An engineer building an eCommerce search feature set out to replicate Algolia's sub-50ms global speed using familiar tools. After four failed architectures, the key insight emerged: database location trumps compute location. Edge functions failed when queries still crossed oceans to a centralized database.

The breakthrough came with Turso's embedded replicas on Fly.io. By replicating the SQLite database across multiple regions, each VM could serve queries with near-instant local data access. This setup finally hit the sub-50ms target, proving that co-locating data with compute is critical for global performance.

The final stack uses Fly.io VMs, Turso's embedded replicas, and Bun for a performance boost. The author learned that modern SQLite, with extensions like FTS5 and vector support, can handle production search workloads effectively. Choosing the right tool for specific requirements beats forcing one solution to do everything.