Pipenv is a Python packaging tool that does one thing reasonably well — application dependency management. However, it is also plagued by issues, limitations and a break-neck development process. In the past, Pipenv’s promotional material was highly misleading as to its purpose and backers.
In this post, I will explore the problems with Pipenv. Was it really recommended by Python.org? Can everyone — or at least, the vast majority of people — benefit from it?
“Pipenv — the officially recommended Python packaging tool from Python.org, free (as in freedom).”
Pipenv’s README used to have a version of the above line in their README for many months: it was added on 2017-08-31 and eventually disappeared on 2018-05-19. For a short while (2018-05-16), it was clarified (managing application dependencies, and PyPA instead of Python.org), and for about 15 minutes, the tagline called Pipenv the world’s worst or something along these lines (this coming from the maintainer).
The README tagline claimed that Pipenv is the be-all, end-all of Python packaging. The problem is: it isn’t that. There are some use cases that benefit from Pipenv, but for many others, trying to use that tool will only lead to frustration. We will explore this issue later.
Another issue with this tagline was the Python.org and official parts. The thing that made it “official” was a short tutorial  on packaging.python.org, which is the PyPA’s packaging user guide. Also of note is the Python.org domain used. It makes it sound as if Pipenv was endorsed by the Python core team. PyPA (Python Packaging Authority) is a separate organization — they are responsible for the packaging parts (including pypi.org, setuptools, pip, wheel, virtualenv, etc.) of Python. This made the endorsement misleading. Of course, PyPA is a valued part of the Python world; an endorsement by the core team — say, inclusion in official Python distributions — is something far more important.
This tagline has led to many discussions and flamewars, perhaps with this Reddit thread from May being the most heated and most important. The change was the direct result of this Reddit thread. I recommend reading this thread in full.
We’ve already learned that Pipenv is used to manage application dependencies. Let’s learn what that term really means.
Here is an example use case for Pipenv: I’m working on a website based on Django. I create ~/git/website and run pipenv install Django in that directory. Pipenv:
- automatically creates a virtualenv somewhere in my home directory
- writes a Pipfile, which lists Django as my dependency
- installs Django using pip
- proceeds to write Pipfile.lock, which stores the exact version and source file hash  of each package installed (including pytz, Django’s dependency).
The last part of the process was the most time consuming. At one point, while locking the dependency versions, Pipenv hangs for 46 seconds. This is one of Pipenv’s notable issues: it’s slow. Of course, this isn’t the only one, but it defintely doesn’t help. Losing 46 seconds isn’t much, but when we get to the longer waits in the timing test section later, we’ll see something that could easily discourage users from using a package.
But let’s continue with our workflow. pipenv run django-admin startproject foobanizer is what I must use now, which is rather unwieldy to type, and requires running pipenv even for the smallest things. (The manage.py script has /usr/bin/env python in its shebang.) I can run pipenv shell to get a new shell which runs the activate script by default, giving you the worst of both worlds when it comes to virtualenv activation: the unwieldiness of a new shell, and the activate script, which the proponents of the shell spawning dislike.
Using pipenv shell means spawning a new subshell, executing the shell startup scripts (eg. .bashrc), and requiring you to exit with exit or ^D. If you type deactivate, you are working with an extra shell, but now outside of the virtualenv. Or you can use the --fancy mode that manipulates $PATH before launching the subshell, but it requires a specific shell configuration, in which $PATH is not overridden in non-login shells — and also often changing the config of your terminal emulator to run a login shell, as many of the Linux terminals don’t do it.
Now, why does all this happen? Because a command cannot manipulate the environment of the shell it spawns. This means that Pipenv must pretend what it does is a reasonable thing instead of a workaround. This can be solved with manual activation using source $(pipenv --venv)/bin/activate (can be made into a neat alias), or shell wrappers (similar to what virtualenvwrapper does).
Anyway, I want a blog on my site. I want to write them in Markdown syntax, so I run pipenv install Markdown, and a few long seconds later, it’s added to both Pipfiles. Another thing I can do is pipenv install --dev ipython and get a handy shell for tinkering, but it will be marked as a development dependency — so, not installed in production. That last part is an important advantage of using Pipenv.
When I’m done working on my website, I commit both Pipfiles to my git repository, and push it to the remote server. Then I can clone it to, say, /srv/website. Now I can just pipenv install to get all the production packages installed (but not the development ones — Django, pytz, Markdown will be installed, but IPython and all its million dependencies won’t). There’s just one caveat: by default, the virtualenv will still be created in the current user’s home directory. This is a problem in this case, since it needs to be accessible by nginx and uWSGI, which do not have access to my (or root’s) home directory, and don’t have a home directory of their own. This can be solved with export PIPENV_VENV_IN_PROJECT=1. But note that I will now need to export this environment variable every time I work with the app in /srv via Pipenv. The tool supports loading .env files, but only when running pipenv shell and pipenv run. You can’t use it to configure Pipenv. And to run my app with nginx/uWSGI, I will need to know the exact virtualenv path anyway, since I can’t use pipenv run as part of uWSGI configuration.
The workflow I mentioned above looks pretty reasonable, right? There are some deficiencies, but other than that, it seems to work well. The main issue with Pipenv is: it works with one workflow, and one workflow only. Try to do anything else, and you end up facing multiple obstacles.
Pipenv only concerns itself with managing dependencies. It isn’t a packaging tool. If you want your thing up on PyPI, Pipenv won’t help you with anything. You still need to write a setup.py with install_requires, because the Pipfile format only specifies the dependencies and runtime requirements (Python version), there is no place in it for the package name, and Pipenv does not mandate/expect you to install your project. It can come in handy to manage the development environment (as a requirements.txt replacement, or something used to write said file), but if your project has a setup.py, you still need to manually manage install_requires. Pipenv can’t create wheels on its own either. And pip freeze is going to be a lot faster than Pipenv ever will be.
Another issue with Pipenv is the use of the working directory to select the virtual environment.  Let’s say I’m a library author. A user of my foobar library has just reported a bug and attached a repro.py file that lets me reproduce the issue. I download that file to ~/Downloads on my filesystem. With plain old virtualenv, I can easily confirm the reproduction in a spare shell with:
And then I can launch my fancy IDE to fix the bug. I don’t have to cd into the project. But with Pipenv, I can’t really do that. If I put the virtualenv in .venv with the command line option, I can type ~/git/foobar/.venv/bin/python ~/Downloads/repro.py. If I use the centralized directory + hashes thing, Tab completion becomes mandatory, if I haven’t memorized the hash.
What if I had two .py files, or repro.py otherwise depended on being in the current working directory?
$ cd ~/git/foobar $ pipenv shell (foobar-Mwd1l2m9)$ cd ~/Downloads (foobar-Mwd1l2m9)$ python repro.py (foobar-Mwd1l2m9)$ exit # (not deactivate!)
This is becoming ugly fairly quickly. Also, with virtualenvwrapper, I can do this:
And let’s not forget that Pipenv doesn’t help me to write a setup.py, distribute code, or manage releases. It just manages dependencies. And it does it pretty badly.
I’m a co-maintainer of a static site generator, Nikola. As part of this, I have the following places where I need to run Nikola:
- ~/website (this blog)
- /Volumes/RAMDisk/n (demo site, used for testing and created when needed, on a RAM disk)
That list is long. End users of Nikola probably don’t have a list that long, but they might just have more than one Nikola site. For me, and for the aforementioned users, Pipenv does not work. To use Pipenv, all those repositories would need to live in one directory. I would also need to have a separate Pipenv environment for nikola-users, because that needs Django. Moreover, the Pipfile would have to be symlinked from ~/git/nikola if we were to make use of those in the project. So, I would have a ~/nikola directory just to make Pipenv happy, do testing/bug reproduction on a SSD (and wear it out faster), and so on… Well, I could also use the virtualenv directly. But in that case, Pipenv loses its usefulness, and makes my workflow more complicated. I can’t use virtualenvwrapper, because I would need to hack a fuzzy matching system onto it, or memorize the random string appended to my virtualenv name. All because Pipenv relies on the current directory too much.
Nikola end users who want to use Pipenv will also have a specific directory structure forced on them. What if the site serves as docs for a project, and lives inside another project’s repo? Two virtualenvs, 100 megabytes wasted. Or worse, Nikola ends up in the other project’s Pipfile, which is technically good for our download stats, but not really good for the other project’s contributors.
Pipenv is famous for being slow. But how slow is it really? I put it to the test. I used two test environments:
- Remote: a DigitalOcean VPS, the cheapest option (1 vCPU), Python 3.6/Fedora 28, in Frankfurt
- Local: my 2015 13” MacBook Pro (base model), Python 3.7, on a rather slow Internet connection (10 Mbps on a good day, and the test was not performed on one of them)
Both were runninng Pipenv 2018.7.1, installed from pip.
And with the following cache setups:
- Removed: ~/.cache/pipenv removed
- Partial: rm -rf ~/.cache/pipenv/depcache-py*.json ~/.cache/pipenv/hash-cache/
- Kept: no changes done from previous run
Well, turns out Pipenv likes doing strange things with caching and locking. A look at the Activity Monitor hinted that there is network activity going on when Pipenv displays its Locking [packages] dependencies... line and hangs. Now, the docs don’t tell you that. The most atrocious example was a local Nikola install that was done in two runs: the first pipenv install Nikola run was interrupted  right after it was done installing packages, so the cache had all the necessary wheels in it. The install took 10 minutes and 7 seconds, 9:50 of which were taken by locking dependencies and installing the locked dependencies — so, roughly nine and a half minutes were spent staring at a static screen, with the tool doing something in the background — and Pipenv doesn’t tell you what happens in this phase.
(Updated 2018-07-22: In the pipenv measurements: the first entry is the total time of pipenv executon. The second is the long wait for pipenv to do its “main” job: locking dependencies and installing them. The timing starts when pipenv starts locking dependencies and ends when the prompt appears. The third item is pipenv’s reported installation time. So, pipenv install ⊇ locking/installing ⊇ Pipfile.lock install.)
|Task||Action||Measurement method||Environment||Cache||Times in seconds|
|Attempt 1||Attempt 2||Attempt 3||Average|
|2||pip install Nikola||time||Remote||Removed||11.562||11.943||11.773||11.759|
|3||pip install Nikola||time||Remote||Kept||7.404||7.681||7.569||7.551|
|4||pipenv install Nikola||time||Remote||Removed||67.536||62.973||71.305||67.271|
|├─ locking/installing from lockfile||stopwatch||42.6||40.5||39.6||40.9|
|└─ Pipfile.lock install||pipenv||14||14||13||13.667|
|5||adding Django to an environment||time||Remote||Kept (only Nikola in cache)||39.576||—||—||39.576|
|├─ locking/installing from lockfile||stopwatch||32||—||—||32|
|└─ Pipfile.lock install||pipenv||14||—||—||14|
|6||adding Django to another environment||time||Remote||Kept (both in cache)||37.978||—||—||37.978|
|├─ locking/installing from lockfile||stopwatch||30.2||—||—||30.2|
|└─ Pipfile.lock install||pipenv||14||—||—||14|
|7||pipenv install Django||time||Remote||Removed||20.612||20.666||20.665||20.648|
|├─ locking/installing from lockfile||stopwatch||6.6||6.4||6||6.333|
|└─ Pipfile.lock install||pipenv||1||1||1||1|
|8||pipenv install Django (new env)||time||Remote||Kept||17.615||—||—||17.615|
|├─ locking/installing from lockfile||stopwatch||3.5||—||—||3.5|
|└─ Pipfile.lock install||pipenv||1||—||—||1|
|9||pipenv install Nikola||time||Remote||Partial||61.507||—||—||61.507|
|├─ locking/installing from lockfile||stopwatch||38.40||—||—||38.40|
|└─ Pipfile.lock install||pipenv||14||—||—||14|
|10||pipenv install Django||time||Local||Removed||73.933||—||—||73.933|
|├─ locking/installing from lockfile||stopwatch||46||—||—||46|
|└─ Pipfile.lock install||pipenv||0||—||—||0|
|12||pip install Nikola (cached)||time||Local||Kept||10.951||—||—||10.951|
|13||pipenv install Nikola||time||Local||Partial, after interruption||607.647||(10m 7s)||607.647|
|├─ locking/installing from lockfile||stopwatch||590.85||(9m 50s)||590.85|
|└─ Pipfile.lock install||pipenv||6||6|
|14||pipenv install||time||Local||Kept||31.399||(L/I: 10.51 s)||31.399|
Python packaging is something with the state of which nobody seems to be satisfied. As such, there are many new contenders for the role of “best new packaging tool”. Apart from Pipenv, there are Hatch (by Ofek Lev) and Poetry (by Sébastien Eustace). Both are listed in the “official” tutorial as alternate options.
Hatch tries to take care of everything in the packaging process. This is mostly an asset, as it helps replace other tools. However, it can also be argued that it adds a single point of failure. Hatch works on already standard files, such as requirements.txt and setup.py, so it can be replaced with something else quite easily. It doesn’t use as much magic as Pipenv and is more configurable. Some choices made by Hatch are questionable (such as manually parsing pkg/__init__.py for a version number, installing test suites to site-packages (a rather common oversight), or its shell feature which is as ugly as Pipenv’s), and it does not do anything to manage dependencies. It doesn’t necessarily work for the Django use case I mentioned earlier, or for end-users of software.
Poetry is somewhere in between. Its main aim is close to Pipenv, but it also makes it possible to distribute things to PyPI. It tries really hard to hide that it uses Pip behind the scenes. Its README comes with an extensive “What about Pipenv?” section, which I recommend reading — it has a few more examples of bad Pipenv features. Poetry claims to use the standardized (PEP 518) pyproject.toml file to replace the usual lot of files. Unfortunately, the only thing that is standardized is the file name and syntax. Poetry uses custom [tool.poetry] sections, which means that one needs Poetry to fully use the packages created with it, leading to vendor lock-in. (The aforementioned Hatch tool also produces a pyproject.tmpl, which contains a metadata section…) There is a build feature to produce a sdist with setup.py and friends.
In a simple poetry add Nikola test, it took 24.4s/15.1s/15.3s to resolve dependencies (according to Poetry’s own count, Remote environment, caches removed), complete with reassuring output and no quiet lockups. Not as good as pip, but it’s more reasonable than Pipenv. Also, the codebase and its layout are rather convoluted. Poetry produces packages instead of just managing dependencies, so it’s generally more useful than Pipenv.
But in all the talk about new tools, we’re forgetting about the old ones, and they do their job well — so well in fact, that the new tools still need them under the covers.
Pip is fast. It does its job well enough. It lacks support for splitting packages between production and development (as Pipenv and Poetry do). This means that pip freeze and pip install are instant, at the cost of (a) needing two separate environments, or (b) installing development dependencies in production (which should only be a waste of HDD space and nothing more in a well-architected system).
The virtualenv management features can be provided by virtualenvwrapper. That tool’s main advantage is the shell script implementation, which means that workon foo activates the foo virtualenv without spawning a new subshell (an issue with Pipenv, Hatch, and Poetry, that I already covered when describing Pipenv’s operation in the Running scripts (badly) chapter.) An argument often raised by Pipenv proponents is that one does not need to concern itself with creating the virtualenv, and doesn’t need to care where it is. Unfortuntately, many tools require this knowledge from their user, or force a specific location, or require it to be different to the home directory.
And for a reasonable project template with release automation — well, I have my own entry in that category, called (rather unoriginally) the Python Project Template (PyPT).
Yes, setup.py files are not ideal, since they use .py code and a function execution, making access to meta information hard (./setup.py egg_info creates tool-accessible text files). Their main advantage is that they are the only format that is widely supported — pip is the de-facto default Python package manager (which is pre-installed on Windows and Mac), and other tools would require installation/bootstrapping first.
A good packaging tool is stable. In other words, it doesn’t change often, and it strives to support existing environments. It wouldn’t be fun to re-download everything on your system, because someone decided that /usr is now called /stuff, and all the files in /usr would become forgotten and not removed. Well, this is what Pipenv did:
|2017-01-31 22:01||v3.2.14 released. pipenv --three creates ./.venv (eg. ~/git/foo/.venv). Last version with the original behavior of pipenv.|
|2017-02-01 05:36||v3.3.0 released. pipenv --three creates ~/.local/share/virtualenvs/foo (to be precise, $WORKON_HOME/foo).|
|2017-02-01 06:10||Issue #178 is reported regarding the behavior change.|
|2017-02-01 06:18||Kenneth Reitz responds: “no plans for making it configurable.” and closes the issue.|
|2017-02-02 03:05||Kenneth Reitz responds: “added PIPENV_VENV_IN_PROJECT mode for classic operation. Not released yet.”|
|2017-02-02 04:29||v3.3.3 released. The default is still uses a “remote” location, but .venv can now be used.|
|2017-03-02 13:48||v3.5.0 released. The new default path is $WORKON_HOME/foo-HASH, eg. ~/.local/share/virtualenvs/foo-7pl2iuUI.|
Over the course of a month, the location of the virtualenv changed twice. If the user didn’t read the changelog and didn’t manually intervene (also of note, the option name was mentioned in the issue and in v3.3.4’s changelog), they would have a stale .venv directory, since the new scheme was adopted for them. And then, after switching to v3.5.0, they would have a stale virtualenv hidden somewhere in their home directory, because pipenv decided to add hashes.
Pipenv is a very opinionated tool, and if the dev team changes their mind, the old way is not supported.
Pipenv moves fast and doesn’t care if anything breaks. As an example, between 2018-03-13 13:21 and 2018-03-14 13:44 (a little over 24 hours), Pipenv had 10 releases, ranging from v11.6.2 to v11.7.3. The changelog is rather unhelpful when it comes to informing users what happened in each of the releases.
- Pipenv, contrary to popular belief and (now removed) propaganda, is not an officially recommended tool of Python.org. It merely has a tutorial written about it on packaging.python.org (page run by the PyPA).
- Pipenv solves one use case reasonably well, but fails at many others, because it forces a particular workflow on its users.
- Pipenv does not handle any parts of packaging (cannot produce sdists and wheels). Users who want to upload to PyPI need to manage a setup.py file manually, alongside and independently of Pipenv.
- Pipenv produces lockfiles, which are useful for reproducibility, at the cost of installation speed. The speed is a noticeable issue with the tool. pip freeze is good enough for this, even if there are no dependency classes (production vs development) and no hashes (which have minor benefits) 
- Hatch attempts to replace many packaging tools, but some of its practices and ideas can be questionable.
- Poetry supports the same niche Pipenv does, while also adding the ability to create packages and improving over many gripes of Pipenv. A notable issue is the use of a custom all-encompassing file format, which makes switching tools more difficult (vendor lock-in).
- Pip, setup.py, and virtualenv — the traditional, tried-and-true tools — are still available, undergoing constant development. Using them can lead to a simpler, better experience. Also of note, tools like virtualenvwrapper can manage virtualenvs better than the aforementioned new Python tools, because it is based on shell scripts (which can modify the enivironment).