Best practices

There are a number of good and bad practices that may not be immediately obvious when developing a plugin. This page covers some known practices that could affect the ability to install or use your plugin effectively.

Don’t include PySide2 or PyQt5 in your plugin’s dependencies.

This is important!

Napari supports both PyQt and PySide backends for Qt. It is up to the end-user to choose which one they want. If they installed napari with pip install napari[all], then the [all] extra will (currently) install PyQt5 for them from pypi. If they installed via conda install napari, then they’ll have PyQt5, but via anaconda cloud instead of pypi. Lastly, they may have installed napari with PySide2.

Here’s what can go wrong if you also declare one of these backends in the install_requires section of your plugin metadata:

  • If they installed via conda install napari and then they install your plugin via pip (or via the builtin plugin installer, which currently uses pip), then there will be a binary incompatibility between their conda pyqt installation, and the new pip “PyQt5” installation. This will very likely lead to a broken environment, forcing the user to re-create their entire environment and re-install napari. This is an unfortunate consequence of package naming decisions, and it’s not something napari can fix.

  • Alternatively, they may end up with both PyQt and PySide in their environment, and while that’s not always guaranteed to break things, it can lead to unexpected and difficult to debug problems.

  • Don’t import from PyQt5 or PySide2 in your plugin: use qtpy.

    If you use from PyQt5 import QtCore (or similar) in your plugin, but the end-user has chosen to use PySide2 for their Qt backend — or vice versa — then your plugin will fail to import. Instead use from qtpy import   QtCore. qtpy is a Qt compatibility layer that will import from whatever backend is installed in the environment.

Try not to depend on packages that require C compilation if these packages do not offer wheels

Tip

This requires some awareness of how your dependencies are built and distributed…

Some python packages write a portion of their code in lower level languages like C or C++ and compile that code into “C Extensions” that can be called by python at runtime. This can greatly improve performance, but it means that the package must be compiled for each platform (i.e. Windows, Mac, Linux) that the package wants to support. Some packages do this compilation step ahead of time, by distributing “wheels” on PyPI… or by providing pre-compiled packages via conda. Other packages simply distribute the source code (as an “sdist”) and expect the end-user to compile it on their own computer. Compiling C code requires software that is not always installed on every computer. (If you’ve ever tried to pip install a package and had it fail with a big wall of red text saying something about gcc, then you’ve run into a package that doesn’t distribute wheels, and you didn’t have the software required to compile it).

As a plugin developer, if you depend on a package that uses C extensions but doesn’t distribute a pre-compiled wheel, then it’s very likely that your users will run into difficulties installing your plugin:

  • What is a “wheel”?

    Briefly, a wheel is a built distribution, containing code that is pre-compiled for a specific operating system.

    For more detail, see What Are Python Wheels and Why Should You Care?

  • How do I know if my dependency offers a wheel

    There are many ways, but a sure-fire way to know is to go to the respective package on PyPI, and click on the “Download Files” link. If the package offers wheels, you’ll see one or more files ending in .whl. For example, napari offers a wheel. If a package doesn’t offer a wheel, it may still be ok if it’s just a pure python package that doesn’t have any C extensions…

  • How do I know if one of my dependencies uses C Extensions?

    There’s no one right way, but more often than not, if a package uses C extensions, the setup() function in their setup.py file will use the ext_modules argument. (for example, see here in pytorch)

What about conda?

conda also distributes & installs pre-compiled packages, though they aren’t wheels. While this is definitely a fine way to install binary dependencies in a reliable way, the built-in napari plugin installer doesn’t currently work with conda. If your dependency is only available on conda, but does not offer wheels,you may guide your users in using conda to install your package or one of your dependencies. Just know that it may not work with the built-in plugin installer.

Don’t import heavy dependencies at the top of your module

Note

This point will be less relevant when we move to the second generation manifest-based plugin declaration, but it’s still a good idea to delay importing your plugin-specific dependencies and modules until after your hookspec has been called. This helps napari stay quick and responsive at startup.

Consider the following example plugin:

[options.entry_points]
napari.plugin =
  plugin-name = mypackage.napari_plugin

In this example, my_heavy_dependency_like_tensorflow will be imported immediately when napari is launched, and we search the entry_point mypackage.napari_plugin for decorated hook specifications.

# mypackage/napari_plugin.py
from napari_plugin_engine import napari_hook_specification
from qtpy.QtWidgets import QWidget
from my_heavy_dependency_like_tensorflow import something_amazing

class MyWidget(QWidget):
    def do_something_amazing(self):
        return something_amazing()

@napari_hook_specification
def napari_experimental_provide_dock_widget():
    return MyWidget

This can deterioate the end-user experience, and make napari feel slugish. Best practice is to delay heavy imports until right before they are used. The following slight modification will help napari load much faster:

# mypackage/napari_plugin.py
from napari_plugin_engine import napari_hook_specification
from qtpy.QtWidgets import QWidget

class MyWidget(QWidget):
    def do_something_amazing(self):
        # import has been moved here, will happen only after the user
        # has opened and used this widget.
        from my_heavy_dependency_like_tensorflow import something_amazing

        return something_amazing()

(again, the second gen napari plugin engine will help improve this situation, but it’s still a good idea!)

Don’t leave resources open

It’s always good practice to clean up resources like open file handles and databases. As a napari plugin it’s particularly important to do this (and especially for Windows users). If someone tries to use the built-in plugin manager to uninstall your plugin, open file handles and resources may cause the process to fail or even leave your plugin in an “installed-but-unuseable” state.

Don’t do this:

# my_plugin/module.py
import json

data_file = open("some_data_in_my_plugin.json")
data = json.load(data_file)

Instead, make sure to close your resource after grabbing the data (ideally by using a context manager, but manually otherwise):

with open("some_data_in_my_plugin.json") as data_file:
    data = json.load(data_file)

Write extensive tests for your plugin!

Programmer and author Bruce Eckel famously wrote:

“If it’s not tested, it’s broken”

It’s true. High test coverage is one way to show your users that you are dedicated to the stability of your plugin. Aim for 100%!

Of course, simply having 100% coverage doesn’t mean your code is bug-free, so make sure that you test all of the various ways that your code might be called.

See our Tips for testing napari plugins.

How to check test coverage?

The cookiecutter template is already set up to report test coverage, but you can test locally as well, using pytest-cov

  1. pip install pytest-cov

  2. Run your tests with pytest --cov=<your_package> --cov-report=html

  3. Open the resulting report in your browser: open htmlcov/index.html

  4. The report will show line-by-line what is being tested, and what is being missed. Continue writing tests until everything is covered! If you have lines that you know never need to be tested (like debugging code) you can exempt specific lines from coverage with the comment # pragma: no cover

  5. In the cookiecutter, coverage tests from github actions will be uploaded to codecov.io