Hugging Face infrastructure engineer Arek Borucki explained on the AI Engineer podcast that the platform's search system broke under scale, as models grew from roughly 20,000 to 3 million — a 150x increase — and datasets rose from 10,000 in 2022 to 100,000 in 2024 to 1 million at the time of recording. The platform now has 14 million users, including more than 30% of the Fortune 500, and 50,000 organizations. Borucki noted that a MongoDB regex query scanning a search_tokens array worked for 20,000 models but "regex doesn't scale well," causing latency problems at millions. The fix moved computational cost from query time to insert time: when a model is uploaded, the Hub tokenizes its full ID into sub-tokens stored in a dedicated denormalized read collection. At query time, a MongoDB aggregation pipeline with Atlas Search uses the autocomplete operator against this precomputed array, sorting by trending_score. Borucki stated, "So far scale well. So we don't have any more latency issues in our search bar." He emphasized that "P99 is much more important than P50," calculating that with 14 million users, if 1% of search queries experience slow latency, that means 140,000 people affected per query.