1 unstable release
|0.1.4||Sep 5, 2018|
2 downloads per month
Py-Spy is a sampling profiler for Python programs. It lets you visualize what your Python program is spending time on without restarting the program or modifying the code in any way. Py-Spy is extremely low overhead: it is written in Rust for speed and doesn't run in the same process as the profiled Python program, nor does it interrupt the running program in any way. This means Py-Spy is safe to use against production Python code.
Py-Spy works on Linux, OSX and Windows, and supports profiling all recent versions of the CPython interpreter (versions 2.3-2.7 and 3.3-3.7).
Prebuilt binary wheels can be installed from PyPI with:
pip install py-spy
If you're a Rust user, py-spy can also be installed with:
cargo install py-spy
py-spy works from the command line and takes either the PID of the program you want to sample from or the command line of the python program you want to run:
py-spy --pid 12345 # OR py-spy -- python myprogram.py
The default visualization is a top-like live view of your python program:
There is also support for generating flame graphs from the running process:
py-spy --flame profile.svg --pid 12345 # OR py-spy --flame profile.svg -- python myprogram.py
Which will generate a SVG file looking like:
It also possible to dump out the current call stack for each thread by passing
--dump to the command line.
This project aims to let you profile and debug any running Python program, even if the program is serving production traffic.
While there are many other python profiling projects, almost all of them require modifying the profiled program in some way. Usually, the profiling code runs inside of the target python process, which will slow down and change how the program operates. This means it's not generally safe to use these profilers for debugging issues in production services since they will usually have a noticeable impact on performance. The only other Python profiler that runs totally in a separate process is pyflame, which profiles remote python processes by using the ptrace system call. While pyflame is a great project, it doesn't support Python 3.7 yet and doesn't work on OSX or Windows.
Figuring out the call stack of the Python program is done by looking at the global PyInterpreterState variable to get all the Python threads running in the interpreter, and then iterating over each PyFrameObject in each thread to get the call stack. Since the Python ABI changes between versions, we use rust's bindgen to generate different rust structures for each Python interpreter class we care about and use these generated structs to figure out the memory layout in the Python program.
Getting the memory address of the Python Interpreter can be a little tricky due to Address Space Layout Randomization. If the target python interpreter ships
with symbols it is pretty easy to figure out the memory address of the interpreter by dereferencing the
_PyRuntime variables depending on the Python version. However, many Python
versions are shipped with either stripped binaries or shipped without the corresponding PDB symbol files on Windows. In
these cases we scan through the BSS section for addresses that look like they may point to a valid PyInterpreterState
and check if the layout of that address is what we expect.
Since we're getting the call stacks of the python program by looking at the PyInterpreterState we don't yet get information about non-python threads and can't profile native extensions like those written in languages like Cython or C++. Native code will instead show up as spending time in the line of Python that calls the native function, rather than as its own entry right now.
Py-spy works by reading memory from a different python process, and this might not be allowed for security reasons depending on your OS and system settings. In many cases, running as a root user (with sudo or similar) gets around these security restrictions. OSX always requires running as root, but on Linux it depends on how you are launching py-spy and the system security settings.
On Linux the default configuration is to require root permissions when attaching to a process that isn't a child.
For py-spy this means you can profile without root access by getting py-spy to create the process (
py-spy -- python myprogram.py) but attaching to an existing process by specifying a PID will usually require root (
sudo py-spy -pid 123456).
You can remove this restriction on linux by setting the ptrace_scope sysctl variable.
OSX has a feature called System Integrity Protection that prevents even the root user from reading memory from any binary located in /usr/bin. Unfortunately, this includes the python interpreter that ships with OSX.
There are a couple of different ways to deal with this:
- You can install a different Python distribution (you probably want to migrate away from python2 anyways =)
- You can use virtualenv to run the system python in an environment where SIP doesn't apply.
- You can disable System Integrity Protection.
Running py-spy inside of a docker container will also usually bring up a permissions denied error even when running as root.
This error is caused by docker restricting the process_vm_readv system call we are using. This can be overridden by setting
--cap-add SYS_PTRACE when starting the docker container.
SYS_PTRACE to be able to read process memory. Kubernetes drops that capability by default, resulting in the error
Permission Denied: Try running again with elevated permissions by going 'sudo env "PATH=$PATH" !!'
The recommended way to deal with this is to edit the spec and add that capability. For a deployment, this is done by adding this to
securityContext: capabilities: add: - SYS_PTRACE
More details on this here: https://kubernetes.io/docs/tasks/configure-pod-container/security-context/#set-capabilities-for-a-container Note that this will remove the existing pods and create those again.
Py-spy doesn't pause the target python program we are sampling from, and reads the interpreter state from the python process as it is running. Since the calls we use to read memory from are not atomic, and we have to issue multiple calls to get a stack trace this means that occasionally we get errors when sampling. If you run the flame graph code it will report the number of errors at the end (and the top view will report the error rate if it goes beyond 1%). This happens infrequently enough that it shouldn't affect the output that much though. Julia Evans wrote a great blog post talking about this trade off with rbspy
Ok, so no-one has ever actually asked me this - but I wanted to share since it's a pretty terrible hack that might be useful to other people.
I really wanted to distribute this package over PyPI, since installing with pip will make this much easier for most Python programmers to get installed on their system. Unfortunately, installing executables as python scripts isn't something that setuptools supports.
To get around this I'm using setuptools_rust package to build the py-spy binary, and then overriding the distutils install command to copy the built binary into the python scripts folder. By doing this with prebuilt wheels for supported platforms means that we can install py-spy with pip, and not require a Rust compiler on the machine that this is being installed onto.
Does this run on BSD? Support 32-bit Windows? Integrate with PyPy? Work with USC-16 versions of Python2?
Not yet =).
py-spy is heavily inspired by Julia Evans excellent work on rbspy. In particular, the code to generate the flamegraphs is taken directly from rbspy, and this project uses the (read-process-memory and proc-maps) crates that were spun off from rbspy.
Py-spy is released under the GNU General Public License v3.0, see LICENSE file for the full text.
- macos libproc 0.3.1
- unix termios 0.2.2
+winbase +consoleapi +wincon +handleapi +timeapi +processenv
- benfred-read-process-memory 0.1.3
- clap 2.0
- console 0.6.1
- env_logger 0.5
- failure 0.1.1
- goblin 0.0.15
- indicatif 0.9
- lazy_static 1.1
- libc 0.2.34
- log 0.4
- proc-maps 0.1
- regex 1.0
- tempdir 0.3
- tempfile 3.0