shenzi 0.0.3

A greedy Python standalone application bundler

shenzi

shenzi helps you create standalone Python applications from your development virtual environment. Using shenzi, you can create standalone folders which can be distributed to any machine, and the application will work (even when python is not installed on the target system).

The python packaging problem

Given a development environment (a virtual environment), we want to produce a single directory containing ALL the dependencies that the application needs. Other languages like rust and go provide easy way to create statically linked executables, which makes them very easy to distribute.
Python struggles in this area mainly because of how flexible it is when it comes to delegating work to C code (shared libraries on your system).

Out in the wild, python libraries regularly links to shared libraries in your system:

• C Extensions
• loading shared libraries using dlopen and equivalents

Even creating a development environment for some pip package might require you to install some system dependencies (a good example is weasyprint)
It becomes difficult to ship applications if we need to install system dependencies in target machines. Docker solves this problem by packaging everything in a single docker image.
shenzi does not compete with docker, if you can use docker, you should. shenzi is useful for shipping desktop applications.

Getting Started

First install shenzi in your virtual environment.

pip install shenzi

Initializing the workspace

If you have a project run using poetry, run

# only poetry package manager is supported
shenzi init

It will ask you some questions and generate shenzi_workspace.toml file. The TOML file looks like this.

# shenzi_workspace.toml
# all relative paths are relative to the directory containing this file

# you can add a list of binaries that your application calls
# something like calling aws cli. Shenzi would try to find all these in your path and add them to the distribution
binaries = ["tesseract"]

[packaging]
kind = "poetry"
config_file = "<relative-path-to-poetry.lock>"
# you can add the dependency groups you want in the distribution (dev, or other custom groups)
groups = ["main"]

[execution]
main = "<relative-path-to-main-python-script>"

Intercepting

You need to first configure shenzi to listen to all the imports that your python application makes. You can either do this by running your application in your development environment and testing it. Or running tests.

Running an application

In you main script, add the following lines

import os

if os.environ.get("SHENZI_INIT_DISCOVERY", "False") == "True":
    from shenzi.discovery import shenzi_init_discovery
    shenzi_init_discovery()

In pytest

If you are running tests in pytest, you can add this function in your root conftest.py

# root conftest.py


# this function is run by pytest in the beginning
def pytest_configure():
    from shenzi.discover import shenzi_init_discovery
    shenzi_init_discovery()

Run your application as you normally do/or run tests. shenzi will start intercepting all shared libraries that your code is importing.
You should run as much of your application code as possible, like running all the tests. This allows shenzi to detect every dependency linked to your application at runtime.

Once you stop the application, a file shenzi.json (called the manifest) will be dumped in the current directory. This file contains all the shared library loads that shenzi detected. It also contains some information about your virtual environment.
Now run the shenzi CLI with this manifest file

Building the application

From the directory containing shenzi_workspace.toml (your project's root directory), run this command:

RUST_LOG=INFO shenzi build ./shenzi.json

This can take a moment, after it is done, your application would be packaged in a dist folder.
You can ship this dist folder to any target machine and it should work out of the box. The only required dependency is bash.

Note: by default shenzi would try to validate if some warnings are actually errors. It needs to scan the whole file system to do that, it would print a log like this: shenzi will now validate if any of your warnings are errors, this can take time (it will scan your whole file system). You can skip this by passing --skip-warning-checks. If you feel its taking too long, you can skip it by passing --skip-warning-checks. You should however, at least have one successful build with all warnings validated.

Run dist/bootstrap.sh to run your application.

# bootstrap.sh is the entrypoint for your application
# you can run this from any directory generally
bash dist/bootstrap.sh

Note that if you don't specify main file in your shenzi_workspace.toml, shenzi would try to dynamically query that file, this can be annoying if you are running tests, so setting the file in workspace config is useful.

Next steps

You should at least read the doc which describes the structure of shenzi.json here.

If you use this, feel free to raise an issue on any problem, I need feedback for this :)

How is this different?

I will add a small comparison to PyInstaller, which I feel is the most mature tool in the ecosystem.
From what I've seen, PyInstaller statically analyses your python code (and does some imports too) to create the smallest possible packaged application. It is smarter than shenzi.

• shenzi is much simpler. It tries to intercept all linker activity during runtime.
  • During packaging, shenzi will faithfully analyze all dependencies in the same order as done by the linker. Following the linker might solve a class of edge cases (not proved though, for all I know, this algorithm might end up performing very poorly)
• It also packages everything in your python path (all data+code in your site-packages).
  • This makes shenzi faster in some cases (where you have complex applications, as we do not do any static analysis), but slower in others (mainly if your virtual environment is huge, and not all dependencies are used by your application normally)

Apart from that, there are some other internal differences that may or may not matter

• The structure of the final application (described here). It's slightly similar to how pnpm organizes node_modules as far as I'm aware.
• The bootstrap script in shenzi is pretty a simple bash script, it simply sets up the correct Python environment variables and starts the interpreter. PyInstaller has a very sophisticated bootstrapping CLI written in C

Supported Platforms

Currently only Mac and Linux are supported.
The project is very new right now, I've tested it on Ubuntu 20.04 and MacOS Sequoia with Python 3.9