Edit me on GitHub

Venusian

Venusian is a library which allows you to defer the action of decorators. Instead of taking actions when a function, method, or class decorator is executed at import time, you can defer the action until a separate “scan” phase.

This library is most useful for framework authors. It is compatible with CPython versions 2.6, 2.7, and 3.2+. It is also known to work on PyPy 1.5.

Note

The name “Venusian” is a riff on a library named Martian (which had its genesis in the Grok web framework), from which the idea for Venusian was stolen. Venusian is similar to Martian, but it offers less functionality, making it slightly simpler to use.

Overview

Offering a decorator that wraps a function, method, or class can be a convenience to your framework’s users. But the very purpose of a decorator makes it likely to impede testability of the function or class it decorates: use of a decorator often prevents the function it decorates from being called with the originally passed arguments, or a decorator may modify the return value of the decorated function. Such modifications to behavior are “hidden” in the decorator code itself.

For example, let’s suppose your framework defines a decorator function named jsonify which can wrap a function that returns an arbitrary Python data structure and renders it to a JSON serialization:

1
2
3
4
5
6
7
8
 import json

 def jsonify(wrapped):
     def json_wrapper(request):
         result = wrapped(request)
         dumped = json.dumps(result)
         return dumped
     return json_wrapper

Let’s also suppose a user has written an application using your framework, and he has imported your jsonify decorator function, and uses it to decorate an application function:

1
2
3
4
5
 from theframework import jsonify

 @jsonify
 def logged_in(request):
    return {'result':'Logged in'}

As a result of an import of the module containing the logged_in function, a few things happen:

  • The user’s logged_in function is replaced by the json_wrapper function.
  • The only reference left to the original logged_in function is inside the frame stack of the call to the jsonify decorator.

This means, from the perspective of the application developer that the original logged_in function has effectively “disappeared” when it is decorated with your jsonify decorator. Without bothersome hackery, it can no longer be imported or retrieved by its original author.

More importantly, it also means that if the developer wants to unit test the logged_in function, he’ll need to do so only indirectly: he’ll need to call the json_wrapper wrapper decorator function and test that the json returned by the function contains the expected values. This will often imply using the json.loads function to turn the result of the function back into a Python dictionary from the JSON representation serialized by the decorator.

If the developer is a stickler for unit testing, however, he’ll want to test only the function he has actually defined, not the wrapper code implied by the decorator your framework has provided. This is the very definition of unit testing (testing a “unit” without any other integration with other code). In this case, it is also more convenient for him to be able to test the function without the decorator: he won’t need to use the json.loads function to turn the result back into a dictionary to make test assertions against. It’s likely such a developer will try to find ways to get at the original function for testing purposes.

To do so, he might refactor his code to look like this:

1
2
3
4
5
6
7
8
 from theframework import jsonify

 @jsonify
 def logged_in(request):
    return _logged_in(request)

 def _logged_in(request):
    return {'result':'Logged in'}

Then in test code he might import only the _logged_in function instead of the decorated logged_in function for purposes of unit testing. In such a scenario, the concentious unit testing app developer has to define two functions for each decorated function. If you’re thinking “that looks pretty tedious”, you’re right.

To give the intrepid tester an “out”, you might be tempted as a framework author to leave a reference to the original function around somewhere that the unit tester can import and use only for testing purposes. You might modify the jsonify decorator like so in order to do that:

1
2
3
4
5
6
7
8
 import json
 def jsonify(wrapped):
     def json_wrapper(request):
         result = wrapped(request)
         dumped = json.dumps(result)
         return dumped
     json_wrapper.original_function = wrapped
     return json_wrapper

The line json_wrapper.original_function = wrapped is the interesting one above. It means that the application developer has a chance to grab a reference to his original function:

1
2
3
 from myapp import logged_in
 result = logged_in.original_func(None)
 self.assertEqual(result['result'], 'Logged in')

That works. But it’s just a little weird. Since the jsonify decorator function has been imported by the developer from a module in your framework, the developer probably shouldn’t really need to know how it works. If he needs to read its code, or understand documentation about how the decorator functions for testing purposes, your framework might be less valuable to him on some level. This is arguable, really. If you use some consistent pattern like this for all your decorators, it might be a perfectly reasonable solution.

However, what if the decorators offered by your framework were passive until activated explicitly? This is the promise of using Venusian within your decorator implementations. You may use Venusian within your decorators to associate a wrapped function, class, or method with a callback. Then you can return the originally wrapped function. Instead of your decorators being “active”, the callback associated with the decorator is passive until a “scan” is initiated.

Using Venusian

The most basic use of Venusian within a decorator implementation is demonstrated below.

1
2
3
4
5
6
7
import venusian

def jsonify(wrapped):
    def callback(scanner, name, ob):
        print 'jsonified'
    venusian.attach(wrapped, callback)
    return wrapped

As you can see, this decorator actually calls into venusian, but then simply returns the wrapped object. Effectively this means that this decorator is “passive” when the module is imported.

Usage of the decorator:

1
2
3
4
5
from theframework import jsonify

@jsonify
def logged_in(request):
    return {'result':'Logged in'}

Note that when we import and use the function, the fact that it is decorated with the jsonify decorator is immaterial. Our decorator doesn’t actually change its behavior.

1
2
3
4
>>> from theapp import logged_in
>>> logged_in()
{'result':'Logged in'}
>>>

This is the intended result. During unit testing, the original function can be imported and tested despite the fact that it has been wrapped with a decorator.

However, we can cause something to happen when we invoke a scan.

1
2
3
4
5
import venusian
import theapp

scanner = venusian.Scanner()
scanner.scan(theapp)

Above we’ve imported a module named theapp. The logged_in function which we decorated with our jsonify decorator lives in this module. We’ve also imported the venusian module, and we’ve created an instance of the venusian.Scanner class. Once we’ve created the instance of venusian.Scanner, we invoke its venusian.Scanner.scan() method, passing the theapp module as an argument to the method.

Here’s what happens as a result of invoking the venusian.Scanner.scan() method:

  1. Every object defined at module scope within the theapp Python module will be inspected to see if it has had a Venusian callback attached to it.
  2. For every object that does have an Venusian callback attached to it, the callback is called.

We could have also passed the scan method a Python package instead of a module. This would recursively import each module in the package (as well as any modules in subpackages), looking for callbacks.

Note

During scan, the only Python files that are processed are Python source (.py) files. Compiled Python files (.pyc, .pyo files) without a corresponding source file are ignored.

In our case, because the callback we defined within the jsonify decorator function prints jsonified when it is invoked, which means that the word jsonified will be printed to the console when we cause venusian.Scanner.scan() to be invoked. How is this useful? It’s not! At least not yet. Let’s create a more realistic example.

Let’s change our jsonify decorator to perform a more useful action when a scan is invoked by changing the body of its callback.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
import venusian

def jsonify(wrapped):
    def callback(scanner, name, ob):
        def jsonified(request):
            result = wrapped(request)
            return json.dumps(result)
        scanner.registry.add(name, jsonified)
    venusian.attach(wrapped, callback)
    return wrapped

Now if we invoke a scan, we’ll get an error:

1
2
3
4
5
6
7
import venusian
import theapp

scanner = venusian.Scanner()
scanner.scan(theapp)

AttributeError: Scanner has no attribute 'registry'.

The venusian.Scanner class constructor accepts any key-value pairs; for each key/value pair passed to the scanner’s constructor, an attribute named after the key which points at the value is added to the scanner instance. So when you do:

1
2
import venusian
scanner = venusian.Scanner(a=1)

Thereafter, scanner.a will equal the integer 1.

Any number of key-value pairs can be passed to a scanner. The purpose of being able to pass arbitrary key/value pairs to a scanner is to allow cooperating decorator callbacks to access these values: each callback is passed the scanner constructed when a scan is invoked.

Let’s fix our example by creating an object named registry that we’ll pass to our scanner’s constructor:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
import venusian
import theapp

class Registry(object):
    def __init__(self):
       self.registered = []

    def add(self, name, ob):
       self.registered.append((name, ob))

registry = Registry()
scanner = venusian.Scanner(registry=registry)
scanner.scan(theapp)

At this point, we have a system which, during a scan, for each object that is wrapped with a Venusian-aware decorator, a tuple will be appended to the registered attribute of a Registry object. The first element of the tuple will be the decorated object’s name, the second element of the tuple will be a “truly” decorated object. In our case, this will be a jsonify-decorated callable.

Our framework can then use the information in the registry to decide which view function to call when a request comes in.

Venusian callbacks must accept three arguments:

scanner

This will be the instance of the scanner that has had its scan method invoked.

name

This is the module-level name of the object being decorated.

ob

This is the object being decorated if it’s a function or an instance; if the object being decorated is a method, however, this value will be the class.

If you consider that the decorator and the scanner can cooperate, and can perform arbitrary actions together, you can probably imagine a system where a registry will be populated that informs some higher-level system (such as a web framework) about the available decorated functions.

Scan Categories

Because an application may use two separate Venusian-using frameworks, Venusian allows for the concept of “scan categories”.

The venusian.attach() function accepts an additional argument named category.

For example:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
import venusian

def jsonify(wrapped):
    def callback(scanner, name, ob):
        def jsonified(request):
            result = wrapped(request)
            return json.dumps(result)
        scanner.registry.add(name, jsonified)
    venusian.attach(wrapped, callback, category='myframework')
    return wrapped

Note the category='myframework' argument in the call to venusian.attach(). This tells Venusian to attach the callback to the wrapped object under the specific scan category myframework. The default scan category is None.

Later, during venusian.Scanner.scan(), a user can choose to activate all the decorators associated only with a particular set of scan categories by passing a categories argument. For example:

1
2
3
import venusian
scanner = venusian.Scanner(a=1)
scanner.scan(theapp, categories=('myframework',))

The default categories argument is None, which means activate all Venusian callbacks during a scan regardless of their category.

onerror Scan Callback

New in version 1.0.

By default, when Venusian scans a package, it will propagate all exceptions raised while attempting to import code. You can use an onerror callback argument to venusian.Scanner.scan() to change this behavior.

The onerror argument should either be None or a callback function which behaves the same way as the onerror callback function described in http://docs.python.org/library/pkgutil.html#pkgutil.walk_packages .

Here’s an example onerror callback that ignores all ImportError exceptions:

1
2
3
4
  import sys
  def onerror(name):
      if not issubclass(sys.exc_info()[0], ImportError):
          raise # reraise the last exception

Here’s how we’d use this callback:

1
2
3
import venusian
scanner = venusian.Scanner()
scanner.scan(theapp, onerror=onerror)

The onerror callback should execute raise at some point if any exception is to be propagated, otherwise it can simply return. The name passed to onerror is the module or package dotted name that could not be imported due to an exception.

ignore Scan Argument

New in version 1.0a3.

The ignore to scan allows you to ignore certain modules, packages, or global objects during a scan. It should be a sequence containing strings and/or callables that will be used to match against the full dotted name of each object encountered during the scanning process. If the ignore value you provide matches a package name, global objects contained by that package as well any submodules and subpackages of the package (and any global objects contained by them) will be ignored. If the ignore value you provide matches a module name, any global objects in that module will be ignored. If the ignore value you provide matches a global object that lives in a package or module, only that particular global object will be ignored.

The sequence can contain any of these three types of objects:

  • A string representing a full dotted name. To name an object by dotted name, use a string representing the full dotted name. For example, if you want to ignore the my.package package and any of its subobjects during the scan, pass ignore=['my.package']. If the string matches a global object (e.g. ignore=['my.package.MyClass']), only that object will be ignored and the rest of the objects in the module or package that contains the object will be processed.
  • A string representing a relative dotted name. To name an object relative to the package passed to this method, use a string beginning with a dot. For example, if the package you’ve passed is imported as my.package, and you pass ignore=['.mymodule'], the my.package.mymodule module and any of its subobjects will be omitted during scan processing. If the string matches a global object (e.g. ignore=['my.package.MyClass']), only that object will be ignored and the rest of the objects in the module or package that contains the object will be processed.
  • A callable that accepts a full dotted name string of an object as its single positional argument and returns True or False. If the callable returns True or anything else truthy, the module, package, or global object is ignored, if it returns False or anything else falsy, it is not ignored. If the callable matches a package name, the package as well as any of that package’s submodules and subpackages (recursively) will be ignored. If the callable matches a module name, that module and any of its contained global objects will be ignored. If the callable mactches a global object name, only that object name will be ignored. For example, if you want to skip all packages, modules, and global objects that have a full dotted name that ends with the word “tests”, you can use ignore=[re.compile('tests$').search].

Here’s an example of how we might use the ignore argument to scan to ignore an entire package (and any of its submodules and subpackages) by absolute dotted name:

1
2
3
import venusian
scanner = venusian.Scanner()
scanner.scan(theapp, ignore=['theapp.package'])

Here’s an example of how we might use the ignore argument to scan to ignore an entire package (and any of its submodules and subpackages) by relative dotted name (theapp.package):

1
2
3
import venusian
scanner = venusian.Scanner()
scanner.scan(theapp, ignore=['.package'])

Here’s an example of how we might use the ignore argument to scan to ignore a particular class object:

1
2
3
import venusian
scanner = venusian.Scanner()
scanner.scan(theapp, ignore=['theapp.package.MyClass'])

Here’s an example of how we might use the ignore argument to scan to ignore any module, package, or global object that has a name which ends with the string tests:

1
2
3
4
import re
import venusian
scanner = venusian.Scanner()
scanner.scan(theapp, ignore=[re.compile('tests$').search])

You can mix and match the three types in the list. For example, scanner.scan(my, ignore=['my.package', '.someothermodule', re.compile('tests$').search]) would cause my.package (and all its submodules and subobjects) to be ignored, my.someothermodule to be ignored, and any modules, packages, or global objects found during the scan that have a full dotted path that ends with the word tests to be ignored beneath the my package.

Packages and modules matched by any ignore in the list will not be imported, and their top-level code will not be run as a result.

Limitations and Audience

Venusian is not really a tool that is maximally useful to an application developer. It would be a little silly to use it every time you needed a decorator. Instead, it’s most useful for framework authors, in order to be able to say to their users “the frobozz decorator doesn’t change the output of your function at all” in documentation. This is a lot easier than telling them how to test methods/functions/classes decorated by each individual decorator offered by your frameworks.

API Documentation / Glossary

Indices and tables