Machine Learning System Design Interview #10 - The SOTA Trap
The trap almost every ML engineer falls into when designing LinkedIn-scale recommendation systems.
You’re in a Senior ML System Design interview at LinkedIn. The interviewer sets a specific trap:
“We have 500 million users in a social graph. We need a real-time model to recommend new connections. Design the architecture.”
90% of candidates walk right into the buzzword trap.
The reflex is immediate. “It’s a graph structure, so we need a 𝘎𝘳𝘢𝘱𝘩 𝘕𝘦𝘶𝘳𝘢𝘭 𝘕𝘦𝘵𝘸𝘰𝘳𝘬 (𝘎𝘕𝘕). I would implement 𝘎𝘳𝘢𝘱𝘩𝘚𝘈𝘎𝘌 or 𝘎𝘈𝘛 to aggregate neighbor features and capture the non-Euclidean topology.”
It feels like the “Textbook” answer. It’s theoretically perfect. It’s also wrong.
The interviewer looks at you and says:
