Tsutsumu: A Python Module Bundler and Runtime

つつむ (tsutsumu), Japanese for bundle

Tsutsumu makes it easy to create a module bundle, i.e., a single file that, like a file archive, contains many more modules and supporting resources and, like a shell script, imports and executes such bundled modules. That way, Tsutsumu enables self-contained scripts that run anywhere a suitable Python interpreter is available—without creating a virtual environment or installing packages first. Tsutsumu comes with its own, simple, textual bundle format for when trust is lacking and recipients prefer to inspect code before running. Tsutsumu can also target Python's builtin zipapp format, which has the benefits of being more compact and hence more suitable for larger bundles. Independent of format, Tsutsumu automatically determines the packages—and also the package extras—to include in a bundle, something zipapp doesn't know how to do. (zipapp support will ship with release 0.2. So does automatic dependency resolution.)

Having said that, Tsutsumu isn't the only option for more easily distributing Python code and it may very well not be the right option for your use case. In addition to the already mentioned zipapp module in the standard library, there also are, for example, PEX and PyInstaller, which combine bundling with virtual environments (PEX) and further include the Python runtime as well (PyInstaller). That makes them more sophisticated but also significantly more resource-intensive. Tsutsumu's simplicity makes it best suited to tools and applications of modest size that need to run in untrusted or resource-limited environments.

The rest of this document covers Tsutsumu thusly:

1. Just Download and Run!
2. Make a Bundle
3. The Workings of a Bundle
   1. Layout of Bundled Files
   2. A Bundle's Manifest: Offsets and Lengths
   3. On-Disk vs In-Memory
   4. Meta-Circular Bundling
   5. importlib.resources Considered Harmful
   6. Add a Resource Manifest Instead
4. Coming Soon

1. Just Download and Run!

There is nothing to install. There is no virtual environment to set up. Just download this one Python script and run it:

% curl -o tsutsumu.py \
    "https://raw.githubusercontent.com/apparebit/tsutsumu/boss/bundles/bundler.py"
% python tsutsumu.py -h
usage: tsutsumu [-h] [-b] [-m MODULE] [-o FILENAME] [-r] [-v]
                PKGROOT [PKGROOT ...]

Combine Python modules and related resources into a single,
...

Yup. I used Tsutsumu to bundle its own modules into bundler.py. As a result, getting started with Tsutsumu is as easy as downloading a file and running it. Bundled scripts can be this easy and convenient!

The Complicated Route Still Works

But just in case that you prefer to take the slow and familiar route, you can do that, too. It just requires 3.5 times more command invocations and takes quite a bit longer. But sure, here you go:

% mkdir tsutsumu
% cd tsutsumu
% python -m venv .venv
% source .venv/bin/activate
(.venv) % pip install --upgrade pip
(.venv) % pip install tsutsumu
(.venv) % python -m tsutsumu -h
usage: tsutsumu [-h] [-b] [-m MODULE] [-o FILENAME] [-r] [-v]
                PKGROOT [PKGROOT ...]

Combine Python modules and related resources into a single,
...

So how about bundling your Python modules?

2. Make a Bundle

The only challenge in making a bundle is in selecting the right directories for inclusion. Right now, you need to list every package that should be included in the bundle as a separate directory argument to Tsutsumu. Alas, for most Python tools and applications, that's just the list of regular dependencies. While module-level tree-shaking might still be desirable, automating package selection based on a project's pyproject.toml is an obvious next step.

When Tsutsumu traverses provided directories, it currently limits itself to a few textual formats based on file extension. In particular, it includes plain text, Markdown, ReStructured Text, HTML, CSS, JavaScript, and most importantly Python sources. Out of these, Tsutsumu only knows how to execute Python code. The other files serve as resources. Adding support for Base85-encoded binary formats seems like another obvious next step.

3. The Workings of a Bundle

This section is a hands-on exploration of Tsutsumu's inner workings. Its implementation is split across the following modules:

• tsutsumu for Tsutsumu's __version__ and nothing else;
• tsutsumu.__main__ for the main() entry point and command line interface;
• tsutsumu.debug for validating the manifest and contents of bundle scripts;
• tsutsumu.maker for generating bundles with the BundleMaker class;
• tsutsumu.bundle for importing from bundles with the Bundle class.

As that breakdown should make clear, tsutsumu.maker and tsutsumu.bundle provide the critical two classes that do all the heavy lifting. Hence I'll be focusing on them in this section. To illustrate their workings, I rely on the spam package also contained in Tsutsumu's source repository. In addition to its own __init__ package module, the package contains two Python modules, __main__ and bacon, as well as a very stylish webpage, ham.html, that also includes an image, bacon.jpg.

All subsequent code examples have been validated with Python's doctest tool. Running the tool over this file is part of Tsutsumu's test suite.

Let's get started making a bundle with the contents of the spam directory:

>>> import tsutsumu.maker
>>> maker = tsutsumu.maker.BundleMaker(['spam'])
>>> maker
<tsutsumu-maker spam>
>>>

The bundle maker ultimately needs to produce a Python script. To get there, the bundle maker processes data from byte to file granularity, which is quite the spread. At the same time, it's easy enough to format strings that are entire lines and, similarly, break down larger blobs into individual lines. Hence, most bundle maker methods treat the source line as the common unit of abstraction. However, since files are stored as byte string, not character strings, and byte counts do matter, those source lines are bytes, not str, and include newlines, just \n.

Having said that, the bundle maker starts out by iterating over the contents of directories, yielding the files to be bundled. For each such file, it yields an operating-system-specific Path—suitable for reading the file's contents from the local file system—as well as a relative str key with forward slashes—for identifying the file in the bundle's manifest. Here are the bundle maker's keys for spam:

>>> files = list(sorted(maker.list_files(), key=lambda f: f.key))
>>> for file in files:
...     print(file.key)
...
spam/__init__.py
spam/__main__.py
spam/bacon.jpg
spam/bacon.py
spam/ham.html
>>>

Those are just the five files we expect:

• spam/__init__.py contains spam's package module;
• spam/__main__.py is the package's main entry point;
• spam/bacon.jpg is a package resource;
• spam/bacon.py contains the spam.bacon submodule;
• spam/ham.html is a package resource, too.

3.1 Layout of Bundled Files

Now that we know which files to include in the bundle, we can turn to their layout in bundle scripts. The current format tries to reconcile two contradictory requirements: First, the layout must be valid Python source code. That pretty much limits us to string literals for file names and contents. Furthermore, since the collection of file names and contents obviously forms a mapping, we might as well use a dict literal for the file data.

Second, the code must not retain the bundled data. Otherwise, all bundled files are loaded into memory at startup and remain there for the duration of the application's runtime. Ideally, the Python runtime doesn't even instantiate the dict literal and just scans for its end. To facilitate that, the bundle script does not assign the dict literal to a variable and, on top of that, includes it only inside an if False: branch.

>>> writeall = tsutsumu.maker.BundleMaker.writeall
>>> writeall(tsutsumu.maker._BUNDLE_START.splitlines(keepends=True))
if False: {
>>>

I don't know whether CPython does optimize parsing along those lines. Though I do know that Donald Knuth's TeX (which dates back to the late 1970s) does optimize just this case: Once TeX knows that a conditional branch is not taken, it simply scans upcoming tokens, taking only \if (and variations thereof), \else, and \fi into account, until it has found the end of the branch, after which TeX resumes regular processing.

Let's hope that Python is just as clever and fill in the file name, content pairs for each bundled file. We start with spam/__init__.py:

>>> writeall(maker.emit_file(*files[0]))
# ------------------------------------------------------------------------------
"spam/__init__.py": b"print('spam/__init__.py')\n",
>>>

As shown, the file name or key is a str literal, whereas the file contents are bytes. We use bytes instead of characters for the latter because, at their most basic, files are just that, bytestrings. We get characters only after decoding, nowadays typically from UTF-8.

Beware of bytestring literals in Python: They are limited to ASCII characters and require that all other code points be escaped. In other words, the majority of code points in bytestring literals must be escaped. That would make for a rather verbose encoding if values were more evenly distributed. However, in the case of Tsutsumu, the bundled files are mostly text files, in particular Python source code. That strongly biases bundled files towards ASCII and makes this an efficient and human-readable encoding.

The spam package's __main__ module isn't so different from the __init__ module, except that it takes up several lines and hence uses triple-quotes:

>>> writeall(maker.emit_file(*files[1]))
# ------------------------------------------------------------------------------
"spam/__main__.py":
b"""print('spam/__main__.py')
import spam.bacon
<BLANKLINE>
print('also:', __file__)
""",
>>>

Unlike the other files in the bundle, the next file contains bitmap image and hence is binary. Its contents are represented by a bytestring literal too, but the contents have been encoded in Base85. For readability, the literal adds newlines every 76 characters. By design, Base85 uses only ASCII characters and hence there should be no escape sequences in the bytestring literals for binary files.

>>> writeall(maker.emit_file(*files[2]))   # doctest: +ELLIPSIS
# ------------------------------------------------------------------------------
"spam/bacon.jpg": b"""
s4IA0!"_al8O`[\!<<,,!42_+s5<sN7<iNY!!#_f!%IsK!!iQ.!>5A7!!!!"!!*'"!?(qA!!!!"!
!!!k!?2"B!!!!"!!!!s!AOQU!!!!5!!!"&LM6_k!!!!"!!!":z!!!#+!!!!"!!!#+!!!!"6"FnCA
KXHVEb0H5Ebf_=6W5c@!!Akp!!<3$!!*'#!!&Yn!!E9%!!*'"!;`>j!!E9%!!*'"!/U[U!!*&d!'
!egDffo=BQ%i41G1?]3'p22"9\])z3'p22"=4$J!!!!1e/aM$NrZHgl$s)-m.`nrs1eUH#QT\]q?
...
>>>

Note that the complete encoded image is larger than what would fit into four measly lines of Base85, a bit more than 13,000 bytes larger.

For the fourth and fifth file, we are back to text again:

>>> writeall(maker.emit_file(*files[3]))
# ------------------------------------------------------------------------------
"spam/bacon.py": b"print('spam/bacon.py')\n",
>>> writeall(maker.emit_file(*files[4]))
# ------------------------------------------------------------------------------
"spam/ham.html":
b"""<!DOCTYPE html>
<html lang=en>
<meta charset=utf-8>
<title>Ham?</title>
<style>
* {
    margin: 0;
    padding: 0;
}
html {
    height: 100%;
}
body {
    min-height: 100%;
    display: grid;
    justify-content: center;
    align-content: center;
}
img {
    height: calc(15vmin + 1vmax);
    width: auto;
    display: block;
    position: relative;
    left: -20%;
    top: 40%;
}
p {
    font-family: system-ui, sans-serif;
    font-size: calc(30vmin + 3vmax);
    font-weight: bolder;
}
</style>
<img src=bacon.jpg><p>Ham!
""",
>>>

With that, we can close the dictionary again:

>>> writeall(tsutsumu.maker._BUNDLE_STOP.splitlines(keepends=True))
}
<BLANKLINE>
>>>

3.2 A Bundle's Manifest: Offsets and Lengths

A bundle's files are encoded as a dict literal by design—so that the script parses—but are not assigned to any variable by design as well—so that the script does not retain access to the data, which would only increase memory pressure. So if the script doesn't retain a reference to the data, how does it access the data when it's needed?

I've already hinted at the solution: While turning file names and contents into yielded lines of the bundle script, the bundle maker tracks the byte offset and length of each content literal. It helps that the bundle maker is implemented as a class with several methods that are generators instead of as a bunch of generator functions. That way, accumulating state while yielding lines only requires another method call, with the state stored by the bundle maker instance. It also helps that the bundle maker emits bundle contents first, at the beginning of the content script and that it relies on named string constants for the boilerplate before, between, and right after the file contents dictionary.

Once the bundle maker is done with the file contents, it emits the manifest with the offset and length for each file included in the bundle:

>>> writeall(maker.emit_manifest())
__manifest__ = {
    "spam/__init__.py": ("t", 305, 30),
    "spam/__main__.py": ("t", 438, 77),
    "spam/bacon.jpg": ("b", 620, 13_003),
    "spam/bacon.py": ("t", 13_726, 27),
    "spam/ham.html": ("t", 13_853, 534),
}
>>>

The data collected while yielding the file contents is one datum more granular than offset and length. But the generator for the manifest consumes the output of another generator that accumulates the original three length values per file. As you can see, Tsutsumu's not so secret sauce are generator functions and methods!

Tsutsumu's source repository does not just include the spam package. But its so far tiny collection of prebundled scripts includes can.py, which already bundles the package. If you check can.py's contents, you should see the exact same files in the same order with the same offsets and lengths. That means that we can use the bundle to illustrate how the bundle runtime reads a file such as spam/bacon.py:

>>> with open('bundles/can.py', mode='rb') as file:
...     _ = file.seek(13_726)
...     data = file.read(27)
...
>>> data
b'b"print(\'spam/bacon.py\')\\n"'
>>>

As you can see, the returned bytes aren't just the file contents, but also the leading and trailing characters necessary for turning the contents into a valid Python bytestring literal. We need those "decorations" in the script, so that Python knows to parse the bytestring. But why read those extra characters?

Python bytestring literals represent 256 values per byte with ASCII characters only. As a result, some code points necessarily require escape sequences. In fact, there are more code points that require escaping than printable ASCII characters. Nonetheless, this is a reasonable encoding for this domain because Python source code draws on ASCII mostly and remains human-readable under the encoding.

Still, we can't escape escape sequences—as the above example illustrates. Notice the trailing \\n? That's an escaped newline taking up two bytes in the bytestring. So why read a bytestring, as indicated by the leading b', containing a bytestring literal, as indicated by the subsequent b", when we really want proper bytes?

Here's why:

>>> eval(data)
b"print('spam/bacon.py')\n"
>>>

It only takes an eval to turn two consecutive bytestring prefixes and backslash characters into one each, producing real bytes.

3.3 On-Disk vs In-Memory

As presented so far, bundled files are named by relative paths with forward slashes. That makes sense for bundle scripts while they are inert and being distributed. After all, the raison d'être for Tsutsumu's bundle scripts is to be easily copied to just about any computer and run right there. That wouldn't be practical if the names used in the bundle were tied to the originating file system or limited to some operating system only.

However, the naming requirements change fundamentally the moment a bundle starts to execute on some computer. That instance should seamlessly integrate with the local Python runtime and operating system, while also tracking provenance, i.e., whether modules originate from the bundle or from the local machine. In other words, a running bundle uses absolute paths with the operating system's path segment separator. Sure enough, the constructor for tsutsumu.bundle.Bundle performs the translation from relative, system-independent paths to absolute, system-specific paths by joining the absolute path to the bundle script with each key.

Let's see how that plays out in practice on the example of the can.py bundle:

>>> import bundles.can
>>> manifest = bundles.can.__manifest__
>>> for key in manifest.keys():
...     print(key)
...
spam/__init__.py
spam/__main__.py
spam/bacon.jpg
spam/bacon.py
spam/ham.html
>>>

Clearly, the __manifest__ is using relative paths.

Since bundles.can isn't __main__, importing the bundle resulted in the definition of the __manifest__ dictionary and the Bundle class but it did not install a new Bundle instance in the module loading machinery. Before we manually install the bundle, there's a bit of housekeeping to do. We need to cut off our ability to load modules from the regular file system. Otherwise, we might inadvertently import the spam package from its sources and get mightily confused. (Not that that ever happened to me...)

>>> bundles.can.Toolbox.restrict_sys_path()
>>> import spam
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
ModuleNotFoundError: No module named 'spam'
>>>

No more readily available spam? Time to open bundles.can:

>>> from pathlib import Path
>>> can_path = Path('.').absolute() / 'bundles' / 'can.py'
>>> version = bundles.can.__version__
>>> can_content = bundles.can.Bundle.install(can_path, version, manifest)
>>> import spam
spam/__init__.py
>>>

W00t, our supply of spam is secured. That's great. But how does it work? What did Bundle.install() do exactly?

Well, a Bundle is what importlib's documentation calls an importer, a class that is both a meta path finder and a loader. When Python tries to load a module that hasn't yet been loaded, it (basically) invokes find_spec(name) on each object in sys.meta_path, asking that meta path finder whether it recognizes the module. If the meta path finder does, it returns a description of the module. Most fields of that spec are just informative, i.e., strings, but one field, surprisingly called loader, is an object with methods for loading and executing the module's Python code. It just happens that Bundle does not delegate to a separate class for loading but does all the work itself.

In short, Bundle.install() creates a new Bundle() and makes that bundle the first entry of sys.meta_path.

Ok. But what about the bundle using absolute paths?

>>> for key in can_content._manifest.keys():
...     path = Path(key)
...     assert path.is_absolute()
...     print(str(path.relative_to(can_path)).replace('\\', '/'))
...
spam/__init__.py
spam/__main__.py
spam/bacon.jpg
spam/bacon.py
spam/ham.html
>>>

Clearly, the installed can_content bundle is using absolute paths. Also, each key now starts with the bundle script's path, which we recreated in CAN. While we usually don't worry much about these paths when importing modules in Python, we do need to use them when loading resources from a package:

>>> data = can_content.get_data(can_path / 'spam' / 'ham.html')
>>> data[-5:-1]
b'Ham!'
>>>

Ham! it is.

My apologies to vegetarians. You probably are tired of all this ham-fisted humor by now. So let's make sure we stop right here:

>>> can_content.uninstall()
>>> import spam.bacon
Traceback (most recent call last):
  ...
ModuleNotFoundError: No module named 'spam.bacon'
>>>

Alas, already imported modules are much harder to expunge. In fact, it may just be impossible. In this case, however, it is feasible:

>>> import sys
>>> 'spam' in sys.modules
True
>>> import spam
>>> del sys.modules['spam']
>>> 'spam' in sys.modules
False
>>> import spam
Traceback (most recent call last):
  ...
ModuleNotFoundError: No module named 'spam'
>>>

3.4 Meta-Circular Bundling

Tsutsumu can bundle any application that is not too large and written purely in Python. That includes itself. Tsutsumu can bundle itself because it avoids the file system when including its own tsutsumu/bundle.py in the bundle script. Instead, it uses the module loader's get_data() method, which is designed for accessing packaged resources and whose use I just demonstrated.

One drawback of Tsutsumu treating its own source code just like other Python files is the effective duplication of tsutsumu/bundle.py, once as part of the bundled files and once as part of the bundle script itself. While that may be desirable, for example, when experimenting with a new version of Tsutsumu, it also wastes almost 8 kb. To avoid that overhead, you can use the -r/--repackage command line option when bundling Tsutsumu. Under that option, Tsutsumu special cases the tsutsumu and tsutsumu.bundle modules and recreates them during startup—with tsutsumu's bundle attribute referencing the tsutsumu.bundle module and tsutsumu.bundle's Bundle attribute referencing the corresponding class.

3.5 importlib.resources Considered Harmful

While Tsutsumu does support Loader.get_data(), it does not support the more recent Loader.get_resource_reader() and probably never will. The API simply is too complex for what it does, i.e., providing yet another way of traversing a hierarchy of directory-like and file-like entities. Furthermore, the documentation's claims about the benefits of integration with Python's import machinery seem farfetched at best.

A look at the 8 (!) modules implementing importlib.resources in the standard library bears this out: In addition to the documented ResourceReader, Traversable, and TraversableReader abstract classes, there are undocumented FileReader, ZipReader, and NamespaceReader implementations, the SimpleReader fallback implementation, and the CompatibilityFiles adapter. Furthermore, since Traversable is designed to have "a subset of pathlib.Path methods," the code in importlib.resources makes heavy use of the Path implementations provided by pathlib and zipfile. Taken together, that's a lot of code for exposing a hierarchy of directory- and file-like entities. Worse, despite the documentation's claims to the contrary, none of this code leverages core importlib machinery—besides hanging off loaders and hence touching on ModuleType and ModuleSpec. In fact, it doesn't even integrate with the previous resource API, the much simpler get_data() method on loaders. In summary, importlib.resources does not offer what it claims and is far too complex for what it offers. It should be scrapped!

3.6 Add a Resource Manifest Instead

When you compare the two ways of accessing resources, Loader.get_data() and Loader.get_resource_reader(), the latter obviously wins on traversing a package's namespace. But that's a non-feature when it comes to resource access. When code needs a resource, it shouldn't need to search for the resource by searching them all, it should be able to just access the resource, possibly through one level of indirection. In other words, if a package's resources may vary, the package should include a resource manifest at a well-known location, say, manifest.toml relative to the package's path. Once the package includes a manifest, Loader.get_data() more than suffices for retrieving resources. Loader.get_resource_reader() only adds useless complexity.

4. Coming Soon

I believe that Tsutsumu is ready for real-world use. However, since it hasn't seen wide usage, I'd hold off on mission-critical deployments for now. Meanwhile, Tsutsumu could use a few more features. I can think of three:

• Automatically determine module dependencies
• Support inclusion of binary files in bundles
• [-] Support the bundling of namespace packages

The first two features are mostly implemented and the third also has basic support. Alas none of them have been released. They are scheduled for v0.2. Since zipapp also lacks the automatic dependency discovery, Tsutsumu will likely add support for that format, too.

What else?

---

Tsutsumu is © 2023 Robert Grimm and has been released under the Apache 2.0 license.