a smol (<1k loc) faas platform. inspired by CloudFlare Workers/Vercel Edge Functions. adventures in embedded V8.
try it out here: https://smolfaas.jchap.me/docs#/default/invoke_source_invoke_post
this is an educational project aimed at better understanding V8 and serverless compute. also checked another box for me that i've been interested in which is doing a python project with a native extension
- FastAPI API
- embedded V8 runtime
- uses
scikit-build-coreandpybind11for c++ <> python bridge - SQLite for storing function/deployment data and source code
- a "function" is an ESM module with a default export that is invoked as the function handler
- functions are invoked in V8 isolates
- isolates are kept warm and reused for subsequent invocations of the same functions (ie lambda cold/warm starts)
- warm isolates are kept in an lru cache (128 of them) - arbitrary and simple
- invocations will always hit the latest deployment of a function (even with a warm function running - it'll get evicted)
make sure you have uv and homebrew installed
gh repo clone thejchap/smolfaas
cd smolfaas
cd server
# get up and running with a mac build of v8 from brew
make
# alternatively, build+run with docker
make docker- getting cmake working with vscode was funky - uv setup/
uv runworked out of the box, needed to tweak cmakelists a bit for vscode to pick up pybind11 - i've only written a little c++ professionally, and no rust professionally, but have worked on i think what qualifies as a non-trivial project in rust (implementing a programming language), and c++ scares me - it fails in such funky, unhelpful ways. i'm sure there's tooling to help protect you from the footguns, but i don't know what it is. segfaults with no error messages, super unhelpful error messages, stuff like that. rust on the other hand is harder to get it to compile, but once it compiles you can be pretty sure it's gonna fail exactly where and how you think its gonna fail and be loud and obvious about it.
- was an adventure getting v8 built for linux (docker) and getting cmake
configured to build on both platforms. homebrew packages an older build of
v8 that is NOT the one that ships with
libnode-devwhich was the easiest way to install on linux, and smolfaas wouldn't compile on the latter due to some breaking api changes between the 2 versions. i ended up building from source off a release branch for both platforms. side note - building v8 was the first time i'd heard my fan turn on on my m4 pro. also kept running into "cannot allocate memory" from docker - realize i had a resource limit set, removed that and it worked but was slow. this was when building for amd64 as a result of this.
- log from functions
- ship cli to pypi (ditching cli for now)
- blow up if homebrew v8 version not docker version
- sqlite docker
- check how promises (dont) work
- docker
- invoke from stdin
- resource limits
- exception handling
- json to and from invocations
- cache compiled module
- store deployments
- logging
- https://chromium.googlesource.com/v8/v8.git/+/4.5.103.9/test/cctest/test-serialize.cc#661
- https://v8.dev/blog/custom-startup-snapshots
- https://v8.dev/blog/code-caching
- https://chromium.googlesource.com/v8/v8/+/branch-heads/11.9/samples/hello-world.cc
- https://blog.cloudflare.com/cloud-computing-without-containers/
- https://johnfoster.pge.utexas.edu/blog/posts/debugging-cc%2B%2B-libraries-called-by-python/
- https://stackoverflow.com/questions/61490100/debugging-pybind11-extension-with-visual-studio-code-macos
- https://github.com/cloudflare/workerd/blob/00d8c87890edb836bfd1445157f2a3960eb5bc5e/src/workerd/jsg/setup.c%2B%2B#L155
- https://chromium.googlesource.com/v8/v8/+/branch-heads/11.9/samples/hello-world.cc
- https://gist.github.com/surusek/4c05e4dcac6b82d18a1a28e6742fc23e?permalink_comment_id=4472429
- https://github.com/nodejs/node-v0.x-archive/blob/master/lib/console.js
- rogchap/v8go#308
- https://v8.github.io/api/head/classv8_1_1ResourceConstraints.html
