
Talk Python To Me #565: Tachyon, Python 3.15's Built-in Sampling Profiler
44 snips
Oct 2, 2026 László Kiss Kollár, a Python infrastructure engineer focused on tooling and performance, and Pablo Galindo Salgado, a CPython core developer and Steering Council member, explore Tachyon, Python 3.15’s built-in sampling profiler. They discuss why tracing can distort results, safer live-process debugging, profiling async and production workloads, flame graphs, and the journey from a tiny prototype to extreme sampling speeds.
AI Snips
Chapters
Transcript
Episode notes
Profiling Exposed A Simple Data Structure Mistake
- Michael Kennedy expected complex mathematics to cause a slowdown, but profiling revealed an inefficient data structure instead.
- Replacing operations such as list searches with set or dictionary lookups solved the problem without rewriting thousands of lines.
Unused Imports Can Slow The Whole Application
- Eager imports can impose substantial memory and garbage-collection costs even when the imported functionality is never used.
- Lazy imports avoid creating unused classes, functions, and constants, keeping long-running applications more linear under garbage collection.
Python Speedups Now Require Better Measurement
- Faster CPython delivered major gains, but the remaining optimizations are harder because JIT work, free threading, and interpreter stability interact.
- Profiling becomes essential for verifying whether a speedup came from CPython, NumPy, pandas, or another dependency.


