Internet-Scale Profiling Without the Headaches

Pinterest collaborated with Intel to significantly enhance gProfiler, an open-source continuous profiling tool, to handle the immense scale of Pinterest's infrastructure. Pinterest runs hundreds of thousands of instances for its 640 million users across multiple programming languages. The core problem was fragmentation: engineers were stuck using separate, specialized profilers for each language (like perf for Linux, async-profiler for Java), making it nearly impossible to get a unified view of performance across their entire system.

The solution was to extend gProfiler as an orchestration layer that automatically discovers processes, runs the right language-specific profiler underneath, and merges the results into a single, comprehensive flame graph. This gives engineers a complete picture of where an application spends its time, from the high-level code down through libraries and into the Linux kernel. The real-world impact is turning profiling from a manual, expert-only debugging task into a routine operational capability that can be triggered remotely for any service, saving massive engineering time and uncovering efficiency gains that translate directly to cost savings at Pinterest's scale.