FastAPI plus Socket.IO gives async-friendly endpoints and persistent connections, but CPU-heavy scoring can still stall the loop if you let it run inline. We treat match scoring as a bounded worker path: vectorized NumPy/Pandas where possible, with clear batching.
Sub-200ms scoring is a product requirement when users expect instant feedback. That means profiling the actual hot functions, not guessing–then pushing allocation-heavy work off the critical WebSocket ack path.
PostgreSQL gets composite indexes aligned to real filters: location, preferences, activity–whatever the queries actually use in production. Connection pooling prevents thundering herds when thousands of sockets are active.
The architecture that survives is explicit: async IO for I/O, controlled parallelism for CPU, and the database indexed for the patterns we run at scale–not the ones we imagined in design docs.
