The idea, to load the plugins later, originated from pyexcel project  which uses loosely coupled plugins to extend the main package to read more file formats. During its code growth, the code in pyexcel packages to manage the external and internal plugins becomes a independent library, lml.
Lml is similar to Factories in Zope Component Architecture . It provides functionalities to discover, register and load lml based plugins. It cares how the meta data were written but it does NOT care how the plugin interface is written.
Simply, lml promises to load your external dependency when they are used, but only when you follow lazy-loading design principle below. Otherwise, lml does immediate import and takes away the developer’s responsibility to manage plugin registry and discovery.
In terms of extensibility of your proud package, lml keeps the door open even if you use lml for immediate import. As a developer, you give the choice to other contributor to write up a plugin for your package. As long as the user would have installed community created extensions, lml will discover them and use them.
Prior to lml, three different ways of loading external plugins have been tried in pyexcel. namespace package  comes from Python 3 or pkgutil style in Python 2 and 3. It allows the developer to split a bigger packages into a smaller ones and publish them separately. sphinxcontrib  uses a typical namespace package based method. However, namespace package places a strict requirement on the module’s __init__.py: nothing other than name space declaration should be present. It means no module level functions can be place there. This restriction forces the plugin to be driven by the main package but the plugin cannot use the main package as its own library to do specific things. So namespace package was ruled out.
The Flask extension management system was used early versions of pyexcel(=<0.21). This system manipulates sys.path so that your plugin package appears in the namespace of your main package. For example, there is a xls plugin called pyexcel-xls. To import it, you can use “import pyexcel.ext.xls”. The shortcomings are:
- explicit statement “import pyexcel.ext.xls” becomes a useless statement in your code. static code analyser(flake8/pep8/pycharm) would flag it up.
- you have to explicitly import it. Otherwise, your plugin is not imported. PR 7 of pyexcel-io has extended discussion on this topic.
- flask extension management system become deprecated by itself in Flask’s recent development since 2016.
In order to overcome those shortcomings, implicit imports were coded into module’s __init__.py. By iterating through currently installed modules in your python environment, the relevant plugins are imported automatically.
lml uses implicit import. In order to manage the plugins, pip can be used to install cherry-picked plugins or to remove unwanted plugins. In the situation where two plugins perform the same thing but have to co-exist in your current python path, you can nominate one plugin to be picked.
In terms of plugin registrations, three different approaches have been tried. Monkey-patching was easy to implement. When a plugin is imported, it loads the plugin dictionary from the main package and add itself. But it is generally perceived as a “bad” idea. Another way of doing it is to place the plugin code in the main component and the plugin just need to declare a dictionary as the plugin’s meta data. The main package register the meta data when it is imported. tablib  uses such a approach. The third way is to use meta-classes. M. Alchin (2008)  explained how meta class can be used to register plugin classes in a simpler way.
lml uses meta data for plugin registration. Since lml load your plugin later, the meta data is stored in the module’s __init__.py. For example, to load plugins later in tablib, the ‘exports’ variable should be taken out from the actual class file and replace the hard reference to the classes with class path string.
yapsy  and GEdit plugin management system  load plugins from file system. To install a plugin in those systems, is to copy and paste the plugin code to a designated directory. zope components, namespace packages and flask extensions can be installed via pypi. lml support the latter approach. lml plugins can be released to pypi and be installed by your end developers.
To use lml, it asks you to avoid importing your “heavy” dependencies in __init__.py. lml respects the independence of individual packages. You can put modular level functions in your __init__.py as long as it does not trigger immediate import of your dependency. This is to allow the individual plugin to become useful as it is, rather to be integrated with your main package. For example, pyexcel-xls can be an independent package to read and write xls data, without pyexcel.
With lml, as long as your third party developer respect the plugin name prefix, they could publish their plugins as they do to any normal pypi packages. And the end developer of yours would only need to do pip install.