Profiling in CPython 3.15: Low-Overhead Tachyon Profiler for Real Python Systems

Keynote
Shanghai
11:30 a.m. - 12:10 p.m.
  • Petr Andreev Russian Professor · CPython Contribution Team Lead · PSF Board of Directors Nominee

    Specializes in CPython internals, optimization, and high-performance computing.

    Driven by GPU acceleration, CPU vectorization. Evolved from ML systems to CPython core research engineer.

    8+ years leading teams in AI, maths, and physics. PyCon speaker.

    Lecturer at Moscow Institute of Physics and Technology – top 1 Russian university.

    Open to talks and collaboration.

    anonymous

Abstract

CPython 3.15 brings a clearer profiling story with Tachyon Profiler — a serious sampling profiler designed for finding real bottlenecks with low overhead.

Details

This talk focuses only on profiling: how Python developers should measure performance, why traditional tracing profilers are not always enough, and how Tachyon changes the debugging workflow for production-scale systems. We'll go deep on: sampling vs tracing — when each model works and where each one lies; Tachyon Profiler internals — how low-overhead sampling works conceptually; production debugging — how to profile services without destroying the signal; async, threads, and free-threading — why modern Python profiling is harder than "find the slow function"; and profiler interpretation — how to read results without chasing fake bottlenecks.

This talk will be interactive: we'll walk through several real profiling outputs live, and invite the audience to spot the likely bottlenecks and guess which code paths are the actual "usual suspects" eating up runtime. Beyond just showing how to use Tachyon Profiler, this talk aims to help developers build a performance engineer's mindset — learning to locate performance issues efficiently and accurately in complex Python systems.