Preamble:
Python setuptools are used for the package distribution. I have a Python package (let us call it my_package
), that has several extra_require
packages to it. Everything works just find (installation and build of the package, as well as extras, if were requested), as all extra_require
were python packages themselves and pip correctly resolved everything. A simple pip install my_package
worked like a charm.
Setup:
Now, for one of the extras (let us call it extra1
) I need to call a binary of a non-python library X
.
Module X
itself (source code) was added to the my_package
codebase and was included in the distribution my_package
. Sadly for me, to be utilized, X
needs to be compiled first into a binary on the target machine (C++ implementation; I assume such compilation shall happen on the build stage of my_package
installation). There is a Makefile
in the X
library optimized for different platform compilation, so all that is needed, is to run make
in the respective directory of X
library in the my_package
when the build process is running.
Question #1: how to run a terminal command (i.e., make
in my case) during the build process of the package, using setuptools/distutils?
Question #2: how to ensure, that such terminal command is executed only if the corresponding extra1
is specified during the installation process?
Example:
- If someone runs
pip install my_package
, no such additional compilation of library X
shall happen.
- If someone runs
pip install my_package [extra1]
, module X
needs to be compiled, so the corresponding binary would be created and available on the target machine.
This question came back to haunt me long after I commented on it two years ago! I had almost the same problem myself recently, and I found the documentation VERY scarce, as I think most of you must have experienced. So I tried to research a bit of the source code of setuptools and distutils to see if I could find a more or less standard approach to both the questions you asked.
The first question you asked
Question #1: how to run a terminal command (i.e., make
in my case) during the build process of the package, using setuptools/distutils?
has many approaches and all of them involve setting a cmdclass
when calling setup
. The parameter cmdclass
of setup
must be a mapping between command names that will execute depending on the build or install needs of the distribution, and classes that inherit from distutils.cmd.Command
base class (as a side note, the setuptools.command.Command
class is derived from distutils
' Command
class so you can derive directly from setuptools
implementation).
The cmdclass
allows you to define any command name, like what ayoon did and then execute it specifically when calling python setup.py --install-option="customcommand"
from the command line. The problem with this, is that it is not the standard command that will be executed when trying to install a package through pip
or by calling python setup.py install
. The standard way to approach this is to check what commands will setup
try to execute in a normal install and then overload that particular cmdclass
.
From looking into setuptools.setup
and distutils.setup
, setup
will run the commands it found in the command line, which lets assume is just a plain install
. In the case of setuptools.setup
, this will trigger a series of tests that will see whether to resort to a simple call to the distutils.install
command class, and if this does not occur, it will attempt to run bdist_egg
. In turn, this command does many things but crucially decides on whether to call the build_clib
, build_py
and/or the build_ext
commands. The distutils.install
simply runs build
if necessary which also runs build_clib
, build_py
and/or build_ext
. This means that regardless of whether you use setuptools
or distutils
, if it is necessary to build from source, the commands build_clib
, build_py
, and/or build_ext
will be runned, so these are the ones that we will want to overload with the cmdclass
of setup
, the question becomes which of the three.
build_py
is used to "build" pure python packages, so we can safely ignore it.
build_ext
is used to build declared Extension modules that are passed through the ext_modules
parameter of the call to the setup
function. If we wish to overload this class, the main method that builds each extension is build_extension
(or here for distutils)
build_clib
is used to build declared libraries that are passed through the libraries
parameter of the call to the setup
function. In this case, the main method that we should overload with our derived class is the build_libraries
method (here for distutils
).
I'll share an example package that builds a toy c static library through a Makefile by using setuptools
build_ext
command. The approach can be adapted to using the build_clib
command, but you'll have to checkout the source code of build_clib.build_libraries
.
setup.py
import os, subprocess
import setuptools
from setuptools.command.build_ext import build_ext
from distutils.errors import DistutilsSetupError
from distutils import log as distutils_logger
extension1 = setuptools.extension.Extension('test_pack_opt.test_ext',
sources = ['test_pack_opt/src/test.c'],
libraries = [':libtestlib.a'],
library_dirs = ['test_pack_opt/lib/'],
)
class specialized_build_ext(build_ext, object):
"""
Specialized builder for testlib library
"""
special_extension = extension1.name
def build_extension(self, ext):
if ext.name!=self.special_extension:
# Handle unspecial extensions with the parent class' method
super(specialized_build_ext, self).build_extension(ext)
else:
# Handle special extension
sources = ext.sources
if sources is None or not isinstance(sources, (list, tuple)):
raise DistutilsSetupError(
"in 'ext_modules' option (extension '%s'), "
"'sources' must be present and must be "
"a list of source filenames" % ext.name)
sources = list(sources)
if len(sources)>1:
sources_path = os.path.commonprefix(sources)
else:
sources_path = os.path.dirname(sources[0])
sources_path = os.path.realpath(sources_path)
if not sources_path.endswith(os.path.sep):
sources_path+= os.path.sep
if not os.path.exists(sources_path) or not os.path.isdir(sources_path):
raise DistutilsSetupError(
"in 'extensions' option (extension '%s'), "
"the supplied 'sources' base dir "
"must exist" % ext.name)
output_dir = os.path.realpath(os.path.join(sources_path,'..','lib'))
if not os.path.exists(output_dir):
os.makedirs(output_dir)
output_lib = 'libtestlib.a'
distutils_logger.info('Will execute the following command in with subprocess.Popen: \n{0}'.format(
'make static && mv {0} {1}'.format(output_lib, os.path.join(output_dir, output_lib))))
make_process = subprocess.Popen('make static && mv {0} {1}'.format(output_lib, os.path.join(output_dir, output_lib)),
cwd=sources_path,
stdout=subprocess.PIPE,
stderr=subprocess.PIPE,
shell=True)
stdout, stderr = make_process.communicate()
distutils_logger.debug(stdout)
if stderr:
raise DistutilsSetupError('An ERROR occured while running the '
'Makefile for the {0} library. '
'Error status: {1}'.format(output_lib, stderr))
# After making the library build the c library's python interface with the parent build_extension method
super(specialized_build_ext, self).build_extension(ext)
setuptools.setup(name = 'tester',
version = '1.0',
ext_modules = [extension1],
packages = ['test_pack', 'test_pack_opt'],
cmdclass = {'build_ext': specialized_build_ext},
)
test_pack/__init__.py
from __future__ import absolute_import, print_function
def py_test_fun():
print('Hello from python test_fun')
try:
from test_pack_opt.test_ext import test_fun as c_test_fun
test_fun = c_test_fun
except ImportError:
test_fun = py_test_fun
test_pack_opt/__init__.py
from __future__ import absolute_import, print_function
import test_pack_opt.test_ext
test_pack_opt/src/Makefile
LIBS = testlib.so testlib.a
SRCS = testlib.c
OBJS = testlib.o
CFLAGS = -O3 -fPIC
CC = gcc
LD = gcc
LDFLAGS =
all: shared static
shared: libtestlib.so
static: libtestlib.a
libtestlib.so: $(OBJS)
$(LD) -pthread -shared $(OBJS) $(LDFLAGS) -o $@
libtestlib.a: $(OBJS)
ar crs $@ $(OBJS) $(LDFLAGS)
clean: cleantemp
rm -f $(LIBS)
cleantemp:
rm -f $(OBJS) *.mod
.SUFFIXES: $(SUFFIXES) .c
%.o:%.c
$(CC) $(CFLAGS) -c $<
test_pack_opt/src/test.c
#include <Python.h>
#include "testlib.h"
static PyObject*
test_ext_mod_test_fun(PyObject* self, PyObject* args, PyObject* keywds){
testlib_fun();
return Py_None;
}
static PyMethodDef TestExtMethods[] = {
{"test_fun", (PyCFunction) test_ext_mod_test_fun, METH_VARARGS | METH_KEYWORDS, "Calls function in shared library"},
{NULL, NULL, 0, NULL}
};
#if PY_VERSION_HEX >= 0x03000000
static struct PyModuleDef moduledef = {
PyModuleDef_HEAD_INIT,
"test_ext",
NULL,
-1,
TestExtMethods,
NULL,
NULL,
NULL,
NULL
};
PyMODINIT_FUNC
PyInit_test_ext(void)
{
PyObject *m = PyModule_Create(&moduledef);
if (!m) {
return NULL;
}
return m;
}
#else
PyMODINIT_FUNC
inittest_ext(void)
{
PyObject *m = Py_InitModule("test_ext", TestExtMethods);
if (m == NULL)
{
return;
}
}
#endif
test_pack_opt/src/testlib.c
#include "testlib.h"
void testlib_fun(void){
printf("Hello from testlib_fun!\n");
}
test_pack_opt/src/testlib.h
#ifndef TESTLIB_H
#define TESTLIB_H
#include <stdio.h>
void testlib_fun(void);
#endif
In this example, the c library that I want to build using the custom Makefile just has one function which prints "Hello from testlib_fun!\n"
to stdout. The test.c
script is a simple interface between python and this library's single function. The idea is that I tell setup
that I want to build a c extension named test_pack_opt.test_ext
, which only has a single source file: the test.c
interface script, and I also tell the extension that it must link against the static library libtestlib.a
. The main thing is that I overload the build_ext
cmdclass using specialized_build_ext(build_ext, object)
. The inheritance from object
is only necessary if you want to be able to call super
to dispatch to parent class methods. The build_extension
method takes an Extension
instance as its second argument, in order to work nice with other Extension
instances that require the default behavior of build_extension
, I check if this extension has the name of the special one and if it doesn't I call the super
's build_extension
method.
For the special library, I call the Makefile simply with subprocess.Popen('make static ...')
. The rest of the command passed to the shell is just to move the static library to a certain default location in which the library should be found to be able to link it to the rest of the compiled extension (which is also just compiled using the super
's build_extension
method).
As you can imagine there are just sooo many ways in which you could organize this code differently, it does not make sense to list them all. I hope this example serves to illustrate how to call the Makefile, and which cmdclass
and Command
derived class you should overload to call make
in a standard installation.
Now, onto question 2.
Question #2: how to ensure, that such terminal command is executed only if the corresponding extra1 is specified during the installation process?
This was possible with the deprecated features
parameter of setuptools.setup
. The standard way is to try to install the package depending on the requirements that are met. install_requires
lists the mandatory requirements, the extras_requires
lists the optional requirements. For example from the setuptools
documentation
setup(
name="Project-A",
...
extras_require={
'PDF': ["ReportLab>=1.2", "RXP"],
'reST': ["docutils>=0.3"],
}
)
you could force the installation of the optional required packages by calling pip install Project-A[PDF]
, but if for some reason the requirements for the 'PDF'
named extra were satisfied before hand, pip install Project-A
would end up with the same "Project-A"
functionality. This means that the way in which "Project-A" is installed is not customized for each extra specified at the command line, "Project-A" will always try to install in the same way and may end up with reduced functionality because of unavailable optional requirements.
From what I understood, this means that in order to get your module X to be compiled and installed only if [extra1] is specified, you should ship module X as a separate package and depend on it through an extras_require
. Lets imagine module X will be shipped in my_package_opt
, your setup for my_package
should look like
setup(
name="my_package",
...
extras_require={
'extra1': ["my_package_opt"],
}
)
Well, I'm sorry that my answer ended up being so long but I hope it helps. Don't hesitate in pointing out any conceptual or naming error, as I mostly tried to deduce this from the setuptools
source code.
Unfortunately, the docs are extremely scarce around the interaction between setup.py and pip, but you should be able to do something like this:
import subprocess
from setuptools import Command
from setuptools import setup
class CustomInstall(Command):
user_options = []
def initialize_options(self):
pass
def finalize_options(self):
pass
def run(self):
subprocess.call(
['touch',
'/home/{{YOUR_USERNAME}}/'
'and_thats_why_you_should_never_run_pip_as_sudo']
)
setup(
name='hack',
version='0.1',
cmdclass={'customcommand': CustomInstall}
)
This gives you a hook into running arbitrary code with commands, and also supports a variety of custom option parsing (not demonstrated here).
Put this in a setup.py
file and try this:
pip install --install-option="customcommand" .
Note that this command is executed after the main install sequence, so depending on exactly what you're trying to do, it may not work. See the verbose pip install output:
(.venv) ayoon:tmp$ pip install -vvv --install-option="customcommand" .
/home/ayoon/tmp/.venv/lib/python3.6/site-packages/pip/commands/install.py:194: UserWarning: Disabling all use of wheels due to the use of --build-options / -
-global-options / --install-options.
cmdoptions.check_install_build_global(options)
Processing /home/ayoon/tmp
Running setup.py (path:/tmp/pip-j57ovc7i-build/setup.py) egg_info for package from file:///home/ayoon/tmp
Running command python setup.py egg_info
running egg_info
creating pip-egg-info/hack.egg-info
writing pip-egg-info/hack.egg-info/PKG-INFO
writing dependency_links to pip-egg-info/hack.egg-info/dependency_links.txt
writing top-level names to pip-egg-info/hack.egg-info/top_level.txt
writing manifest file 'pip-egg-info/hack.egg-info/SOURCES.txt'
reading manifest file 'pip-egg-info/hack.egg-info/SOURCES.txt'
writing manifest file 'pip-egg-info/hack.egg-info/SOURCES.txt'
Source in /tmp/pip-j57ovc7i-build has version 0.1, which satisfies requirement hack==0.1 from file:///home/ayoon/tmp
Could not parse version from link: file:///home/ayoon/tmp
Installing collected packages: hack
Running setup.py install for hack ... Running command /home/ayoon/tmp/.venv/bin/python3.6 -u -c "import setuptools, tokenize;__file__='/tmp/pip-j57ovc7
i-build/setup.py';f=getattr(tokenize, 'open', open)(__file__);code=f.read().replace('\r\n', '\n');f.close();exec(compile(code, __file__, 'exec'))" install --
record /tmp/pip-_8hbltc6-record/install-record.txt --single-version-externally-managed --compile --install-headers /home/ayoon/tmp/.venv/include/site/python3
.6/hack customcommand
running install
running build
running install_egg_info
running egg_info
writing hack.egg-info/PKG-INFO
writing dependency_links to hack.egg-info/dependency_links.txt
writing top-level names to hack.egg-info/top_level.txt
reading manifest file 'hack.egg-info/SOURCES.txt'
writing manifest file 'hack.egg-info/SOURCES.txt'
Copying hack.egg-info to /home/ayoon/tmp/.venv/lib/python3.6/site-packages/hack-0.1-py3.6.egg-info
running install_scripts
writing list of installed files to '/tmp/pip-_8hbltc6-record/install-record.txt'
running customcommand
done
Removing source in /tmp/pip-j57ovc7i-build
Successfully installed hack-0.1