How to create a dataset plugin (pip)

From the end-user’s perspective, a Dataset is a object created using cml.load_dataset(name, *args) with the appropriate name and arguments, which provides data.

From the plugin’s developer perspective, a Dataset is a Python class that inherits from the CliMetLab class climetlab.Dataset. This class contains the Python code providing specific helper functions and curated access to the data. Dataset can also be defined from yaml files if they have no specific Python code and rely on (yet to defined) standard conventions.

CliMetLab has build-in example datasets for demo purposes. And more examples can be found in the non-exhaustive list of CliMetLab plugins.


Naming convention

  • A plugin package name (pip) should preferably start with climetlab- and use dashes “-“.

  • The Python package to import should start with climetlab_ and must use underscores “_”.

  • A CliMetLab dataset defined by a plugin should start with the plugin name and must use dashes “-“.


Automatic creation script

While creating the package manually from the documentation and from the example above is possible, there is also a semi-automated way to generate a pip package from a template. The generated package has a predefined dataset and is ready to be shared on Github and distributed with pip.

$ pip install climetlab-plugin-tools
$ climetlab plugin_create_dataset
# Answer the questions
# Only the first question (plugin name) and the latest (licence) are required.
# Other have sensible default values.

Here is a verbose output of running the plugin creation script.


“Unknown command plugin_create_dataset” This error happens if you have not installed the package climetlab-plugin-tools.

$ climetlab plugin_create_dataset
Unknown command plugin_create_dataset. Type help for the list of known command names.

Dataset names

The plugin mechanism relies on using entry_points. The three lines highlighted below are registering the class climetlab_dataset_plugin.rain_observations:RainObservations with entry_points to a CliMetLab dataset called dataset-plugin-rain-observations. The python plugin mechanism is exhaustively described in the Python reference documentation and here are more details about how on CliMetLab uses it.

 description="Example climetlab external dataset plugin",



Once entry_point has registered the class, the end-user can use this external plugin to access it

import climetlab as cml

Automatic testing of the plugin

In the folder tests are set up automatic tests of the plugin using pytest. If the repository is hosted on github, the tests triggers automatically when pushing to the repository. Additionally, code quality is enabled using black, isort and flake.

All tests could be disabled or adapted in the .github/workflows/ folder.

Notebooks as documentation

The folder notebooks in each plugin can be used to store usage example or demo on how to use the data, such as this notebook, Notebook are automatically tested if the repository is on github.

Links on the README file are pointing to binder, colab, etc. to run the automatically created notebook.

Manually creating the Python package

Here is a minimal example of pip package defining a dataset plugin :

Adapting plugin code

Renaming a dataset

The dataset name can be changed by changing the file.

- ["dataset-plugin-rain-observations=climetlab_dataset_plugin.rain_observations:RainObservations"]
+ ["dataset-plugin-new-name         =climetlab_dataset_plugin.rain_observations:RainObservations"]

A good practice is to change keep the class name in sync with the dataset name.

Adding a dataset to a plugin

New datasets can be added to the plugin, as long as the corresponding class is created:

- ["dataset-plugin-rain-observations=climetlab_dataset_plugin.rain_observations:RainObservations"]
+ ["dataset-plugin-rain-observations=climetlab_dataset_plugin.rain_observations:RainObservations",
+  "dataset-plugin-rain-forecast    =climetlab_dataset_plugin.rain_observations:RainForecast"]

CliMetLab hooks


Document .source attribute, to_xarray(), to_pandas(), to_etc() Point to decorator