Скачать или смотреть GopherCon 2025: Supercharging ML Pipelines with Go - Vaidehi Thete

Building scalable and efficient machine learning pipelines often requires overcoming bottlenecks in data transfer, feature retrieval, and orchestration. This talk showcases how Go was leveraged to operationalize an ML model service, transforming it into a high-performance, real-time system. By utilizing Go’s powerful concurrency model, shared memory for inter-process communication, and efficient queuing mechanisms, we reduced inference times from hours to just 10-15 minutes.

We’ll explore the architecture that decouples Go’s operational responsibilities—such as feature retrieval, queuing, and shared memory management—from the ML model’s ranking tasks. Attendees will learn how shared memory was used to transfer millions of scores efficiently, how Go’s interfaces enabled rapid prototyping, and how its strong typing and safety ensured robust system design. Real-world benchmarks, implementation details, and lessons learned will provide actionable insights for engineers tackling similar challenges in high-performance computing and distributed systems.

Whether you’re a software engineer, system architect, or distributed systems enthusiast, this talk will demonstrate how Go can be the backbone of scalable and efficient ML inference pipelines.