Skip to content

Hosting

Where each part of plainml can live online:

What Where Why
The website, in the browser Vercel (free), or any static host plainml runs in the visitor's browser, so plain files are enough
This documentation Vercel, next to the website Static pages, rebuilt on every push
The Python package PyPI pip install plainml. Release steps are in RELEASING.md
The website, on a server Render, Railway, Fly.io For big files and long jobs: see below
A trained model's API Any container host plainml deploy builds the image

The website in the browser

plainml web --export DIR writes the website as static files: no server needed. When someone opens it, a background worker in their browser starts Python (via Pyodide), installs plainml, and does all the work right there: - Free to host anywhere: Vercel, GitHub Pages, Netlify, a Hugging Face Static Space. - Private: visitors' data never leaves their computer. - Always on: it never sleeps, and runs are saved in each visitor's browser between visits. - Complete: every task works, including LightGBM and XGBoost. - Trade-offs: - The first visit downloads about 50 MB of Python libraries (cached afterwards). - Uploads are limited to 200 MB. - Training uses one CPU core, so it's slower than the command line on big data.

From a plainml source checkout, the export includes plainml built from that code. Otherwise the browser installs the same plainml version from PyPI.

To try it locally:

plainml web --export site

python -m http.server --directory site

Then open http://localhost:8000.

On Vercel (website and docs)

The repository's vercel.json builds both parts: the website at the root of the address and this documentation under /docs. plainml's own site is deployed exactly this way, at plainml-tpua.vercel.app.

{
  "framework": null,
  "installCommand": "python3 -m venv .venv && .venv/bin/python -m pip install --quiet \"mkdocs-material>=9.5\" \"mkdocs<2\"",
  "buildCommand": ".venv/bin/python -m plainml.web.static_site site && .venv/bin/python -m mkdocs build -d site/docs",
  "outputDirectory": "site"
}

The build tools go into a throwaway virtual environment (.venv) because Vercel's own Python refuses package installs (it's managed by uv, per PEP 668).

  1. Sign in at vercel.com with GitHub.
  2. Choose Add New → Project and import pranay-obla/plainml. The settings come from vercel.json, so you don't need to change anything.
  3. Click Deploy. The site appears at a *.vercel.app address, with the docs at /docs. To use your own domain, go to Settings → Domains.
  4. Every push to main redeploys both, and pull requests get their own preview link.
  5. Put the addresses in mkdocs.yml (site_url: https://…/docs/) and in the [project.urls] of pyproject.toml (Homepage and Documentation).

The documentation can also go to GitHub Pages with .github/workflows/docs.yml. That workflow only runs when started by hand from the Actions tab. To use it, turn on Settings → Pages → Source: GitHub Actions.

On a Hugging Face Static Space

Static Spaces are free on Hugging Face.

  1. Export the website:
plainml web --export space
  1. Copy hosting/huggingface/README.md into the space folder. Its header tells Hugging Face to serve the folder as a static site.
  2. On huggingface.co, choose New Space, give it a name, and choose Static with a Blank template.
  3. Upload everything in space/ to it (Files → Add file → Upload files, keeping the wheels folder), or git push it to the Space's repository.

Export again and re-upload whenever you want a newer plainml.

On a server (Render, Railway, Fly.io)

The server version (plainml web) suits big files and long jobs, and keeps runs on the server. hosting/docker/Dockerfile runs it on any host that runs containers: - Tell the host the app listens on port 7860. - Set PLAINML_WEB_TOKEN as a secret environment variable to require an access token. - Attach a persistent disk at /data so runs survive restarts and redeploys. - Upgrade by changing PLAINML_VERSION in the Dockerfile.

Render has a free plan that sleeps after 15 minutes without visitors and has no disk. That's fine for a demo, since runs are lost on restart. An always-on instance with a disk costs a few dollars a month.

The server version can't run on Vercel itself: Vercel runs Python only as short request handlers, with 4.5 MB request bodies, at most 5 minutes per request on the free plan, and nothing kept between requests. That's why Vercel hosts the in-browser version instead.

A model's prediction API

plainml deploy MODEL writes a Docker build folder for one trained model's REST API (see Deploying). Build and push that image to any container host: Google Cloud Run, AWS App Runner, Azure Container Apps, Fly.io, Render or Kubernetes. Set PLAINML_API_KEY there to require an API key.