Pyflame: A Ptracing Profiler For Python
Pyflame is a high performance profiling tool that generates flame graphs for Python. Pyflame is implemented in C++, and uses the Linux ptrace(2) system call to collect profiling information. It can take snapshots of the Python call stack without explicit instrumentation, meaning you can profile a program without modifying its source code. Pyflame is capable of profiling embedded Python interpreters like uWSGI. It fully supports profiling multi-threaded Python programs.
Pyflame usually introduces significantly less overhead than the builtin
profile
(or cProfile
) modules, and emits richer profiling data. The
profiling overhead is low enough that you can use it to profile live processes
in production.
Full Documentation: https://pyflame.readthedocs.io
Quickstart
Building And Installing
For Debian/Ubuntu, install the following:
# Install build dependencies on Debian or Ubuntu.
sudo apt-get install autoconf automake autotools-dev g++ pkg-config python-dev python3-dev libtool make
Once you have the build dependencies installed:
./autogen.sh
./configure
make
The make
command will produce an executable at src/pyflame
that you can run
and use.
Optionally, if you have virtualenv
installed, you can test the executable you
produced using make check
.
Using Pyflame
The full documentation for using Pyflame is here. But here's a quick guide:
# Attach to PID 12345 and profile it for 1 second
pyflame -p 12345
# Attach to PID 768 and profile it for 5 seconds, sampling every 0.01 seconds
pyflame -s 5 -r 0.01 -p 768
# Run py.test against tests/, emitting sample data to prof.txt
pyflame -o prof.txt -t py.test tests/
In all of these cases you will get flame graph data on stdout (or to a file if
you used -o
). This data is in the format expected by flamegraph.pl
, which
you can find here.
FAQ
The full FAQ is here.
What's The Deal With (idle) Time?
Full
answer
here.
tl;dr: use the -x
flag to suppress (idle) output.
What About These Ptrace Errors?
See here.
How Do I Profile Threaded Applications?
Use the --threads
option.
Is There A Way To Just Dump Stack Traces?
Yes, use the -d
option.