Larkspur Video asks for a personalized home feed for 8,640,000 daily users and 40 million videos. Deliver the scope, a peak load estimate, a candidate table, the batch, streaming and serving paths, a ranking fallback, a retrieval approach and an exploration policy.
How interviewers level this design
- Junior: separates offline candidate generation from online ranking and keeps a fallback.
- Mid-level: sizes the request path, adds a streaming path for freshness and explains two-stage retrieval.
- Senior: designs exploration, measures feedback loops and plans for training and serving feature consistency.
Larkspur Video builds a personalized home feed for 8.6 million daily users. Scoring every video in the catalog for every visit is far too slow, so feeds are built in two stages: cheap candidate generation narrows millions of videos to a few hundred, and a heavier ranking model orders those few hundred within the request budget.
Candidates come from a nightly batch job and from a streaming job that adds fresh uploads within the hour. You will estimate feed traffic, define the candidate table, draw the batch, streaming and serving paths, add a fallback for when ranking times out, choose the two-stage approach and keep the feed from only showing what is already popular.
Narrow candidates cheaply, rank a few hundred within budget, keep a fallback and leave room for exploration.