#873 tulip-python crashes when headless: "QPixmap: Must construct a QApplication before a QPaintDevice"

[jim_jom@web.de]

I use tulip for rendering graphs to png and svg. The script reads a gexf, creates a view, applies algorithms, saves a png screenshot and exports a svg. It works fine when I use it from my desktop envrironment. I use the createNodeLinkDiagramView(g, show=False), so the script works without any displayed windows. It should therefore also work in a headless environment.

But when executed in a headless environment, it crashes.
The minimal complete example below shows the bug. First we create a LXD container, then we install tulip via pip3. When loading the c++ plugins, the interpreter crashes.

# lxc launch ubuntu:18.04
# lxc exec close-manatee bash
root@close-manatee:~# apt install python3-pip
root@close-manatee:~# pip3 install --upgrade pip
root@close-manatee:~# apt install libxrender1 libsm6 libgl1
root@close-manatee:~# python3 -m pip install tulipgui-python tulip-python
root@close-manatee:~# python3
Python 3.6.5 (default, Apr  1 2018, 05:46:30) 
[GCC 7.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import tulipgui
>>> import tulip
>>> tulip.tlp.loadPlugins()
QPixmap: Must construct a QApplication before a QPaintDevice
Aborted (core dumped)
root@close-manatee:~# 

Can you help me debug this problem? Am I missing dependencies?

[Patrick Mary]

Most of the Tulip gui plugins (views, interactors, perspectives) needs a QApplication to be instantiated, so they cannot be loaded without a previously existing QApplication. By the way, saving a png screenshot from a node-link diagram view also needs a QApplication object which in turn needs to connect to an X server when instantiated on Linux. So from my point of view it cannot works in an headless environment.

[jim_jom@web.de]

Thanks for your reply. I managed to get it somewhat working in LXD/Docker by using Xvfb and installing tons of dependencies (probably more than needed).
The script imports the GEXF, creates a hidden node-link window and then exports a PNG - which is working fine. The final PNG file contains the picture just like I need it.

But somehow the SVG export still segfaults. Do you know if the SVG Export module is different to the PNG export in any way? How can I narrow down why it is faulting?

The Dockerfile:

FROM python:3.6

RUN mkdir /graphs
VOLUME /graphs

RUN apt-get update && apt-get install -y libgl1 tulip libtulip-core-4.8 libtulip-dev libtulip-gui-4.8 libtulip-ogdf-4.8 libtulip-ogl-4.8 libtulip-python-4.8 libxrender1 libsm6 libgl1 xvfb libegl1-mesa libegl1-mesa-drivers libgl1-mesa-dri libgl1-mesa-glx libglapi-mesa libgles2-mesa libglu1-mesa mesa-utils

COPY mylib.whl /tmp/
RUN pip3 install /tmp/mylib.whl # also installs Python dependencies

CMD rm -f /tmp/.X1-lock && Xvfb :1 -screen 0 1024x768x24 +extension GLX +render -noreset & DISPLAY=:1 python3 -m "mylib.graph_to_image"

The script:

if "GEXF" not in tlp.getImportPluginsList():
    tlp.loadPlugins() # segfault here was mitigated by installing tons of packages via apt-get
assert("GEXF" in tlp.getImportPluginsList())

importedgraph = tlp.importGraph("GEXF", { "filename" : "/tmp/graph.gexf" })

# Layout FM^3
viewLayout = importedgraph.getLayoutProperty("viewLayout")
params = tlp.getDefaultPluginParameters("FM^3 (OGDF)", importedgraph)
importedgraph.applyLayoutAlgorithm("FM^3 (OGDF)", viewLayout, params)

# Show graph info in the view
viewport = tlpgui.createNodeLinkDiagramView(importedgraph, show=False)
pars = viewport.getRenderingParameters()
pars.setViewArrow(True)
pars.setViewEdgeLabel(True)
pars.setViewNodeLabel(True)
pars.setLabelsDensity(100)
pars.setLabelScaled(True)
viewport.setRenderingParameters(pars)

# Save PNG
viewport.saveSnapshot("/tmp/graph.png", png_width, png_height)
PIL.Image.open("/tmp/graph.png").save("/tmp/graph.png", format="PNG", optimize=True, compress_level=9)

# Save SVG
params = tlp.getDefaultPluginParameters("SVG Export", importedgraph)
params["Edge extremities"] = True
params["Makes SVG output human readable"] = False
params["Export edge labels"] = True
tlp.exportGraph('SVG Export', importedgraph, "/tmp/graph.svg", params ) # Segfault still happens

[Patrick Mary]

The saveSnapshort feature makes a rendering of the graph layout in a QPixmap used to create a QImage which then will be saved in a png file. While the "SVG export" works directly from the Tulip graph structure to describe the graph layout in the svg format; and so some layout properties values may not be properly handled.

[jim_jom@web.de]

The ExportSvg.cpp has been changed May 2018. The PyPi Package is from November 2017. Maybe the bug is already fixed.

Do you work on getting Tulip 5.2 into PyPi?