matplotlib · timhoffm · Nov 21, 2022 · May 22, 2022
diff --git a/doc/users/explain/backends.rst b/doc/users/explain/backends.rst
@@ -237,3 +237,6 @@ More generally, any importable backend can be selected by using any of the
 methods above. If ``name.of.the.backend`` is the module containing the
 backend, use ``module://name.of.the.backend`` as the backend name, e.g.
 ``matplotlib.use('module://name.of.the.backend')``.
+
+Information for backend implementers is available at
+:doc:`/users/explain/writing_a_backend_pyplot_interface`.
diff --git a/doc/users/explain/index.rst b/doc/users/explain/index.rst
@@ -10,6 +10,7 @@ Explanations
 
     api_interfaces.rst
     backends.rst
+    writing_a_backend_pyplot_interface.rst
     interactive.rst
     fonts.rst
     event_handling.rst

diff --git a/doc/users/explain/writing_a_backend_pyplot_interface.rst b/doc/users/explain/writing_a_backend_pyplot_interface.rst
@@ -0,0 +1,82 @@
+=========================================
+Writing a backend -- the pyplot interface
+=========================================
+
+This page assumes general understanding of the information in the
+:doc:`/users/explain/backends` page, and is instead intended as reference for
+third-party backend implementers.  It also only deals with the interaction
+between backends and `.pyplot`, not with the rendering side, which is described
+in `.backend_template`.
+
+There are two APIs for defining backends: a new canvas-based API (introduced in
+Matplotlib 3.6), and an older function-based API.  The new API is simpler to
+implement because many methods can be inherited from "parent backends".  It is
+recommended if back-compatibility for Matplotlib < 3.6 is not a concern.
+However, the old API remains supported.
+
+Fundamentally, a backend module needs to provide information to `.pyplot`, so
+that
+
+1. `.pyplot.figure()` can create a new `.Figure` instance and associate it with
+   an instance of a backend-provided canvas class, itself hosted in an instance
+   of a backend-provided manager class.
+2. `.pyplot.show()` can show all figures and start the GUI event loop (if any).
+
+To do so, the backend module must define a ``backend_module.FigureCanvas``
+subclass of `.FigureCanvasBase`.  In the canvas-based API, this is the only
+strict requirement for backend modules.  The function-based API additionally
+requires many module-level functions to be defined.
+
+Canvas-based API (Matplotlib >= 3.6)
+------------------------------------
+
+1. **Creating a figure**: `.pyplot.figure()` calls
+   ``figure = Figure(); FigureCanvas.new_manager(figure, num)``
+   (``new_manager`` is a classmethod) to instantiate a canvas and a manager and
+   set up the ``figure.canvas`` and ``figure.canvas.manager`` attributes.
+   Figure unpickling uses the same approach, but replaces the newly
+   instantiated ``Figure()`` by the unpickled figure.
+
+   Interactive backends should customize the effect of ``new_manager`` by
+   setting the ``FigureCanvas.manager_class`` attribute to the desired manager
+   class, and additionally (if the canvas cannot be created before the manager,
+   as in the case of the wx backends) by overriding the
+   ``FigureManager.create_with_canvas`` classmethod.  (Non-interactive backends
+   can normally use a trivial ``FigureManagerBase`` and can therefore skip this
+   step.)
+
+   After a new figure is registered with `.pyplot` (either via
+   `.pyplot.figure()` or via unpickling), if in interactive mode, `.pyplot`
+   will call its canvas' ``draw_idle()`` method, which can be overridden as
+   desired.
+
+2. **Showing figures**: `.pyplot.show()` calls
+   ``FigureCanvas.manager_class.pyplot_show()`` (a classmethod), forwarding any
+   arguments, to start the main event loop.
+
+   By default, ``pyplot_show()`` checks whether there are any ``managers``
+   registered with `.pyplot` (exiting early if not), calls ``manager.show()``
+   on all such managers, and then, if called with ``block=True`` (or with
+   the default ``block=None`` and out of IPython's pylab mode and not in
+   interactive mode), calls ``FigureCanvas.manager_class.start_main_loop()``
+   (a classmethod) to start the main event loop.  Interactive backends should
+   therefore override the ``FigureCanvas.manager_class.start_main_loop``
+   classmethod accordingly (or alternatively, they may also directly override
+   ``FigureCanvas.manager_class.pyplot_show`` directly).
+
+Function-based API
+------------------
+
+1. **Creating a figure**: `.pyplot.figure()` calls
+   ``new_figure_manager(num, *args, **kwargs)`` (which also takes care of
+   creating the new figure as ``Figure(*args, **kwargs)``); unpickling calls
+   ``new_figure_manager_given_figure(num, figure)``.
+
+   Furthermore, in interactive mode, the first draw of the newly registered
+   figure can be customized by providing a module-level
+   ``draw_if_interactive()`` function.  (In the new canvas-based API, this
+   function is not taken into account anymore.)
+
+2. **Showing figures**: `.pyplot.show()` calls a module-level ``show()``
+   function, which is typically generated via the ``ShowBase`` class and its
+   ``mainloop`` method.