pythonloc is a drop in replacement for python and pip that automatically recognizes a __pypackages__ directory and prefers importing packages installed in this location over user or global site-packages. If you are familiar with node, __pypackages__ works similarly to node_modules.

So instead of running python you run pythonloc and the __pypackages__ path will automatically be searched first for packages. And instead of running pip you run piploc and it will install/uninstall from __pypackages__.

This is an alternative to using Virtual Environments.

This is a Python implementation of PEP 582, "Python local packages directory". The goal of pythonloc is to make an accessible tool while discussion takes place around adding this functionality to CPython itself. If you prefer, you can build your own CPython with these changes instead of using pythonloc.

Please note that PEP 582 has not been accepted. It may or not be accepted in the long term. pythonloc is experimental and its API may change in the future.

# myapp.py
import requests
print(requests)

> piploc install requests
Installing collected packages: urllib3, certifi, chardet, idna, requests
Successfully installed certifi-2018.11.29 chardet-3.0.4 idna-2.8 requests-2.21.0 urllib3-1.24.1

> pipfreezeloc
requests==2.21.0

> pythonloc myapp.py  # works!
<module 'requests' from '/tmp/demo/__pypackages__/3.6/lib/requests/__init__.py'>

Featured on episode #117 of the Python bytes podcast.

"Chad has been working and writing some exciting python tools and articles in the packaging/pip space."

— Jeff Triplett, Python Software Foundation Director

"I’m very enthusiastic about how __pypackages__ could help simplify and streamline the Python dependencies workflow. Well done on bringing an early prototype implementation for people to test!"

— Florimond Manca, Creator of Bocadillo Project

• Python 2.7+
• pip

After installing with

pip install --user pythonloc

or

python3 -m pip install --user pythonloc

you will have four CLI tools available to you: pythonloc, piploc, pipx, and pipfreezeloc.

Short for "python local", it is a drop-in replacement for python with one important difference: the local directory __pypackages__/<version>/lib is added to the front of sys.path. <version> is the Python version, something like 3.7. All arguments are forwarded to python.

So instead of running

python ...

you would run

pythonloc ...

If PEP 582 is adopted, python itself will have this behavior.

Short for "pip local", it invokes pip with the same sys.path as pythonloc. If installing a package, the target installation directory is modified to be __pypackages__ instead of the global site-packages.

If __pypackages__ directory does not exist it will be created.

All arguments are forwarded to pip.

So instead of running

pip ...

you would run

piploc ...

If PEP 582 is adopted, I think pip should default to working in the appropriate __pypackages__ directory. A flag can be added to install to site-packages, if desired.

Note: pipx is included with pythonloc for Python 3.6+ only.

Installing packages that have so called "entry points" to __pypackages__ presents a problem. The entry points, or "binaries", are no longer available on your $PATH as they would be if you installed in a virtual environment or to your system. These binaries are massively popular and useful. Examples of binaries are black, pytest, tox, flake8, mypy, poetry, and pipenv (and indeed pythonloc itself).

pipx is a binary installer and runner for Python that, when run, searches for a binary in the appropriate __pypackages__ location and runs it. If you are familiar with JavaScript's npx, it's similar to that.

So instead of running

BINARY [BINARY ARGS]

you would run

pipx run BINARY [BINARY ARGS]

If not found, pipx will install and run it from a temporary directory. If you require the binary to be found in the __pypackages__ directory, you can run

pipx run --pypackages BINARY [BINARY ARGS]

If the binary is not found, and error will be presented.

Note: When installing a new package to an existing __pypackages__ directory, the entry points will not be created in .../3.6/lib/bin, for example, if something is already there. To do that, you need to run piploc install -U PACKAGE. When you do that, the entire contents of the directory will be replaced. Fixing this would require a modification to pip itself.

If PEP 582 is adopted, pipx will be a good companion tool to run binaries.

Running pip freeze presents a problem because it shows all installed python packages: those in site-packages as well as in __pypackages__. You likely only want to output the packages installed to __pypackages__ and that is exactly what pipfreezeloc does.

It is the equivalent of pip freeze but only outputs packages in __pypackages__. This is required because there is no built-in way to do this with standard pip. For example, the command pip freeze --target __pypackages__ does not exist.

No arguments are handled with pipfreezeloc.

So instead of running

pip freeze > requirements.txt

you would run

pipfreezeloc > requirements.txt

If PEP 582 is adopted, a more robust solution to freezing the state of __pypackages__ should be created.

This works just like it does in pip. You just need a requirements.txt file to install from.

piploc install -r requirements.txt
pythonloc <app>

pip cannot read poetry.lock files, so you'll have to generate a requirements.txt file.

Poetry 1.x provides export command:

poetry self:update --preview #  install 1.x version of Poetry
poetry export -f requirements.txt
piploc install -r requirements.txt
pythonloc <app>

For Poetry 0.x you can adopt the next approach:

poetry run pip freeze > requirements.txt
piploc install -r requirements.txt
pythonloc <app>

pip cannot read Pipfiles yet, only pipenv can. So you will need to generate requirements.txt using pipenv.

pipenv lock --requirements
pipenv lock --requirements --dev
piploc install -r requirements.txt
pythonloc <app>

In the long term tools will be able to install directly to __pypackages__ or piploc will be able to read various lockfile formats.

You can run any python command with pythonloc and it will just run python under the hood:

> pythonloc --help
> pythonloc --version

Another example showing how imports work:

> ls

> pythonloc -c "import requests; print(requests)"
Traceback (most recent call last):
  File "<string>", line 1, in <module>
ModuleNotFoundError: No module named 'requests'

> piploc install requests  # installs to __pypackages__/3.6/lib/requests
Installing collected packages: urllib3, certifi, chardet, idna, requests
Successfully installed certifi-2018.11.29 chardet-3.0.4 idna-2.8 requests-2.21.0 urllib3-1.24.1

> pythonloc -c "import requests; print(requests)"  # requests is now found
<module 'requests' from '/tmp/demo/__pypackages__/3.6/lib/requests/__init__.py'>

> piploc uninstall requests  # uninstalls from __pypackages__/3.6/lib/requests
Successfully uninstalled requests-2.21.0

> piploc install cowsay
Collecting cowsay
  Using cached https://files.pythonhosted.org/packages/e7/e7/e93f311adf63ac8936beb962223771b1ab61227ae3d9ec86e8d9f8f9da1c/cowsay-2.0-py2.py3-none-any.whl
Installing collected packages: cowsay
Successfully installed cowsay-2.0

> pipx run cowsay moooo from local __pypackages__!
  ________________________________
< moooo from local __pypackages__! >
  ================================
                                     \
                                      \
                                        ^__^
                                        (oo)\_______
                                        (__)\       )\/       ||----w |
                                            ||     ||

While this PEP is pretty exciting, there are a some things it doesn't solve.

• OS-dependent packages: The directory structure in __pypackages__ is namespaced on python version, so packages for Python 3.6 will not mix with 3.7, which is great. But sometimes packages install differently for different OS's, so Windows may not match mac, etc.
• site-packages: This PEP first looks to __pypackages__ but will fall back to looking in site-packages. This is not entirely hermetic and could lead to some confusion around which packages are being used. I would prefer the default search path be only __pypackages__ and nothing else.
• perceived downside -- bloat: Many have brought this up in various forums, comparing it to node_modules, but I don't think it applies here. For one, the same if not more content is installed into a virtual environment, so this just moves it into a local directory. No additional bloat. In fact, it is more obvious and can be deleted because it's not hidden away in a virtual env directory. But more importantly, I think the assumption that it is bloated or will be abused stems from JavaScript's ecosystem. JavaScript has a notoriously limited standard library, and developers need to reach for third party packages more often. In addition, the JavaScript community heavily relies on many plugins and transpilation. Python does not. I do not find the bloat argument convincing.
• Some pip installation idiosyncracies. For example, pip install with --target will wipe out content in the lib/bin directory when the -U flag is passed, but not put anything there when it's not passed.

• A virtual environment may or may not include system packages, whereas pythonloc will first look for packages in ., then __pypackages__, then in other locations such as user or site-packages.
• pythonloc does not require activation or deactivation
• pythonloc only looks for a local directory called __pypackages__. On the other hand, virtual environment activation modifies your PATH so you can access virtual environment packages no matter which directory you're in.

It's quite simple and clocks in at less than 100 lines of code. It uses features already built into Python and pip.

All it does is provide a slight level of indirection when invoking Python and pip. It modifies the PYTHONPATH environment variable when running Python to include __pypackages__.

If you consult the output of python --help, you'll see this:

PYTHONPATH is a ':'-separated list of directories prefixed to the default module search path. The result is sys.path.

pythonloc is an alias for PYTHONPATH=.:__pypackages__/<version>/lib:$PYTHONPATH python PYTHONARGS

To install packages to the __pypackages__ directory, it uses pip and runs

PYTHONPATH=.:__pypackages__/<version>/lib:$PYTHONPATH python -m pip PIPARGS

where PIPARGS are whatever arguments you pass it, such as piploc install requests.

It will insert the arguments --target __pypackages__ if you are installing a package.

The installed packages go there. This includes their source code, for example the requests directory below. metadata about the package is stored in the *.dist-info directories.

If you want to modify or debug the source of an installed package, it's very easy to do so. Just open the appropriate file in __pypackages__ and edit away!

> piploc install requests
Installing collected packages: urllib3, certifi, chardet, idna, requests
Successfully installed certifi-2018.11.29 chardet-3.0.4 idna-2.8 requests-2.21.0 urllib3-1.24.1


> ls  # note __pypackages__ was created
__pypackages__

> ls __pypackages__/3.6/lib
bin                           idna-2.8.dist-info
certifi                       requests
certifi-2018.11.29.dist-info  requests-2.21.0.dist-info
chardet                       urllib3
chardet-3.0.4.dist-info       urllib3-1.24.1.dist-info
idna

piploc will automatically add __pypackages__ to $PYTHONPATH, so

piploc uninstall PACKAGE

will work.

If you get the error

Not uninstalling PACKAGE at ..., outside environment ...

then run deactivate to make sure you are not using a virtual environment, then try again.

An easy way to get this behavior is to create a symlink in your local directory

ln -s `which pythonloc` python
ln -s `which piploc` pip

Then run them with

./python
./pip

Otherwise you'll have to build CPython yourself with the reference implementation on GitHub.

There is more overhead involved in building and distributing a custom CPython build than installing a pip package.

You are encouraged to check it out if you are interested though, it's pretty cool!

If it gets accepted and added to CPython then pythonloc may not be needed anymore.

• PEP 582
• reference CPython implementation on GitHub