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 type, 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.
Simplification of the package loading system¶
Currently Pyodide has two way of loading packages:
loadPackagefor packages built with Pyodide and
micropip.installfor pure Python packages from PyPi.
The relationship between these tools is confusing and could be simplified. Furthermore, the dependency resolution logic and packaging / build system could be improved.
Update SciPy to a more recent version¶
SciPy is a cornerstone of scientific computing in Python. It’s a challenging package to build for WebAssembly because it is large, includes Fortran code, and requires BLAS and LAPACK libraries. Currently Pyodide includes scipy 0.17.1 from 2016. Updating it is a blocker for using more recent versions of packages such as scikit-learn, scikit-image, statsmodels, and MNE.
See issue #549.
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. Luckily @joemarshall has a solution for this involving rewinding the Python stack. He has a prototype implementation here. We would like to bring this into Pyodide as a core feature.
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 of the
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.