Debugging C extensions#

Pandas uses select C extensions for high performance IO operations. In case you need to debug segfaults or general issues with those extensions, the following steps may be helpful.

First, be sure to compile the extensions with the appropriate flags to generate debug symbols and remove optimizations. This can be achieved as follows:

python setup.py build_ext --inplace -j4 --with-debugging-symbols

Using a debugger#

Assuming you are on a Unix-like operating system, you can use either lldb or gdb to debug. The choice between either is largely dependent on your compilation toolchain - typically you would use lldb if using clang and gdb if using gcc. For macOS users, please note that gcc is on modern systems an alias for clang, so if using Xcode you usually opt for lldb. Regardless of which debugger you choose, please refer to your operating systems instructions on how to install.

After installing a debugger you can create a script that hits the extension module you are looking to debug. For demonstration purposes, let’s assume you have a script called debug_testing.py with the following contents:

import pandas as pd

pd.DataFrame([[1, 2]]).to_json()

Place the debug_testing.py script in the project root and launch a Python process under your debugger. If using lldb:

lldb python

If using gdb:

gdb python

Before executing our script, let’s set a breakpoint in our JSON serializer in its entry function called objToJSON. The lldb syntax would look as follows:

breakpoint set --name objToJSON

Similarly for gdb:

break objToJSON

Note

You may get a warning that this breakpoint cannot be resolved in lldb. gdb may give a similar warning and prompt you to make the breakpoint on a future library load, which you should say yes to. This should only happen on the very first invocation as the module you wish to debug has not yet been loaded into memory.

Now go ahead and execute your script:

run <the_script>.py

Code execution will halt at the breakpoint defined or at the occurrence of any segfault. LLDB’s GDB to LLDB command map provides a listing of debugger command that you can execute using either debugger.

Another option to execute the entire test suite under lldb would be to run the following:

lldb -- python -m pytest

Or for gdb

gdb --args python -m pytest

Once the process launches, simply type run and the test suite will begin, stopping at any segmentation fault that may occur.

Improve debugger printing#

By default your debug will simply print the type and memory address of a PyObject. Assuming we passed a list containing ["a", "b"] as an argument to a Cython-generated function with parameter obj, debugging that object would look as follows:

(gdb) p __pyx_v_obj
$1 = (PyObject *) 0x5555558b91e0

Dereferencing this will yield the standard PyObject struct members of the object, which provides some more visibility

(gdb) p *__pyx_v_obj
$2 = {ob_refcnt = 1, ob_type = 0x5555558b91e0 <PyList_Type>}

If you are using gdb, CPython provides an extension that prints out more useful information about the object you are inspecting. The extension can be found in cpython/Tools/gdb/libpython.py; for best results be sure to use the gdb extension from the CPython branch that matches the version of your interpreter.

To activate the extension you will need to execute source <path_to_cpython_source>/Tools/gdb/libpython.py from an actively-running gdb session. After loading you will get more detailed information about the Python object you are inspecting.

(gdb) p __pyx_v_obj
$3 = ['a', 'b']

If you do not wish to explicitly source this file on every gdb run, you can alternately add it as a start up command to your gdbinit file.

Checking memory leaks with valgrind#

You can use Valgrind to check for and log memory leaks in extensions. For instance, to check for a memory leak in a test from the suite you can run:

PYTHONMALLOC=malloc valgrind --leak-check=yes --track-origins=yes --log-file=valgrind-log.txt python -m pytest <path_to_a_test>

Note that code execution under valgrind will take much longer than usual. While you can run valgrind against extensions compiled with any optimization level, it is suggested to have optimizations turned off from compiled extensions to reduce the amount of false positives. The --with-debugging-symbols flag passed during package setup will do this for you automatically.

Note

For best results, you should run use a Python installation configured with Valgrind support (–with-valgrind)