Profiling Python: From cProfile to Sampling
Your Python program is slow, so you run cProfile. It reports exact call counts and instrumented timings, but tracing every function call can change the workload you are trying to understand. Run a sampling profiler on the same program and a different picture emerges. Neither result is inherently wrong: each profiler answers a different question.
In this talk, we will compare cProfile, external sampling tools such as py-spy and austin, and Python 3.15's new standard-library sampling profiler, profiling.sampling, also known as Tachyon.
Using the same mixed workload throughout, we will see why call-heavy code is especially vulnerable to tracing overhead, how periodic stack samples become estimated hotspots, and what sampling can miss. We will then compare Tachyon's wall-clock and CPU modes to reveal an important distinction: time spent waiting can dominate one profile and disappear from the other.
We will also see how Tachyon can profile a script or attach to an existing Python process without modifying or restarting it.
You will leave with a practical framework: use tracing when you need exact call counts, sampling when you need representative hotspots or production diagnostics, and benchmarking when you need to compare small performance differences.
Python 3.15 gives developers two built-in profiling methodologies, but choosing between them requires understanding what each one measures. Tracing captures every call but can distort call-heavy workloads; sampling has much lower impact but produces statistical estimates and can miss short-lived functions.
The talk begins with one workload containing many small function calls, CPU-intensive work, and deliberate waiting. We first inspect it with cProfile, then use sampling to explain how repeated stack snapshots reveal hotspots. Finally, we profile it with Tachyon in wall-clock and CPU modes and compare the resulting views.
Attendees will leave able to:
- Explain the difference between deterministic tracing and statistical sampling
- Recognize when tracing overhead may distort a profile
- Understand what sampling can miss and why results vary between runs
- Use Tachyon to profile a script, attach to a process, and distinguish wall-clock time from CPU time
- Choose between tracing, sampling, and benchmarking for a performance investigation
By day, he works as a Python Software Developer at Ørsted, helping automate simulation workflows for power engineers. By curiosity, he enjoys exploring Python internals, performance tools, and the strange ways developers convince themselves their code is fast before measuring it.