Write canonical example of how to use Matplotlib inside a webserver

[tacaswell]

As shown by #11094 #12612 matplotlib is used inside of webservers to dynamically serve plots. The example should show:

1. importance of picking a non-GUI backend (both by setting the backend and explicitly by not using pyplot). Given that these servers are multi-threaded, not using pyplot at all may be the best option to avoid any coupling between workers via pyplot and memory leaks due to pyplot holding references to figures forever. Should probably show use of new_figure_manager
2. saving to a bytesIO buffer to avoid disk
3. png, svg, and pdf
4. be applicable to both django and flask (not sure off the top of my head what the differences are), bonus points for cherrypy, aiohttp, tornado, and pyramid

[anntzer]

Rather than bothering with any explicit backend at all, I would strongly suggest merging #12450 and have the canonical approach be

from matplotlib.figure import Figure
fig = Figure()
<do the plotting>
fig.savefig(...)

Even with 3.0 (i.e. without that PR) new_figure_manager is totally overkill, just manually attaching a canvas (FigureCanvasAgg(fig) or even just FigureCanvasBase(fig)) is enough.
(... except if you want to use a third-party renderer e.g. mplcairo, but I doubt that's the point of the proposed example)

[mgeier]

Possible duplicate: #4124.

[tacaswell]

Indeed! I'm going to close the older one.

[anntzer]

Do you want to make a decision about #12450, by the way? :) (as noted above)

[anntzer]

@tacaswell I think something like

import base64
from io import BytesIO

from flask import Flask
from matplotlib.figure import Figure


app = Flask(__name__)


@app.route("/")
def hello():
    # Generate the figure **without using pyplot**.
    fig = Figure()
    ax = fig.subplots()
    ax.plot([1, 2])
    # Save it to a temporary buffer.
    buf = BytesIO()
    fig.savefig(buf, format="png")
    # Embed the result in the html output.
    data = base64.b64encode(buf.getbuffer()).decode("ascii")
    return f"<img src='data:image/png;base64,{data}'/>"

is what you're looking for?

[story645]

My version was (stolen from a cookbook)

fig = matplotlib.figure.Figure()
canvas = FigureCanvas(fig)
//plot to figure code
canvas.print_figure(response)

where the outfile is the responsebuffer, like

response = pyramid.response.Response(content_type='image/png')

[anntzer]

• canvas = FigureCanvas(fig) is not necessary anymore since Attach a FigureCanvasBase by default to Figures. #12450 (that's the whole point of that PR).
• I prefer suggesting the use of savefig (and yes you can also pass the Response buffer to it), as it's a regular method that you'd also use in other contents, rather than print_figure, which is a bit an implementation detail.

[tacaswell]

Yes, something as simple as that is what I had in mind.

Agree on preferring savefig over print_figure.

[pjmnoble]

Hi problem goes away really nicely using the AGG backend
I just import matlplotlib at the top of my code and follow this with matplotlib.use('AGG') before importing any other matplotlib components (matplotlib.pyplot in this case)

No need for the matplotlib GUI backend in the web server.

Running on a Mac(High Sierra) with a Django server