CoWasm: Collaborative WebAssembly for Servers and Browsers
URL: https://github.com/sagemathinc/cowasm
DEMOS:
- https://cowasm.org - Python using SharedArrayBuffers
- https://zython.org - Python using Service Workers
- https://cowasm.sh - Dash Shell with Python, Sqlite, Lua, etc., using SharedArrayBuffers
TEST STATUS:
Or Type this if you have nodejs at least version 16 installed:
~$ npx python-wasm@latest
Python 3.11.0 (main, Oct 27 2022, 10:03:11) [Clang 15.0.3 ([email protected]:ziglang/zig-bootstrap.git 0ce789d0f7a4d89fdc4d9571 on wasi
Type "help", "copyright", "credits" or "license" for more information.
>>> 2 + 3
5
>>> import numpy
>>> import sympy
...
NOTE: On Microsoft Windows you have to enter a few times, and there is an issue with terminal echo. We are working on this.
Packages
There are four subdirectories that each contain packages:
- core - core functionality, including dynamic linking, a WASI implementation, openssl, and nontrivial posix extensions to node.js written in Zig. (This is analogous to emscripten.org.)
- python - a WebAssembly build of Python, along with some nontrivial scientific libraries (This is analogous to pyodide.org.)
- web - examples of how to use the other packages in web applications; this includes building https://cowasm.org.
- desktop - a native Electron app that provides a sandboxed WebAssembly Python terminal running on your native filesystem.
- sagemath - beginning of a port of https://sagemath.org to WebAssembly. Vasts amounts of work remain.
What is this?
CoWasm means "collaborative WebAssembly", and goes far beyond just Python. The collaboration aspects will be part of https://cocalc.com at some point in the future. CoWasm will support various technologies (such as libgit2 and realtime sync) that are important foundations for collaboration.
The underlying software components that CoWasm is built on (i.e., that we didn't write) are mostly extremely stable and mature. Zig is less stable, but we mostly use Zig for its amazing cross compilation support and packaging of clang/llvm and musl-libc, which are themselves both very mature. Many other components, such as Python, Dash, Numpy, etc., are ridiculously mature multidecade old projects. Morever, other components of CoWasm such as memfs are libraries with 10M+ downloads per week that are heavily used in production.
The goal of CoWasm is overall somewhat similar to all of emscripten, WebAssembly.sh, wapm.io, and Pyodide combined, in various ways.
- Unlike WebAssembly.sh and wapm.io (but similar to Pyodide), we make VERY heavy use of shared dynamic libraries (e.g.,
-fPIC
code), which is only possible because of a plugin contributed from emscripten to LLVM. The "Co" in CoWasm suggestion "collaboration" or "sharing", which also reflects how the binaries in this project are structured. - We use actual editline (similar to readline) instead of a Javascript terminal. Moreover, unlike other webassembly shells, we just use a real command line shell (dash = Debian Almquest Shell). We also have a userspace including ports of many coreutils, e.g., ls, head, tail, etc.
- Unlike emscripten, we use modern Typescript, our code is more modular, and we make use of existing components when possible (e.g., the nodejs memfs project), instead of using our own.
- A core design constraint is to efficiently run on a wide range of platforms, not mainly in the browser like emscripten, and not mainly on servers like wasmer. CoWasm should run on servers, desktops (e.g., as an electron app), an iPad/iOS app, and in web browsers.
- There is no business unfriendly GPL code in CoWasm. CoWasm itself is extremely liberally licensed and business friendly. The license of all new code and most components is 3-clause BSD. CoWasm will serve as a foundation for other projects with more restrictive licenses:
- CoCalc will build on top of CoWasm to provide a graphical interface and realtime collaboration, and that will be a commercial product.
- Products like GP/PARI SageMath will build on CoWasm to provide GPL-licensed mathematics software.
Python
An exciting package in CoWasm is python-wasm, which is a build of Python for WebAssembly, which supports both servers and browsers. It also supports extension modules such as numpy. See python/README.md for more details.
Try python-wasm
Try the python-wasm REPL under node.js (version at least 16):
~$ npx python-wasm@latest
Python 3.11.0 (main, Oct 27 2022, 10:03:11) [Clang 15.0.3 (git@github.com:ziglang/zig-bootstrap.git 0ce789d0f7a4d89fdc4d9571 on wasi
Type "help", "copyright", "credits" or "license" for more information.
>>> 2 + 3
5
>>> import sys; sys.version
'3.11.0 (main, Oct 27 2022, 10:03:11) [Clang 15.0.3 (git@github.com:ziglang/zig-bootstrap.git 0ce789d0f7a4d89fdc4d9571'
>>> sys.platform
'wasi'
Install python-wasm
Install python-wasm into your project, and try it via the library interface and the node.js terminal.
npm install python-wasm
Then from the nodejs REPL:
~/cowasm/core/python-wasm$ node
Welcome to Node.js v19.0.0.
Type ".help" for more information.
> {syncPython, asyncPython} = require('.')
{
syncPython: [AsyncFunction: syncPython],
asyncPython: [AsyncFunction: asyncPython],
default: [AsyncFunction: asyncPython]
}
> python = await syncPython(); 0;
0
> python.exec('import sys')
undefined
> python.repr('sys.version')
"'3.11.0b3 (main, Jul 14 2022, 22:22:40) [Clang 13.0.1 ([email protected]:ziglang/zig-bootstrap.git 623481199fe17f4311cbdbbf'"
> python.exec('import numpy')
undefined
> python.repr('numpy.linspace(0, 10, num=5)')
'array([ 0. , 2.5, 5. , 7.5, 10. ])'
There is also a Python REPL that is part of python-wasm:
> python.terminal()
Python 3.11.0 (main, Oct 27 2022, 10:03:11) [Clang 15.0.3 (git@github.com:ziglang/zig-bootstrap.git 0ce789d0f7a4d89fdc4d9571 on wasi
Type "help", "copyright", "credits" or "license" for more information.
>>> 2 + 3 # you can edit using readline
5
>>> input('name? ')*3
name? william <-- I just typed "william"
'williamwilliamwilliam'
>>> import time; time.sleep(3) # sleep works
>>> while True: pass # ctrl+c works, but exits this terminal
> # back in node.
You can also use python-wasm in your own web applications via webpack. In the browser, this transparently uses SharedArrayBuffers if available, and falls back to ServiceWorkers.
Build from Source
We support and regularly test building CoWasm from source on the following platforms:
- x86 macOS and aarch64 macOS (Apple Silicon M1)
- x86 and aarch64 Linux
Prerequisites
You need Node.js version at least 16.x, pnpm and several standard dev tools listed below. The dependency you need for every possible package are as follows:
-
On MacOS, install the XCode command line tools.
-
On Linux apt-based system, e.g., on Ubuntu 22.04:
sudo apt-get install git make cmake curl dpkg-dev m4 yasm texinfo python-is-python3 libtool tcl zip libncurses-dev
If you also want to install node v18 and pnpm on Ubuntu, you could do:
sudo curl -sL https://deb.nodesource.com/setup_18.x | bash - \
&& sudo apt-get install -y nodejs \
&& sudo npm install -g npm pnpm
- On Linux RPM based system, e.g., Fedora 37:
dnf install git make cmake curl dpkg-dev m4 yasm texinfo libtool tcl zip ncurses-devel perl
- On ArchLinux:
pacman -Sy binutils git nodejs npm cmake curl m4 yasm texinfo python libtool tcl zip unzip patch binutils diffutils
NOTE: pnpm
is not in the Arch Linux official package sets. You may want to install it from Arch User Repository (AUR) AUR (en) - pnpm.
- Currently, the only way to build CoWasm from source on MS Windows is to use a Docker container running Linux. Using WSL2 (maybe) works but is too slow.
In addition you need to install Node.js version at least 16.x and install the pnpm package manager.
Notes about Zig
You do NOT need to install Zig, and it doesn't matter if you have a random version of Zig on your system already. A zig binary is download automatically. We use zig (instead of any system-wide clang, etc. compilers) for building all compiled code and write some code in the zig language. Since zig is fairly unstable it is critical to use the exact version that we provide.
Build
To build any package in the src/packages
directory, cd into that directory, then:
make
That's it, usually. You do not have to run build at the top level and you can start with building any specific package -- it will automatically cause any required dependencies to get installed or built.
You can also force building of every single package and running its test suite if you want:
~/cowasm$ make test
...
##########################################################
# #
# CONGRATULATIONS -- FULL COWASM TEST SUITE PASSED! #
#
# Fri Oct 28 12:32:19 AM UTC 2022
# Linux aarch64
# Git Branch: main
# #
##########################################################
Depending on your computer, the build should take less than 30 minutes, and about 6GB's of disk space.
This installs a specific tested version of Zig and Nodejs, then builds native and WebAssembly versions of CPython and many dependencies, and also builds all the Typescript code. It also builds many other interesting programs with ports to WebAssembly, though many are not completely finished (e.g., there is the dash shell and ports of tar and BSD coreutils). As mentioned, building from source is regularly tested on Linux and MacOS with both x86_64 and ARM (M1) processors:
- Linux: tested on both x86_64 and aarch64 Ubuntu
- MacOS: tested on both x86_64 and M1 mac with standard XCode command live dev tools installed.
CoWasm does not use the compilers on the system, and instead uses clang/llvm as shipped with Zig. If you're using Windows, you'll have to use Linux via a virtual machine or Docker.
Pull latest code, build and test
At the top level run ./bin/rebuild-all
:
~/cowasm$ ./bin/rebuild-all
This does make clean,
pulls the latest code, then runs the full build and test suite. Fortunately, zig
caches a lot of build artifacts, so subsequent builds are faster.
NOTE/WARNING: Zig's cache is in ~/.cache/zig
and it can get HUGE. As far as I can tell, I think it just grows and grows without bound (it's not an LRU cache), and I think there are no tools to "manage" it besides just rm -rf
it periodically.
What is tested?
Note that running make test
at the top level does NOT run the full test suite of every package, since it takes quite a while and there are still some failing tests, since CoWasm doesn't support enough of what Python expects. It does run a large supported subset of the cpython test suite (it's the part that I got to pass so far, which is over 80%). As an example, the sympy test suite is massive, takes a very long time to run, and doesn't even work for me natively; instead, we just run a handful of small tests to ensure sympy is working at all. Similarly, for Cython, we run all their demos, but not the full test suite. A longer term goal of CoWasm is to support a second more thorough testing regime that runs the full test suite of each package. There will likely always be issues due to WASM not being a multiuser POSIX system, but it's good to know what those issues are!
You can also use the WebAssembly Python REPL directly on the command line.
~/cowasm$ ./bin/python-wasm
Python 3.11.0 (main, Oct 27 2022, 10:03:11) [Clang 15.0.3 ([email protected]:ziglang/zig-bootstrap.git 0ce789d0f7a4d89fdc4d9571 on wasi
Type "help", "copyright", "credits" or "license" for more information.
>>> 2 + 3
5
>>> import sys
>>> sys.platform
'wasi'
>>> sys.executable
'/Users/wstein/build/cocalc/src/data/projects/2c9318d1-4f8b-4910-8da7-68a965514c95/cowasm/core/cpython/bin/python-wasm'
>>> ^D
~/cowasm$
The above directly runs the `python.wasm` executable produced by building cPython. You can instead run an enhanced version (e.g., with signal support) with more options in the python-wasm package:
~/cowasm$ . bin/env.sh
~/cowasm$ cd core/python-wasm/
~/cowasm/core/python-wasm$ ./bin/python-wasm
Python 3.11.0 (main, Oct 27 2022, 10:03:11) [Clang 15.0.3 (git@github.com:ziglang/zig-bootstrap.git 0ce789d0f7a4d89fdc4d9571 on wasi
Type "help", "copyright", "credits" or "license" for more information.
>>> import time; t=time.time(); print(sum(range(10**7)), time.time()-t)
49999995000000 0.8989999294281006
>>> ^D
~/cowasm/core/python-wasm$
As mentioned above, you can use python-wasm as a library in node.js. There is a synchronous api that runs in the same thread as the import, and an asynchronous api that runs in a worker thread.
~/cowasm$ . bin/env.sh
~/cowasm$ cd core/python-wasm/
~/cowasm/core/python-wasm$ node
Welcome to Node.js v19.0.0.
Type ".help" for more information.
> python = require('.')
{
syncPython: [AsyncFunction: syncPython],
asyncPython: [AsyncFunction: asyncPython],
default: [AsyncFunction: asyncPython]
}
> const { exec, repr} = await python.asyncPython();
undefined
> await repr('31**37')
'15148954872646847105498509334067131813327318808179940511'
> await exec('import time; t=time.time(); print(sum(range(10**7)), time.time()-t)')
49999995000000 0.8880000114440918
And yes you can run many async Python's in parallel in the same node.js process, with each running in its own thread:
~/cowasm/core/python-wasm$ nodeWelcome to Node.js v19.0.0.
Type ".help" for more information.
> python = require('.')
{
syncPython: [AsyncFunction: syncPython],
asyncPython: [AsyncFunction: asyncPython],
default: [AsyncFunction: asyncPython]
}
> const { exec, repr} = await python.asyncPython();
undefined
> await exec('import time; t=time.time(); print(sum(range(10**7)), time.time()-t)')
49999995000000 0.8880000114440918
undefined
>
> const v = [await python.asyncPython(), await python.asyncPython(), await python.asyncPython()]; 0;
0
> s = 'import time; t=time.time(); print(sum(range(10**7)), time.time()-t)'
'import time; t=time.time(); print(sum(range(10**7)), time.time()-t)'
> d = new Date(); console.log(await Promise.all([v[0].exec(s), v[1].exec(s), v[2].exec(s)]), new Date() - d)
49999995000000 0.8919999599456787
49999995000000 0.8959999084472656
49999995000000 0.8929998874664307
[ undefined, undefined, undefined ] 905
undefined
What's the goal?
Our initial goal is to create a WebAssembly build of the core Python and dependent packages, which runs both on the command line with Node.js and in the major web browsers (via npm modules that you can include via webpack). This is done. It should also be relatively easy to build from source on both Linux and MacOS (x86_64 and aarch64) and to easily run the cpython test suite,_ with a clearly defined supported list of passing tests. The compilation system is based on Zig, which provides excellent caching and cross compilation, and each package is built using make.
How does CoWasm compare to Emscripten and Pyodide?
Pyodide currently provides far more packages. However, there is no reason that CoWasm couldn't eventually support as much or more than Pyodide.
Our main longterm application is to make CoCalc available on a much wider range of computers. As such, we are building a foundation here on which to support a substantial part of the scientific Python ecosystem and the SageMath packages (a pure math analogue of the scientific Python stack). I'm the founder of SageMath, hence this motivation (some relevant work has been done here).
Some of our code will be written in the Zig language. However, we are mostly targeting just the parts that are used for Python, which is a small subset of the general problem. Our software license -- BSD 3-clause -- is compatible with their's and we hope to at least learn from their solutions to problems.
More about how Pyodide and python-wasm differ...
More about building from source
How to build
Just type make. (Do NOT type make -j8;
it probably won't work.)
...$ make
This runs a single top level Makefile that builds all the packages. The build process for each individual package is also accomplished using a Makefile with two includes that impose some structure. We don't use shell scripts or Python code to orchestrate building anything, since make
is much cleaner and easier to read, maintain and debug... and of course make contains shell scripts in it. (History lesson: homebrew is a lot more successful than Macports.)
What happens
In most subdirectories foo/
of core
, this makefile creates some subdirectories:
core/foo/dist/[native|wasm]
-- a native or WebAssembly build of the package; this has binaries, headers, and libs. These get used by other packages. We rarely build the native version.core/build/[native|wasm]
- build artifacts for the native or WebAssembly build; can be safely deleted
No common prefix directory
Unlike some systems, where everything is built and installed into a single prefix
directory, here we build everything in its own self-contained package dist directory. When a package like cpython
depends on another package like lzma
, our Makefile for cpython
explicitly references packages/lzma/dist
. This makes it easier to uninstall packages, update them, etc., without having to track what files are in any package, whereas using a common directory for everything can be a mess with possibly conflicting versions of files, and makes versioning and dependencies very explicit. Of course, it makes the environment variables and build commands potentially much longer. In some cases, we gather together files from these dist directories in distributions, e.g., see make bin-wasm.
Native and Wasm
The build typically create directories dist/native
and dist/wasm.
The dist/native
artifacts are only of value on the computer where you ran the build, since they are architecture dependent and can easily depend on libraries on your system. In contrast, the dist/wasm
artifacts are platform independent. They can be used nearly everywhere: on servers via WASM, on ARM computers (e.g., aarch64 linux, Apple Silicon, etc.), and in modern web browsers.
Standalone WASM executables
The bin directory has scripts zcc
and z++
that are C and C++ compiler wrappers around Zig + Node. They create binaries that you can run on the command line as normal. Under the hood there's a wrapper script that calls node.js and the wasi runtime.
$ . bin/env.sh
$ echo 'int main() { printf("hello from Web Assembly: %d\n", 2+2); }' > a.c
$ zcc a.c
$ ls -l
a.c a.out a.out.wasm ...
$ ./a.out # this actually runs nodejs + python-wasm
hello from Web Assembly: 4
This isn't currently used here for building python-wasm, but it's an extremely powerful tool. (For example, I used it with JSage to cross compile the NTL library to Web Assembly...)
Run a script from the terminal:
~/cowasm$ echo "import sys; print(f'hi from {sys.platform}')" > a.py
~/cowasm$ bin/python-wasm a.py
hi from wasi
The python-wasm package has a bin/python-wasm script that can run Python programs that including interactive blocking input:
~/cowasm/core/python-wasm$ echo "name = input('name? '); print(name*3)" > a.py
~/cowasm/core/python-wasm$ ./bin/python-wasm a.py
name? william
williamwilliamwilliam
Contact
Email [email protected] or @wstein389 if find this interesting and want to help out.
CoWasm is an open source 3-clause BSD licensed project. It includes components and dependencies that may be licensed in other ways, but nothing is GPL licensed, except some code in the sagemath subdirectory (which nothing else depends on).