Design a recommendation feed

Prompt and requirements

Design the ranking behind a home feed of items (videos, posts or products) for 100 million daily users. Each request returns 50 items in under 200 milliseconds, chosen from a catalog of tens of millions. Clarify the objective first: clicks, watch time, purchases, or a blend with guardrails for quality and diversity. This is an original design exercise reviewed October 11, 2026.

Multi-stage architecture

Request (user, context)
  -> candidate generation: several sources, ~1,000s of items
       two-tower embedding retrieval (approximate nearest neighbours), co-occurrence, follows, trending
  -> lightweight ranker: ~500 items, cheap features
  -> heavy ranker: ~100 items, rich features, multi-task model (p(click), p(long watch), p(hide))
  -> re-ranking: business rules, diversity, freshness, deduplication, policy filters
  -> response + impression logging

Each stage trades recall for precision as the item count drops and the model cost rises.

Data and training

Log impressions with the features used at serving time, then join later engagement as labels. Training on features recomputed later causes training-serving skew; logging served features avoids it. Negative examples come from impressions without engagement, which are biased by what earlier models chose to show; mention exploration traffic or inverse propensity weighting.

Evaluation

Offline: recall@k for retrieval, AUC or log loss for rankers, on a time-based split so the model never trains on the future. Online: A/B tests on the business metric with guardrail metrics such as hides, reports and session length. Offline gains often fail to transfer, so the online test decides.

Cold start and freshness

New users: popularity in their region plus fast onboarding signals. New items: content embeddings so retrieval can find them before they have engagement, plus a small exploration budget.

Follow-up questions

• How do you avoid filter bubbles? (Diversity constraints in re-ranking and exploration.)
• How do you serve embeddings at this scale? (An ANN index such as HNSW or IVF, sharded and refreshed as items change.)