on its own impressions narrows what members see, so exploration traffic is reserved
  • The feature store is derived data; it must be rebuildable from the lake
◀  Observability and Operations All views Graceful Degradation Contract  ▶
Solution Architecture v1.0 · LinkedIn's own stack in LinkedIn-run colos · Enterprise Architecture · 2026-0925-ml-lifecycle