If your project grows beyond ~100k vectors or requires high concurrency, plan for a migration early. When it makes sense: you want to move fast, your dataset fits comfortably in memory/local disk, and ...