Design Spotify's homepage

Asked in: recsys system design at music / video / content companies.

The L4 candidate proposes one ranker. The L6 candidate decomposes the surface into shelves with distinct objectives and a meta-layer that decides what gets shown.

What’s distinctive about a homepage

It’s not one recommendation problem. A homepage like Spotify’s has many shelves (Made For You, Recently Played, New Releases, Daily Mix, Discover Weekly, Podcasts, etc.), each with a distinct purpose:

• Some surface what the user already loves (low novelty, high engagement).
• Some surface new things (high novelty, lower per-item engagement, long-term retention).
• Some are personalized; some are editorial.

The hardest design decision is how the shelves interact: which appear, in what order, with what content.

Learning objective

Trace why ranked item lists become candidates for a second decision that composes the whole homepage.

flowchart TB
  accTitle: Two ranking levels compose a multi-shelf recommendation homepage
  accDescr: At the first level, separate pipelines build or rank coherent shelves for familiar favorites, discovery, new releases, and editorial content. Each produces a completed shelf candidate with its own intent and local objective. At the second level, a page composer receives those completed shelf candidates plus user and session context. It selects which shelves appear and orders them while enforcing cross-shelf diversity, editorial commitments, and repetition constraints. The output is an ordered homepage of intact shelves. Evaluation keeps local shelf outcomes separate from page and long-term outcomes, so a locally strong shelf does not automatically dominate the page.

  subgraph LOCAL["LEVEL 1 · BUILD OR RANK EACH SHELF"]
    Familiar["FAMILIAR FAVORITES<br/>retrieve + rank items<br/>completion · repeat"]
    Discovery["DISCOVERY<br/>retrieve + rank items<br/>appeal · novelty"]
    Fresh["NEW RELEASES<br/>retrieve + rank items<br/>taste · freshness"]
    Editorial["EDITORIAL<br/>curated coherent list<br/>market · cultural context"]
  end

  Context["USER + SESSION CONTEXT<br/>history · device · time · recent activity"]
  Composer{"LEVEL 2 · COMPOSE THE PAGE<br/>select + order whole shelves<br/>cross-shelf diversity · repetition · editorial mix"}
  Homepage["ORDERED HOMEPAGE<br/>shelf 1 · shelf 2 · shelf 3 · …"]
  Evidence["KEEP TWO SCORECARDS<br/>local: shelf engagement<br/>global: session quality · return · satisfaction"]

  Familiar -->|"shelf candidate"| Composer
  Discovery -->|"shelf candidate"| Composer
  Fresh -->|"shelf candidate"| Composer
  Editorial -->|"shelf candidate"| Composer
  Context ==> Composer
  Composer ==> Homepage
  Homepage --> Evidence
  Evidence -.->|"example: improve a learned shelf"| Familiar
  Evidence -.->|"improve page policy"| Composer

  class Familiar,Discovery,Fresh,Editorial viz-input
  class Context viz-state
  class Composer viz-focus
  class Homepage,Evidence viz-output
  class Familiar viz-wide

Read it this way: read the top row as four complete recommendation products, not four item pools. Each emits a completed shelf with a distinct promise; editorial shelves may be curated rather than ranked. Only then does the page composer use session context to select and order those shelves under cross-shelf constraints. A shelf can win its local metric and still lose page position if it repeats another intent or harms longer-term outcomes. Original synthesis checked against Spotify's shelf-generation research and homepage calibration work, plus the multi-carousel evaluation study.

What an L5 answer sounds like

“Treat the homepage as two layers:

Layer 1: per-shelf rankers. Each shelf has its own retrieval and ranking pipeline tuned to its purpose:

• Made For You / Daily Mix: retrieve from user’s listening history embedding space, rank by predicted listen-completion + repeat-listen probability.
• New Releases: retrieve from new content (last 7-30 days) matched to user taste; rank balancing freshness with personalization.
• Discover Weekly: retrieve far from user’s history (exploration); rank by predicted appeal calibrated for novelty.
• Recently Played: simple recency.

Each shelf is its own multi-stage retrieve-then-rank pipeline.

Layer 2: shelf composition. A meta-ranker decides which shelves appear on the homepage and in what order, per user, per session. Inputs: user features, time of day, device, recent shelf engagement. Output: an ordered list of shelves to show.

Eval: each shelf has its own metrics (CTR, listen-time, completion). The composition layer optimizes a session-level metric (total session time, return rate, satisfaction). A/B test composition changes against a fixed baseline shelf order.”

This is L5. Two layers, each shelf as its own ranker, meta-layer for composition.

What an L6 answer adds

“…practical things:

Diversity is critical. Without explicit diversity, the homepage collapses to similar items everywhere. Explicit diversification at three levels: within-shelf (don’t show 10 songs from the same artist), across-shelf (different shelves should serve different intents), and over-time (don’t show the same item every day).

Editorial vs algorithmic mix. Pure algorithmic recsys is fragile (filter bubbles, missed cultural moments). Editorial shelves (Today’s Top Hits, regional charts, curated playlists) anchor the experience. The meta-ranker decides the mix.

Cold-start handling. New users see fewer personalized shelves and more popularity-based / editorial. Transition to personalized as listen history accumulates. New items get a freshness boost in candidate generation; their long-term ranking depends on accumulated engagement.

Multi-task ranking per shelf. Don’t predict ‘click probability’ alone. Predict listen-time, completion, save, share, return-next-day. Combine with weights tuned via offline counterfactual eval and confirmed in A/B.

Latency budget: homepage loads need ~200ms p99. Cache aggressively per user (re-compute every few hours, not per request). For real-time signals (just-finished song), update only the affected shelves.

Long-term metrics gate launches. Short-term engagement is easy to optimize and easy to cannibalize the long-term experience. Track session-quality, return rate, churn proxies; require long-window A/B (4-8 weeks) for any composition change.”

Tells that get you a strong-hire vote

• You decompose into shelves with distinct objectives.
• You add a meta-ranker for composition.
• You bring up editorial / algorithmic mix.
• You discuss diversity at multiple levels.
• You insist on long-window A/B tests for composition changes.

Tells that get you down-leveled

• One ranker for the whole homepage.
• No mention of diversity or editorial content.
• No cold-start handling.
• Optimizing only short-term clicks.

Common follow-up

“How would you order the shelves for a brand-new user?”

The L6 answer:

“For a new user with no history, the meta-ranker has no personal signal, so it falls back on three things: (1) editorial defaults (top playlists, popular new releases), (2) onboarding signal if any (genre selections from sign-up flow), (3) demographic and contextual proxies (region, device, time of day). The order should make engagement easy: hits and recognized content first, exploration second, personalized last (since the personalization will be weak). After a few sessions, transition to the personalized rank.”

---

Related: Design YouTube’s recommender, Two-tower vs cross-encoder: when to use which?, System design case study: building personalized search ranking.