MEP25: Serialization#



This work is important, but this particular effort has stalled.

Branches and Pull requests#

  • development branches:

  • related pull requests:


This MEP aims at adding a serializable Controller objects to act as an Artist managers. Users would then communicate changes to an Artist via a Controller. In this way, functionality of the Controller objects may be added incrementally since each Artist is still responsible for drawing everything. The goal is to create an API that is usable both by graphing libraries requiring high-level descriptions of figures and libraries requiring low-level interpretations.

Detailed description#

Matplotlib is a core plotting engine with an API that many users already understand. It's difficult/impossible for other graphing libraries to (1) get a complete figure description, (2) output raw data from the figure object as the user has provided it, (3) understand the semantics of the figure objects without heuristics, and (4) give matplotlib a complete figure description to visualize. In addition, because an Artist has no conception of its own semantics within the figure, it's difficult to interact with them in a natural way.

In this sense, matplotlib will adopt a standard Model-View-Controller (MVC) framework. The Model will be the user defined data, style, and semantics. The Views are the ensemble of each individual Artist, which are responsible for producing the final image based on the model. The Controller will be the Controller object managing its set of Artist objects.

The Controller must be able to export the information that it's carrying about the figure on command, perhaps via a to_json method or similar. Because it would be extremely extraneous to duplicate all of the information in the model with the controller, only user-specified information (data + style) are explicitly kept. If a user wants more information (defaults) from the view/model, it should be able to query for it.

  • This might be annoying to do, non-specified kwargs are pulled from the rcParams object which is in turn created from reading a user specified file and can be dynamically changed at run time. I suppose we could keep a dict of default defaults and compare against that. Not clear how this will interact with the style sheet [[MEP26]] - @tacaswell

Additional Notes:

  • The "raw data" does not necessarily need to be a list, ndarray, etc. Rather, it can more abstractly just have a method to yield data when needed.

  • Because the Controller will contain extra information that users may not want to keep around, it should not be created by default. You should be able to both (a) instantiate a Controller with a figure and (b) build a figure with a Controller.

Use Cases:

  • Export all necessary informat

  • Serializing a matplotlib figure, saving it, and being able to rerun later.

  • Any other source sending an appropriately formatted representation to matplotlib to open


Here are some examples of what the controllers should be able to do.

  1. Instantiate a matplotlib figure from a serialized representation (e.g., JSON):

    import json
    from matplotlib.controllers import Controller
    with open('my_figure') as f:
        o = json.load(f)
    c = Controller(o)
    fig = c.figure
  2. Manage artists from the controller (e.g., Line2D):

    # not really sure how this should look
    c.axes[0].lines[0].color = 'b'
    # ?
  3. Export serializable figure representation:

    o = c.to_json()
    # or... we should be able to throw a figure object in there too
    o = Controller.to_json(mpl_fig)


  1. Create base Controller objects that are able to manage Artist objects (e.g., Hist)


    • initialization should happen via unpacking **, so we need a copy of call signature parameter for the Artist we're ultimately trying to control. Unfortunate hard-coded repetition...

    • should the additional **kwargs accepted by each Artist be tracked at the Controller

    • how does a Controller know which artist belongs where? E.g., do we need to pass axes references?


  2. Write in protocols for the Controller to update the model.


    • how should containers be dealt with? E.g., what happens to old patches when we re-bin a histogram?

    • in the link from (1), the old line is completely destroyed and redrawn, what if something is referencing it?

  3. Create method by which a json object can be assembled from the Controllers

  4. Deal with serializing the unserializable aspects of a figure (e.g., non-affine transforms?)

  5. Be able to instantiate from a serialized representation

  6. Reimplement the existing pyplot and Axes method, e.g. pyplot.hist and Axes.hist in terms of the new controller class.

> @theengineer: in #2 above, what do you mean by get updates from each Artist?

^ Yup. The Controller shouldn't need to get updated. This just happens in #3. Delete comments when you see this.

Backward compatibility#

  • pickling will change

  • non-affine transformations will require a defined pickling method


PR #3150 suggested adding semantics by parasitically attaching extra containers to axes objects. This is a more complete solution with what should be a more developed/flexible/powerful framework.