~80 → 400+
Concurrent players inside one synchronised live event
Real-time multiplayer scaling on GKE, sized from load tests instead of CPU
A host-led game failed exactly when the room joined. I found that more collaboration pods increased broadcast overhead, split the API from sockets, sized fewer larger pods from load tests, and scaled the frontend before the lobby moved into play.
Client details stay private. I can share the architecture, the constraints, and the reasoning.








