Whatamix: Blendable feed construction

Grace Li, Crystal Lee, Jared Casale7 Nov 2024

Problem

Whatnot, a livestream shopping app, serves many feeds: For You, category feeds, feeds of followed sellers and more. Each needs different content, a different amount of personalization, and display modules such as shelves and banners. When teams build each feed alone, they duplicate code, call external services in inconsistent ways, and end up with disjointed logging and a growing maintenance load. Components were also built one after another, even when independent.

Example from the article: a For You feed needs an event banner and a popular shows shelf next to ranked livestream tiles, while the Ads team separately decides which ads go into the same feed and where.

Solution in brief

Whatamix describes every feed as a directed acyclic graph (DAG) of interchangeable nodes: retrieval, ranking, business logic and budgeting. The graph is built at request time from "feed params", experiment and config values looked up once per request, so A/B tests add or remove nodes without new lookup code. Independent branches run in parallel, and each node gets timeouts, error handling and logging by default.

Questions this case answers
  • How to compose recommendation pipelines from reusable stages so that retrieval, ranking and business rules can be swapped, parallelized and experimented on per request
  • How to let several teams run independent online experiments on different parts of one ranked list without interfering with each other
  • How to build many different feeds (personalized, category, followed, with banners and shelves) without duplicating retrieval, ranking and logging code for each one?
  • How to run several A/B tests on the same feed at once without lookups and code collisions between teams?

Tools and methods

  • DAG of nodes (Whatamix) — wraps each piece of feed logic, runs independent nodes in parallel
  • Feed params — per-request experiment values, configs and defaults
  • Statsig — A/B tests and dynamic configs vendor
  • Elasticsearch — one of the candidate retrieval sources
  • Sub-DAGs — team-owned parts of the graph with their own params slice
Continue reading — it's free

Google or a one-time email code. No password.