Enable plotting in same notebook cell which imports matplotlib.pyplot

[benjimin]

This fixes a bug for Jupyter notebooks, that plot commands often do not display until the second time their cell is executed. jupyter/notebook#3691

Specifically, inline display of plots was broken for the first cell which imports matplotlib.pyplot, but does works in subsequent cells (or even if the same cell is evaluated a second time). This particularly affected libraries that lazily import matplotlib, such as xarray:

In [1]: import xarray, numpy as np
In [2]: x = xarray.DataArray(np.random.random((5,5)))
In [3]: x.plot() # did not work
In [4]: x.plot() # did work

The problem related to circular imports between matplotlib and ipython. The import of the matplotlib.pyplot module executes a call to IPython.core.pylabtools:activate_matplotlib which in turn imports matplotlib.pyplot (while it is still only partially initialised e.g. the matplotlib module may not yet have a pyplot attribute). By using a slightly different import syntax it is possible to avoid causing an exception here.

Note, there is also a partial work-around in ipykernel.pylab.backend_inline:_enable_matplotlib_integration which catches the import/attribute error and in that case reschedules activate_matplotlib until after the cell finishes executing (i.e. after import completes). The problem was that it was late to schedule flush_figures (as the event is already triggering before the callback gets registered), which is why plotting only worked in subsequent cell evaluations.

Tested on python 3.6.7.

[benjimin]

I'd also like to add a regression test, but unsure how to programmatically create a separate kernel instance that thinks it supports inline plotting (so that I can test whether the kernel calls flush_figures after the same cell that runs import matplotlib.pyplot)?

[minrk]

Thanks for the fix! A test like you are describing would probably be best in the ipykernel repo, which has the inline backend and eventloop setup