A Decade of AI Platform at Pinterest (medium.com)

Pinterest’s senior ML platform leader reflects on a decade-long evolution from fragmented, team-specific ML stacks to a company-wide AI platform that now powers recommendation, ranking, and foundation/generative models. The platform serves hundreds of millions of inferences per second, evaluates thousands of models per user request in under 100 ms, and runs on thousands of GPUs across hybrid CPU/GPU clusters. The retrospective traces five eras — from DIY team stacks and the Linchpin DSL and Scorpion inference engine to a scrappy two-engineer platform team (EzFlow, Galaxy) and finally to a GPU/transformer-driven rebuild — showing how technical choices repeatedly collide with organizational incentives and industry timing. Key takeaways for the AI/ML community: unification works best as layered, bottom-up progress (feature DSL → unified serving → orchestration → feature store → training compute) and isn’t permanent — DNNs, GPUs, and LLMs force re-architecture. Practical technical milestones included Linchpin (single-source features/models), Scorpion (C++ high‑fanout scorer), EzFlow (code-first training orchestration with lineage hashing), Galaxy (modular signals/feature store seed), and a Training Compute Platform that moved from Kubernetes + TensorFlow to broader PyTorch/GPU scale. Adoption depended less on fixing pain points and more on alignment with product and exec priorities; efficiency bottlenecks now demand closer modeling–platform co-design as the next frontier.