pyodide/docs/new_packages.md

9.7 KiB

Creating a Pyodide package

Pyodide includes a toolchain to make it easier to add new third-party Python libraries to the build. We automate the following steps:

  • Download a source tarball (usually from PyPI)
  • Confirm integrity of the package by comparing it to a checksum
  • Apply patches, if any, to the source distribution
  • Add extra files, if any, to the source distribution
  • If the package includes C/C++/Cython extensions:
    • Build the package natively, keeping track of invocations of the native compiler and linker
    • Rebuild the package using emscripten to target WebAssembly
  • If the package is pure Python:
    • Run the setup.py script to get the built package
  • Package the results into an emscripten virtual filesystem package, which comprises:
    • A .data file containing the file contents of the whole package, concatenated together
    • A .js file which contains metadata about the files and installs them into the virtual filesystem.

Lastly, a packages.json file is output containing the dependency tree of all packages, so {ref}pyodide.loadPackage <js_api_pyodide_loadPackage> can load a package's dependencies automatically.

mkpkg

If you wish to create a new package for pyodide, the easiest place to start is with the mkpkg tool. If your package is on PyPI, just run:

bin/pyodide mkpkg $PACKAGE_NAME

This will generate a meta.yaml (see below) that should work out of the box for many pure Python packages. This tool will populate the latest version, download link and sha256 hash by querying PyPI. It doesn't currently handle package dependencies, so you will need to specify those yourself.

The meta.yaml file

Packages are defined by writing a meta.yaml file. The format of these files is based on the meta.yaml files used to build Conda packages, though it is much more limited. The most important limitation is that Pyodide assumes there will only be one version of a given library available, whereas Conda allows the user to specify the versions of each package that they want to install. Despite the limitations, keeping the file format as close as possible to conda's should make it easier to use existing conda package definitions as a starting point to create Pyodide packages. In general, however, one should not expect Conda packages to "just work" with Pyodide. (In the longer term, Pyodide may use conda as its packaging system, and this should hopefully ease that transition.)

The supported keys in the meta.yaml file are described below.

package

package/name

The name of the package. It must match the name of the package used when expanding the tarball, which is sometimes different from the name of the package in the Python namespace when installed. It must also match the name of the directory in which the meta.yaml file is placed. It can only contain alpha-numeric characters and -, _.

package/version

The version of the package.

source

source/url

The url of the source tarball.

The tarball may be in any of the formats supported by Python's shutil.unpack_archive: tar, gztar, bztar, xztar, and zip.

source/path

Alternatively to source/url, a relative or absolute path can be specified as package source. This is useful for local testing or building packages which are not available online in the required format.

If a path is specified, any provided checksums are ignored.

source/md5

The MD5 checksum of the tarball. It is recommended to use SHA256 instead of MD5. At most one checksum entry should be provided per package.

source/sha256

The SHA256 checksum of the tarball. It is recommended to use SHA256 instead of MD5. At most one checksum entry should be provided per package.

source/patches

A list of patch files to apply after expanding the tarball. These are applied using patch -p1 from the root of the source tree.

source/extras

Extra files to add to the source tree. This should be a list where each entry is a pair of the form (src, dst). The src path is relative to the directory in which the meta.yaml file resides. The dst path is relative to the root of source tree (the expanded tarball).

build

build/skip_host

Skip building C extensions for the host environment. Default: True.

Setting this to False will result in ~2x slower builds for packages that include C extensions. It should only be needed when a package is a build time dependency for other packages. For instance, numpy is imported during installation of matplotlib, importing numpy also imports included C extensions, therefore it is built both for host and target.

build/cflags

Extra arguments to pass to the compiler when building for WebAssembly.

(This key is not in the Conda spec).

build/ldflags

Extra arguments to pass to the linker when building for WebAssembly.

(This key is not in the Conda spec).

build/post

Shell commands to run after building the library. These are run inside of bash, and there are two special environment variables defined:

  • $SITEPACKAGES: The site-packages directory into which the package has been installed.
  • $PKGDIR: The directory in which the meta.yaml file resides.

(This key is not in the Conda spec).

requirements

requirements/run

A list of required packages.

(Unlike conda, this only supports package names, not versions).

C library dependencies

Some python packages depend on certain C libraries, e.g. lxml depends on libxml.

To package a C library, create a directory in packages/ for the C library. This directory should contain (at least) two files:

  • Makefile that specifies how the library should be be built. Note that the build system will call make, not emmake make. The convention is that the source for the library is downloaded by the Makefile, as opposed to being included in the pyodide repository.

  • meta.yaml that specifies metadata about the package. For C libraries, only three options are supported:

    • package/name: The name of the library, which must equal the directory name.
    • requirements/run: The dependencies of the library, which can include both C libraries and python packages.
    • build/library: This must be set to true to indicate that this is a library and not an ordinary package.

After packaging a C library, it can be added as a dependency of a python package like a normal dependency. See lxml and libxml for an example (and also scipy and CLAPACK).

Remark: Certain C libraries come as emscripten ports, and do not have to be built manually. They can be used by adding e.g. -s USE_ZLIB in the cflags of the python package. See e.g. matplotlib for an example.

Manual creation of a Pyodide package (advanced)

The previous sections describes how to add a python package to the pyodide build.

There are cases where you want to ship additional python libraries without adding it to pyodide itself. For pure python packages, this can be achieved reasonably easily. The most straightforward way is to create a Python wheel and load it with micropip. Alternatively, we can construct a python package manually.

It is helpful to have some understanding of the structure of a Pyodide package. Pyodide is obtained by compiling CPython into web assembly. As such, it loads packages the same way as CPython --- it looks for relevant files .py files in /lib/python3.x/. When creating and loading a package, our job is to put our .py files in the right location in emscripten's virtual filesystem.

Suppose you have a python library that consists of a single directory /PATH/TO/LIB/ whose contents would go into /lib/python3.8/site-packages/PACKAGE_NAME/ under a normal python installation.

The simplest version of the corresponding Pyodide package contains two files --- PACKAGE_NAME.data and PACKAGE_NAME.js. The first file PACKAGE_NAME.data is a concatenation of all contents of /PATH/TO/LIB. When loading the package via pyodide.loadPackage, Pyodide will load and run PACKAGE_NAME.js. The script then fetches PACKAGE_NAME.data and extracts the contents to emscripten's virtual filesystem. Afterwards, since the files are now in /lib/python3.8/, running import PACKAGE_NAME in python will successfully import the module as usual.

To produce these files, download the file_packager.py script from https://github.com/iodide-project/pyodide/blob/master/tools/file_packager.py. You then run the command

$ ./file_packager.py PACKAGE_NAME.data --js-output=PACKAGE_NAME.js --abi=1 --export-name=pyodide._module --use-preload-plugins --preload /PATH/TO/LIB/@/lib/python3.8/site-packages/PACKAGE_NAME/ --exclude "*__pycache__*"

The --preload argument instructs the package to look for the file/directory before the separator @ (namely /PATH/TO/LIB/) and place it at the path after the @ in the virtual filesystem (namely /lib/python3.8/site-packages/PACKAGE_NAME/). Remember to use the correct python version in the target path. At the time of writing, the latest release of Pyodide uses python 3.7 while git master uses python 3.8.

The --exclude argument specifies files to omit from the package. This argument can be repeated, e.g. you can append --exclude README.md to the command.

Remark. The bundled Pyodide packages uses lz4 compression when producing PACKAGE_NAME.data. These instructions skip this step as it requires additional dependencies, which complicates the process. In general, lz4 compression decreases memory usage and can increase performance. On the other hand, if your webserver serves the files with gzip compression, pre-compressing with lz4 could in fact increase the number of bytes transferred.