This document lists general directions that core developers are interested to see developed in Pyodide. The fact that an item is listed here is in no way a promise that it will happen, as resources are limited. Rather, it is an indication that help is welcomed on this topic.
Reducing download sizes and initialization times#
At present a first load of Pyodide requires a 6.4 MB download, and the environment initialization takes 4 to 5 seconds. Subsequent page loads are faster since assets are cached in the browser. Both of these indicators can likely be improved, by optimizing compilation parameters, minifying the Python standard library and packages, reducing the number of exported symbols. To figure out where to devote the effort, we need a better profiling system for the load process.
See issue #646.
Improve performance of Python code in Pyodide#
Across benchmarks Pyodide is currently around 3x to 5x slower than native Python.
At the same time, C code compiled to WebAssembly typically runs between near native speed and 2x to 2.5x times slower (Jangda et al. 2019 PDF). It is therefore very likely that the performance of Python code in Pyodide can be improved with some focused effort.
In addition, scientific Python code would benefit from packaging a high performance BLAS library such as BLIS.
See issue #1120.
Better support and documentation for loading user Python code#
Currently, most of our documentation suggests using
pyodide.runPython to run
code. This makes code difficult to maintain, because it won’t work with
black, or other code analysis tools, doesn’t get good syntax highlighting in
editors, etc. It also may lead to passing “arguments” to code via string
formatting, missing out on the type conversion utilities.
Our goal is to develop and document a better workflow for users to develop Python code for use in Pyodide.
See issue #1940.
Improvements to package loading system#
Currently, Pyodide has two ways of loading packages:
pyodide.loadPackagefor packages built with Pyodide and
micropip.installfor pure Python packages from PyPI.
The relationship between these tools is currently confusing.
Our goal is to have three ways to load packages: one with no dependency resolution at all, one with static dependency resolution which is done ahead of time, and one for dynamic dependency resolution. Ideally most applications can use static dependency resolution and repls can use dynamic dependency resolution.
Find a better way to compile Fortran#
Currently, we use f2c to cross compile Fortran to C. This does not work very well because f2c only fully supports Fortran 77 code. LAPACK has used more modern Fortran features since 2008 and Scipy has adopted more recent Fortran as well. f2c still successfully generates code for all but 6 functions in Scipy + LAPACK, but much of the generated code is slightly wrong and requires extensive patching. There are still a large number of fatal errors due to call signature incompatibilities.
If we could use an LLVM-based Fortran compiler as a part of the Emscripten toolchain, most of these problems would be solved. There are several promising projects heading in that direction including flang and lfortran.
Better project sustainability#
See issue #795.
Improve support for WebWorkers#
WebWorkers are necessary in order to run computational tasks in the browser without hanging the user interface. Currently, Pyodide can run in a WebWorker, however the user experience and reliability can be improved.
See issue #1504.
The majority of existing I/O APIs are synchronous. Unless we can support synchronous IO, much of the existing Python ecosystem cannot be ported. There are several different approaches to this, we would like to support at least one method.
See issue #1503.
Write http.client in terms of Web APIs#
Python packages make an extensive use of packages such as
synchronously fetch data. We currently can’t use such packages since sockets
are not available in Pyodide. We could however try to re-implement some
stdlib libraries with Web APIs, potentially making this possible.
Because http.client is a synchronous API, we first need support for synchronous IO.
See issue #140.