STANDARDwalkthrough
Feed Ranking
Should the newest post always appear first, or should the most relevant one? The constraint: a purely chronological feed buries high-quality content under a flood of low-engagement posts.
We need a ranking layer that surfaces relevant content without losing recency. Our ranking model scores every candidate post using engagement signals: likes (weight 1x), reposts (2x), and replies (1.5x), combined with a recency decay function that reduces the score as the post ages.
“Users following 500 accounts miss 95% of content they would engage with because it scrolls past before they open the app.”
A post with 500 likes and 100 reposts from 2 hours ago can outscore a zero-engagement post from 5 minutes ago. We chose ML-based ranking (not pure chronological) because engagement-weighted feeds increase time-on-site by 20-30%, which directly maps to our product requirement of maximizing content discovery.
Trade-off: we gave up predictability (users cannot reason about feed order) in exchange for relevance. The ranking model runs on dedicated scoring servers that evaluate hundreds of features per candidate in under 100ms.
We keep a chronological tab as an escape valve for users who want deterministic ordering. Facebook's ranking considers over 10,000 signals per post, including relationship strength, content type, and historical engagement patterns.
TikTok takes this further: their entire feed is algorithmically generated with no follow graph required, using watch time as the primary signal. Implication: ranking means we need a feature store that pre-computes engagement metrics and user affinity scores, adding a new infrastructure component to our architecture.
What if the interviewer asks: 'How do we avoid filter bubbles?' We inject a diversity factor that ensures 10-15% of feed content comes from accounts or topics the user has not engaged with recently, trading optimal engagement scores for content breadth.
Related concepts