Repository navigation
Make matplotlib fail more gracefully in headless environments #6739
Description
This is reproducible on any headless environment using matplotlib 1.5.1. On my system, the following script:
from matplotlib import pyplot as plt
plt.plot([1, 2], [1, 2])Will produce the following traceback:
---------------------------------------------------------------------------
TclError Traceback (most recent call last)
<ipython-input-2-8dd816f9ddc7> in <module>()
----> 1 plt.plot([1, 2], [1, 2])
/home/goldbaum/.virtualenvs/yt-dev/local/lib/python2.7/site-packages/matplotlib/pyplot.pyc in plot(*args, **kwargs)
3145 @_autogen_docstring(Axes.plot)
3146 def plot(*args, **kwargs):
-> 3147 ax = gca()
3148 # allow callers to override the hold state by passing hold=True|False
3149 washold = ax.ishold()
/home/goldbaum/.virtualenvs/yt-dev/local/lib/python2.7/site-packages/matplotlib/pyplot.pyc in gca(**kwargs)
926 matplotlib.figure.Figure.gca : The figure's gca method.
927 """
--> 928 return gcf().gca(**kwargs)
929
930 # More ways of creating axes:
/home/goldbaum/.virtualenvs/yt-dev/local/lib/python2.7/site-packages/matplotlib/pyplot.pyc in gcf()
576 return figManager.canvas.figure
577 else:
--> 578 return figure()
579
580
/home/goldbaum/.virtualenvs/yt-dev/local/lib/python2.7/site-packages/matplotlib/pyplot.pyc in figure(num, figsize, dpi, facecolor, edgecolor, frameon, FigureClass, **kwargs)
525 frameon=frameon,
526 FigureClass=FigureClass,
--> 527 **kwargs)
528
529 if figLabel:
/home/goldbaum/.virtualenvs/yt-dev/local/lib/python2.7/site-packages/matplotlib/backends/backend_tkagg.pyc in new_figure_manager(num, *args, **kwargs)
82 FigureClass = kwargs.pop('FigureClass', Figure)
83 figure = FigureClass(*args, **kwargs)
---> 84 return new_figure_manager_given_figure(num, figure)
85
86
/home/goldbaum/.virtualenvs/yt-dev/local/lib/python2.7/site-packages/matplotlib/backends/backend_tkagg.pyc in new_figure_manager_given_figure(num, figure)
90 """
91 _focus = windowing.FocusManager()
---> 92 window = Tk.Tk()
93 window.withdraw()
94
/usr/lib/python2.7/lib-tk/Tkinter.pyc in __init__(self, screenName, baseName, className, useTk, sync, use)
1765 baseName = baseName + ext
1766 interactive = 0
-> 1767 self.tk = _tkinter.create(screenName, baseName, className, interactive, wantobjects, useTk, sync, use)
1768 if useTk:
1769 self._loadtk()
TclError: no display name and no $DISPLAY environment variable
I think it would be much nicer if backends somehow detected these sorts of errors, fell back to the Agg backend (probably with a warning) and then goes ahead and makes the plot.
Activity
tacaswell commented on Jul 14, 2016
This is to some degree a configuration/packaging issue. On servers that are known to be headless, the system wide configuration file should be set to Agg as the default backend.
A way to do this is to require something on the backend be queryable if there are sufficient run-time requirements (which also makes sense for the GUI backends more generally as we do not know until we try if the python wrappers for the backend are even installed).
This should probably be done in pylab_setup in backend.__init__. Maybe add a function backend_usable( ) -> bool to the backends (and fall back to lambda : True if it does not exist).
WeatherGod commented on Jul 14, 2016
One place where this is not a configuration/packaging issue (at least, not
immediately obvious) is user crontabs. Program works fine from the command
line, but can then fail in the crontab. Not a hard problem to solve (set
the MPLBACKEND variable at the top of the crontab to "agg"), but that may
not be obvious to many people. But a solution of setting the matplotlibrc
file to "agg" wouldn't be advisable in this situation because the user may
still want the GUI for their account.
On Thu, Jul 14, 2016 at 3:47 PM, Thomas A Caswell notifications@github.com
wrote:
con-f-use commented on Mar 6, 2017
As and avid user, I just want to say a built-in fallback should be in place here. +1
So far I'm using this in my scripts:
if 'DISPLAY' not in os.environ: matplotlib.use('agg')Is there a better workaround?
anntzer commented on Jul 30, 2017
#8613 (comment) would also solve the issue: one would just need to put Agg at the end of the list of backends, rather than having even more specific fallback code.
anntzer commented on Sep 19, 2018
Closed by #9795 + following PRs.