nbstripout-fast

A much faster version of nbstripout by writing it in rust (of course). This helps strip Jupyter Notebook output and metadata from notebooks. It is very useful as a git filter and is highly configurable.

Installation

pip install nbstripout-fast

Then replace nbstripout-fast with anywhere you use nbstripout.

Key differences

1. While we mirrored most of nbstripout's API, we do not support every nbstripout option.
2. There is no CLI option to install this in git for you
3. We support repository level settings in a .git-nbconfig.yaml file. Check out our examples. On a high level, you can add a git filter in a sitewide/user level and then allow each project to enforce consistent settings.

Why Rust?

nbstripout is a excellent project, but the python startup and import time makes its usage at scale a bit painful. While this means giving up on using nbconvert under the hood and ensuring the notebook is the correct format, it does make things up to 200x faster. This matters when you have a large number of files and git filter is called sometimes more than once per file. Let's look at the data:

The table above shows a large overhead per notebook (mostly python startup time). When you have 100 or more notebooks, nbstripout takes more than 40s while nbstripout-fast takes only 1s!

Example

This example illustrates how nbstripout-fast can be used to automatically clean Jupyter notebooks using Git filters (see e.g. Git Attributes). This keeps your repository clean by removing unnecessary output and clutter, while preserving your local working version. The benefits are minimised diffs and reduced repository size.

1. Install nbstripout-fast as described above.

2. Configure nbstripout-fast

   Create a .git-nbconfig.yaml file at the root of your repository to configure nbstripout-fast, e.g.

   nbstripout_fast:
     keep_count: false
     keep_output: false
     drop_empty_cells: true
     extra_keys: []
     keep_keys: []
   

3. Set Git Attributes

   Create a .gitattributes file at the root of your repository if it doesn't yet exist and add this line:

   *.ipynb filter=jupyter
   

   This instructs Git to use a custom filter named "jupyter" on all .ipynb files.

4. Configure the jupyter Filter

   Run these commands in your terminal to configure the "jupyter" filter:

   git config filter.jupyter.clean nbstripout-fast
   git config filter.jupyter.smudge cat
   

• clean: This filter runs nbstripout-fast when adding notebooks to the version that is checked out, i.e. the clean version.
• smudge: This filter runs cat when checking out notebooks, ensuring your local (smudged) version remains unmodified. Git filters transform files at the time of checkout and commit.

4. Reapply Cleaning to Existing Notebooks (Optional)

   If you already have Jupyter notebooks tracked by Git, you can reapply the cleaning process to them:

   git add --renormalize . git commit -m "Cleaned Jupyter notebooks"
   

Stripping specific cell outputs

To strip cell outputs that match a regular expression, the --strip-regex option can be used in combination with --keep-output. For example, to remove cell outputs that only contain a notebook widget:

nbstripout-fast --keep-output --strip-regex "^Output\(\)$"

or to remove completed tqdm progress bars:

nbstripout-fast --keep-output --strip-regex "100%.*"

See the documentation for regex for information about supported regex syntax.

Developing

You can use cargo which will build + run the CLI:

cargo run -- -t examples/example.ipynb

You can also build with cargo and run the script with the full path:

cargo build # dev build - ./target/debug/nbstripout-fast
cargo build --release # release build - ./target/release/nbstripout-fast

Running unit tests: maturin builds this repo to include pyo3 bindings by default. This allows for us to have an extension python extension mode as well. As of today, we can't have a binary and an extension, so we use the extension only for testing (issue).

pip install -e .
maturin develop
# Should output, this way you can use RUST_LOG=debug
in-venv pytest -rP

Debugging

Use RUST_LOG=debug to debug script for example:

RUST_LOG=debug cargo run -- '--extra-keys "metadata.bar cell.baz" -t foo.ipynb'

Releasing

Manylinux, macos, and windows wheels and sdist are built by github workflows. Builds are triggered upon the creation of a pull request, creating a new release, or with a manual workflow dispatch. The wheels and sdist are only uploaded to PyPI when a new release is published. In order to create a new release:

1. Create a commit updating the version in Cargo.toml and CHANGELOG.md, then create a git tag:

git tag vX.Y.Z
git push --tags

2. Draft a new release in github; select the tag that you just created.
3. Once the new release is created, the wheels and sdist will be built by a github workflow and then uploaded to PyPI automatically using the PYPI_API_TOKEN in the github secrets for the repository.

History

This plugin was contributed back to the community by the D. E. Shaw group.

License

This project is released under a BSD-3-Clause license.

We love contributions! Before you can contribute, please sign and submit this Contributor License Agreement (CLA). This CLA is in place to protect all users of this project.