Under the Hood: How We Serve 3.5 Million Hours of Play Per Day
A deep dive into the engineering behind Poki's infrastructure, from edge caching to real-time game recommendations.
Serving 90 million players per month means our infrastructure needs to handle enormous scale while keeping latency low. In this post, we pull back the curtain on the systems that power Poki, from our global CDN setup to our machine learning recommendation engine.
Our stack runs primarily on Go microservices deployed across multiple cloud regions. Game assets are served through a multi-layer caching strategy that keeps 95% of requests within 50ms of the player. For real-time features like multiplayer matchmaking, we use WebSocket connections managed by a custom load balancer that routes players to the nearest server.
The recommendation engine is perhaps our most complex system. It processes billions of gameplay events daily to build player preference models. Using a combination of collaborative filtering and content-based approaches, it suggests games that match each player's taste while also surfacing new titles from indie developers. This system alone is responsible for over 30% of all game sessions on the platform.