A dynamic code loading framework for building pluggable Python distributions

localstack, updated 🕥 2023-02-21 15:37:54

Plux

CI badge PyPI Version PyPI License Code style: black

plux is the dynamic code loading framework used in LocalStack.

Overview

Plux builds a higher-level plugin mechanism around Python's entry point mechanism. It provides tools to load plugins from entry points at run time, and to discover entry points from plugins at build time (so you don't have to declare entry points statically in your setup.py).

Core concepts

  • PluginSpec: describes a Plugin. Each plugin has a namespace, a unique name in that namespace, and a PluginFactory (something that creates Plugin the spec is describing. In the simplest case, that can just be the Plugin's class).
  • Plugin: an object that exposes a should_load and load method. Note that it does not function as a domain object (it does not hold the plugins lifecycle state, like initialized, loaded, etc..., or other metadata of the Plugin)
  • PluginFinder: finds plugins, either at build time (by scanning the modules using pkgutil and setuptools) or at run time (reading entrypoints of the distribution using stevedore)
  • PluginManager: manages the run time lifecycle of a Plugin, which has three states:
  • resolved: the entrypoint pointing to the PluginSpec was imported and the PluginSpec instance was created
  • init: the PluginFactory of the PluginSpec was successfully invoked
  • loaded: the load method of the Plugin was successfully invoked

architecture

Loading Plugins

At run time, a PluginManager uses a PluginFinder that in turn uses stevedore to scan the available entrypoints for things that look like a PluginSpec. With PluginManager.load(name: str) or PluginManager.load_all(), plugins within the namespace that are discoverable in entrypoints can be loaded. If an error occurs at any state of the lifecycle, the PluginManager informs the PluginLifecycleListener about it, but continues operating.

Discovering entrypoints

To build a source distribution and a wheel of your code with your plugins as entrypoints, simply run python setup.py plugins sdist bdist_wheel.

How it works: For discovering plugins at build time, plux provides a custom setuptools command plugins, invoked via python setup.py plugins. The command uses a special PluginFinder that collects from the codebase anything that can be interpreted as a PluginSpec, and creates from it a plugin index file plux.json, that is placed into the .egg-info distribution metadata directory. When a setuptools command is used to create the distribution (e.g., python setup.py sdist/bdist_wheel/...), plux finds the plux.json plugin index and extends automatically the list of entry points (collected into .egg-info/entry_points.txt). The plux.json file becomes a part of the distribution, s.t., the plugins do not have to be discovered every time your distribution is installed elsewhere.

Examples

To build something using the plugin framework, you will first want to introduce a Plugin that does something when it is loaded. And then, at runtime, you need a component that uses the PluginManager to get those plugins.

One class per plugin

This is the way we went with LocalstackCliPlugin. Every plugin class (e.g., ProCliPlugin) is essentially a singleton. This is easy, as the classes are discoverable as plugins. Simply create a Plugin class with a name and namespace and it will be discovered by the build time PluginFinder.

```python

abstract case (not discovered at build time, missing name)

class CliPlugin(Plugin): namespace = "my.plugins.cli"

def load(self, cli):
    self.attach(cli)

def attach(self, cli):
    raise NotImplementedError

discovered at build time (has a namespace, name, and is a Plugin)

class MyCliPlugin(CliPlugin): name = "my"

def attach(self, cli):
    # ... attach commands to cli object

```

now we need a PluginManager (which has a generic type) to load the plugins for us:

```python cli = # ... needs to come from somewhere

manager: PluginManager[CliPlugin] = PluginManager("my.plugins.cli", load_args=(cli,))

plugins: List[CliPlugin] = manager.load_all()

todo: do stuff with the plugins, if you want/need

in this example, we simply use the plugin mechanism to run a one-shot function (attach) on a load argument

```

Re-usable plugins

When you have lots of plugins that are structured in a similar way, we may not want to create a separate Plugin class for each plugin. Instead we want to use the same Plugin class to do the same thing, but use several instances of it. The PluginFactory, and the fact that PluginSpec instances defined at module level are discoverable (inpired by pluggy), can be used to achieve that.

```python

class ServicePlugin(Plugin):

def __init__(self, service_name):
    self.service_name = service_name
    self.service = None

def should_load(self):
    return self.service_name in config.SERVICES

def load(self):
    module = importlib.import_module("localstack.services.%s" % self.service_name)
    # suppose we define a convention that each service module has a Service class, like moto's `Backend`
    self.service = module.Service()

def service_plugin_factory(name) -> PluginFactory: def create(): return ServicePlugin(name)

return create

discoverable

s3 = PluginSpec("localstack.plugins.services", "s3", service_plugin_factory("s3"))

discoverable

dynamodb = PluginSpec("localstack.plugins.services", "dynamodb", service_plugin_factory("dynamodb"))

... could be simplified with convenience framework code, but the principle will stay the same

```

Then we could use the PluginManager to build a Supervisor

```python

class Supervisor: manager: PluginManager[ServicePlugin]

def start(self, service_name):
    plugin = manager.load(service_name)
    service = plugin.service
    service.start()

```

Functions as plugins

with the @plugin decorator, you can expose functions as plugins. They will be wrapped by the framework into FunctionPlugin instances, which satisfy both the contract of a Plugin, and that of the function.

```python from plugin import plugin

@plugin(namespace="localstack.configurators") def configure_logging(runtime): logging.basicConfig(level=runtime.config.loglevel)

@plugin(namespace="localstack.configurators") def configure_somethingelse(runtime): # do other stuff with the runtime object pass ```

With a PluginManager via load_all, you receive the FunctionPlugin instances, that you can call like the functions

```python

runtime = LocalstackRuntime()

for configurator in PluginManager("localstack.configurators").load_all(): configurator(runtime) ```

Configuring your distribution

If you are building a python distribution that exposes plugins discovered by plux, you need to configure your projects build system so other dependencies creates the entry_points.txt file when installing your distribution.

For a pyproject.toml template this involves adding the build-system section:

```toml [build-system] requires = ['setuptools', 'wheel', 'plux>=1.3.1'] build-backend = "setuptools.build_meta"

...

```

Install

pip install plux

Develop

Create the virtual environment, install dependencies, and run tests

make venv
make test

Run the code formatter

make format

Upload the pypi package using twine

make upload

Issues

[Request] Use tags when creating releases

opened on 2022-08-23 21:48:25 by jonringer

Trying to package localstack for Nixpkgs. Wanted to pull in tests from github to ensure that plux works, however, it's hard to know which commit was used for a particular release.

Provide command to generate setup.py with entrypoints

opened on 2022-06-30 17:55:19 by thrau

Sometimes it can be interesting to not need plux at install time, which can be facilitated by adding the entrypoints directly in the setup.py. It would be great if plux could provide a method similar to setup.py plugins, which just creates or modifies an existing setup.py file by adding the located plugin entrypoints.

Add report log output at end of plugins command

opened on 2022-06-28 16:48:15 by thrau

Currently the log output of the plugins command is a bit of a mess. There's no real good way of seeing the warnings that some files weren't importable and scannable for plugins. A nicely formatted report at the end of the command would be useful, containing

  • all discovered plugins and their files
  • all files that had an error while importing (and the related error)

It would also be useful to configure the loglevel and capture all the output created by simply importing code, similar to pytests' --log-cli-level

LocalStack

Enabling efficient local dev&test loops for Cloud applications

GitHub Repository