Skip to content

disqus/gutter

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

https://api.travis-ci.org/disqus/gutter.png?branch=master

Gutter

NOTE: This repo is the client for Gargoyle 2, known as "Gutter". It does not work with the existing Gargoyle 1 codebase.

Gutter is feature switch management library. It allows users to create feature switches and setup conditions those switches will be enabled for. Once configured, switches can then be checked against inputs (requests, user objects, etc) to see if the switches are active.

For a UI to configure Gutter with see the gutter-django project

Table of Contents

Configuration

Gutter requires a small bit of configuration before usage.

Choosing Storage

Switches are persisted in a storage object, which is a dict or any object which provides the types.MappingType interface (__setitem__ and __getitem__ methods). By default, gutter uses an instance of MemoryDict from the durabledict library. This engine does not persist data once the process ends so a more persistent data store should be used.

Autocreate

gutter can also "autocreate" switches. If autocreate is enabled, and gutter is asked if the switch is active but the switch has not been created yet, gutter will create the switch automatically. When autocreated, a switch's state is set to "disabled."

This behavior is off by default, but can be enabled through a setting. More on "settings" below.

Configuring Settings

To change the storage and/or autocreate settings, simply import the settings module and set the appropriate variables:

from gutter.client.settings import manager as manager_settings
from durabledict.dict import RedisDict
from redis import RedisClient

manager_settings.storage_engine = RedisDict('gutter', RedisClient()))
manager_settings.autocreate = True

In this case, we are changing the engine to durabledict's RedisDict and turning on autocreate. These settings will then apply to all newly constructed Manager instances. More on what a Manager is and how you use it later in this document.

Setup

Once the Manager's storage engine has been configured, you can import gutter's default Manager object, which is your main interface with gutter:

from gutter.client.default import gutter

At this point the gutter object is an instance of the Manager class, which holds all methods to register switches and check if they are active. In most installations and usage scenarios, the gutter.client.gutter manager will be your main interface.

Using a different default Manager

If you would like to construct and use a different default manager, but still have it accessible via gutter.client.gutter, you can construct and then assign a Manager instance to settings.manager.default value:

from gutter.client.settings import manager as manager_settings
from gutter.client.models import Manager

manager_settings.default = Manager({})   # Must be done before importing the default manager

from gutter.client.default import gutter

assert manager_settings.default is gutter

Warning

warning::warning:

Note that the settings.manager.default value must be set before importing the default gutter instance. :warning::warning:

Arguments

The first step in your usage of gutter should be to define your arguments that you will be checking switches against. An "argument" is an object which understands the business logic and object in your system (users, requests, etc) and knows how to validate, transform and extract variables from those business objects for Switch conditions. For instance, your system may have a User object that has properties like is_admin, date_joined, etc. To switch against it, you would then create arguments for each of those values.

To do that, you construct a class which inherits from gutter.client.arguments.Container. Inside the body of the class, you create as many class variable "arguments" that you need by using the gutter.client.arguments function.

from gutter.client import arguments

from myapp import User

class UserArguments(arguments.Container):

    COMPATIBLE_TYPE = User

    name = arguments.String(lambda self: self.input.name)
    is_admin = arguments.Boolean(lambda self: self.input.is_admin)
    age = arguments.Value(lambda self: self.input.age)

There are a few things going on here, so let's break down what they all mean.

  1. The UserArgument class is subclassed from Container. The subclassing is required since Container implements some of the required API.
  2. The class has a bunch of class variables that are calls to arguments.TYPE, where TYPE is the type of variable this argument is. At present there are 3 types: Value for general values, Boolean for boolean values and String for string values.
  3. arguments.TYPE() is called with a callable that returns the value. In the above example, we'll want to make some switches active based on a user's name, is_admin status and age.
  4. Those callables return the actual value, which is derefenced from self.input, which is the input object (in this case a User instance).
  5. Variable objects understand Switch conditions and operators, and implement the correct API to allow themselves to be appropriately compared.
  6. COMPATIBLE_TYPE declares that this argument only works with User instances. This works with the default implementation of applies in the base argument that checks if the type of the input is the same as COMPATIBLE_TYPE.

Since constructing arguments that simply reference an attribute on self.input is so common, if you pass a string as the first argument of argument(), when the argument is accessed, it will simply return that property from self.input. You must also pass a Variable to the variable= kwarg so gutter know what Variable to wrap your value in.

from gutter.client import arguments

from myapp import User

class UserArguments(Container):

    COMPATIBLE_TYPE = User

    name = arguments.String('name')
    is_admin = arguments.Boolean('is_admin')
    age = arguments.Value('age')

Rationale for Arguments

You might be asking, why have these Argument objects at all? They seem to just wrap an object in my system and provide the same API. Why can't I just use my business object itself and compare it against my switch conditions?

The short answer is that Argument objects provide a translation layer to translate your business objects into objects that gutter understands. This is important for a couple reasons.

First, it means you don't clutter your business logic/objects with code to support gutter. You declare all the arguments you wish to provide to switches in one location (an Argument) whose single responsibility it to interface with gutter. You can also construct more savvy Argument objects that may be the combination of multiple business objects, consult 3rd party services, etc. All still not cluttering your main application code or business objects.

Secondly, and most importantly, Arguments return Variable objects, which ensure gutter conditions work correctly. This is mostly relevant to the percentage-based operators, and is best illustrated with an example.

Imagine you have a User class with an is_vip boolean field. Let's say you wanted to turn on a feature for only 10% of your VIP customers. To do that, you would write a condition that says, "10% of the time when I'm called with the variable, I should be true." That line of code would probably do something like this:

return 0 <= (hash(variable) % 100) < 10

The issue is that if variable = True, then hash(variable) % 100 will always be the same value for every User with is_vip of True:

>>> hash(True)
1
>>> hash(True) % 100
1

This is because in Python True objects always have the same hash value, and thus the percentage check doesn't work. This is not the behavior you want.

For the 10% percentage range, you want it to be active for 10% of the inputs. Therefore, each input must have a unique hash value, exactly the feature the Boolean variable provides. Every Variable has known characteristics against conditions, while your objects may not.

That said, you don't absolutely have to use Variable objects. For obvious cases, like use.age > some_value your User instance will work just fine, but to play it safe you should use Variable objects. Using Variable objects also ensure that if you update gutter any new Operator types that are added will work correctly with your ``Variable``s.

Switches

Switches encapsulate the concept of an item that is either 'on' or 'off' depending on the input. The swich determines its on/off status by checking each of its conditions and seeing if it applies to a certain input.

Switches are constructed with only one required argument, a name:

from gutter.client.models import Switch

switch = Switch('my cool feature')

Switches can be in 3 core states: GLOBAL, DISABLED and SELECTIVE. In the GLOBAL state, the Switch is enabled for every input no matter what. DISABLED Switches are not disabled for any input, no matter what. SELECTIVE Switches enabled based on their conditions.

Switches can be constructed in a certain state or the property can be changed later:

switch = Switch('new feature', state=Switch.states.DISABLED)
another_switch = Switch('new feature')
another_switch.state = Switch.states.DISABLED

Compounded

When in the SELECTIVE state, normally only one condition needs be true for the Switch to be enabled for a particular input. If switch.compounded is set to True, then all of the switches conditions need to be true in order to be enabled:

switch = Switch('require alll conditions', compounded=True)

Heriarchical Switches

You can create switches using a specific hierarchical naming scheme. Switch namespaces are divided by the colon character (":"), and hierarchies of switches can be constructed in this fashion:

parent = Switch('movies')
child1 = Switch('movies:star_wars')
child2 = Switch('movies:die_hard')
grandchild = Switch('movies:star_wars:a_new_hope')

In the above example, the child1 switch is a child of the "movies" switch because it has movies: as a prefix to the switch name. Both child1 and child2 are "children of the parent parent switch. And grandchild is a child of the child1 switch, but not the child2 switch.

Concent

By default, each switch makes its "am I active?" decision independent of other switches in the Manager (including its parent), and only consults its own conditions to check if it is enabled for the input. However, this is not always the case. Perhaps you have a cool new feature that is only available to a certain class of user. And of those users, you want 10% to be be exposed to a different user interface to see how they behave vs the other 90%.

gutter allows you to set a concent flag on a switch that instructs it to check its parental switch first, before checking itself. If it checks its parent and it is not enabled for the same input, the switch immediately returns False. If its parent is enabled for the input, then the switch will continue and check its own conditions, returning as it would normally.

For example:

parent = Switch('cool_new_feature')
child = Switch('cool_new_feature:new_ui', concent=True)

For example, because child was constructed with concent=True, even if child is enabled for an input, it will only return True if parent is also enabled for that same input.

Note: Even switches in a GLOBAL or DISABLED state (see "Switch" section above) still consent their parent before checking themselves. That means that even if a particular switch is GLOBAL, if it has concent set to True and its parent is not enabled for the input, the switch itself will return False.

Registering a Switch

Once your Switch is constructed with the right conditions, you need to register it with a Manager instance to preserve it for future use. Otherwise it will only exist in memory for the current process. Register a switch via the register method on a Manager instance:

gutter.register(switch)

The Switch is now stored in the Manager's storage and can be checked if active through gutter.active(switch).

Updating a Switch

If you need to update your Switch, simply make the changes to the Switch object, then call the Manager's update() method with the switch to tell it to update the switch with the new object:

switch = Switch('cool switch')
manager.register(switch)

switch.name = 'even cooler switch'  # Switch has not been updated in manager yet

manager.update(switch)  # Switch is now updated in the manager

Since this is a common pattern (retrieve switch from the manager, then update it), gutter provides a shorthand API in which you ask the manager for a switch by name, and then call save() on the switch to update it in the Manager it was retreived from:

switch = manager.switch('existing switch')
switch.name = 'a new name'  # Switch is not updated in manager yet
switch.save()  # Same as calling manager.update(switch)

Unregistering a Switch

Existing switches may be removed from the Manager by calling unregister() with the switch name or switch instance:

gutter.unregister('deprecated switch')
gutter.unregister(a_switch_instance)

Note: If the switch is part of a hierarchy and has children switches (see the "Hierarchical Switches" section above), all descendent switches (children, grandchildren, etc) will also be unregistered and deleted.

Conditions

Each Switch can have 0+ conditions, which describe the conditions under which that switch is active. Condition objects are constructed with three values: a argument, attribute and operator.

An argument is any Argument class, like the one you defined earlier. From the previous example, UserArgument is an argument object. attribute is the attribute on a argument instance that you want this condition to check. operator is some sort of check applied against that attribute. For instance, is the UserArgument.age greater than some value? Equal to some value? Within a range of values? Etc.

Let's say you wanted a Condition that checks if the user's age is > 65 years old? You would construct a Condition that way:

from gutter.client.operators.comparable import MoreThan

condition = Condition(argument=UserArgument, attribute='age', operator=MoreThan(65))

This Condition will be true if any input instance has an age that is more than 65.

Please see the gutter.operators for a list of available operators.

Conditions can also be constructed with a negative argument, which negates the condition. For example:

from gutter.client.operators.comparable import MoreThan

condition = Condition(argument=UserArgument, attribute='age', operator=MoreThan(65), negative=True)

This Condition is now True if the condition evaluates to False. In this case if the user's age is not more than 65.

Conditions then need to be appended to a switch instance like so:

switch.conditions.append(condition)

You can append as many conditions as you would like to a switch, there is no limit.

Checking Switches as Active

As stated before, switches are checked against input objects. To do this, you would call the switch's enabled_for() method with a User instance, for instance. You may call enabled_for() with any input object, it will ignore inputs for which it knows nothing about. If the Switch is active for your input, enabled_for will return True. Otherwise, it will return False.

gutter.active() API

A common use case of gutter is to use it during the processing of a web request. During execution of code, different code paths are taken depending on if certain switches are active or not. Often times there are multiple switches in existence at any one time and they all need to be checked against multiple arguments. To handle this use case, Gutter provides a higher-level API.

To check if a Switch is active, simply call gutter.active() with the Switch name:

gutter.active('my cool feature')
>>> True

The switch is checked against some number of input objects. Inputs can be added to the active() check one of two ways: locally, passed in to the active() call or globally, configured ahead of time.

To check against local inputs, active() takes any number of input objects after the switch name to check the switch against. In this example, the switch named 'my cool feature' is checked against input objects input1 and input2:

gutter.active('my cool feature', input1, input2)
>>> True

If you have global input objects you would like to use for every check, you can set them up by calling the Manager's input() method:

gutter.input(input1, input2)

Now, input1 and input2 are checked against for every active call. For example, assuming input1 and input2 are configured as above, this active() call would check if the Switch was enabled for inputs input1, input2 and input3 in that order:

gutter.active('my cool feature', input3)

Once you're doing using global inputs, perhaps at the end of a request, you should call the Manager's flush() method to remove all the inputs:

gutter.flush()

The Manager is now setup and ready for its next set of inputs.

When calling active() with a local inputs, you can skip checking the Switch against the global inputs and only check against your locally passed in inputs by passing exclusive=True as a keyword argument to active():

gutter.input(input1, input2)
gutter.active('my cool feature', input3, exclusive=True)

In the above example, since exclusive=True is passed, the switch named 'my cool feature' is only checked against input3, and not input1 or input2. The exclusive=True argument is not persistent, so the next call to active() without exclusive=True will again use the globally defined inputs.

Signals

Gutter provides 4 total signals to connect to: 3 about changes to Switches, and 1 about errors applying Conditions. They are all available from the gutter.signals module

Switch Signals

There are 3 signals related to Switch changes:

  1. switch_registered - Called when a new switch is registered with the Manager.
  2. switch_unregistered - Called when a switch is unregistered with the Manager.
  3. switch_updated - Called with a switch was updated.

To use a signal, simply call the signal's connect() method and pass in a callable object. When the signal is fired, it will call your callable with the switch that is being register/unregistered/updated. I.e.:

from gutter.client.signals import switch_updated

def log_switch_update(switch):
    Syslog.log("Switch %s updated" % switch.name)

switch_updated.connect(log_switch_updated)

Understanding Switch Changes

The switch_updated signal can be connected to in order to be notified when a switch has been changed. To know what changed in the switch, you can consult its changes property:

>>> from gutter.client.models import Switch
>>> switch = Switch('test')
>>> switch.concent
True
>>> switch.concent = False
>>> switch.name = 'new name'
>>> switch.changes
{'concent': {'current': False, 'previous': True}, 'name': {'current': 'new name', 'previous': 'test'}}

As you can see, when we changed the Switch's concent setting and name, switch.changes reflects that in a dictionary of changed properties. You can also simply ask the switch if anything has changed with the changed property. It returns True or False if the switch has any changes as all.

You can use these values inside your signal callback to make decisions based on what changed. I.e., email out a diff only if the changes include changed conditions.

Condition Application Error Signal

When a Switch checks an input object against its conditions, there is a good possibility that the Argument value may be some sort of unexpected value, and can cause an exception. Whenever there is an exception raised during Condition checking itself against an input, the Condition will catch that exception and return False.

While catching all exceptions is generally bad form and hides error, most of the time you do not want to fail an application request just because there was an error checking a switch condition, especially if there was an error during checking a Condition for which a user would not have applied in the first place.

That said, you would still probably want to know if there was an error checking a Condition. To accomplish this, gutter-client provides a condition_apply_error signal which is called when there was an error checking a Condition. The signal is called with an instance of the condition, the input which caused the error and the instance of the Exception class itself:

signals.condition_apply_error.call(condition, inpt, error)

In your connected callback, you can do whatever you would like: log the error, report the exception, etc.

Namespaces

gutter allows the use of "namespaces" to group switches under a single umbrella, while both not letting one namespace see the switches of another namespace, but allowing them to share the same storage instance, operators and other configuration.

Given an existing vanilla Manager instance, you can create a namespaced manager by calling the namespaced() method:

notifications = gutter.namespaced('notifications')

At this point, notifications is a copy of gutter, inheriting all of its:

  • storage
  • autocreate setting
  • Global inputs
  • Operators

It does not, however, share the same switches. Newly constructed Manager instances are in the default namespace. When namespaced() is called, gutter changes the manager's namespace to notifications. Any switches in the previous default namespace are not visible in the notifications namespace, and vice versa.

This allows you to have separate namespaced "views" of switches, possibly named the exact same name, and not have them conflict with each other.

Decorators

Gutter features a @switch_active decorator you can use to decorate your Django views. When decorated, if the switch named as the first argument of the @switch_decorated decorator is False, a Http404 exception is raised. However, if you also pass a redirect_to= kwarg, the decorator will return a HttpResponseRedirect instance, redirecting to that location. If the switch is active, then the view runs as normal.

For example, here is a view decorated with @switch_active:

from gutter.client.decorators import switch_active

@switch_active('cool_feature')
def my_view(request):
    return 'foo'

As stated above, if the cool_feature switch is inactive, this view will raise a Http404 exception.

If, however, the decorator was constructed with a redirect_to= kwarg:

@switch_active('cool_feature', redirect_to=reverse('upsell-page'))

Then a HttpResponseRedirect instance will be returned, redirecting to reverse('upsell-page').

Testing Utilities

If you would like to test code that uses gutter and have the gutter manager return predictable results, you can use the switches object from the testutils module.

The swtiches object can be used as both a context manager and a decorator. It is passed kwargs of switch names and their``active`` return values.

For instance, with this code here, by passing cool_feature=True to the switches object as a context manager, any call to gutter.active('cool_feature') will return True. Calls to active() with other switch names will return their actual live switch status:

from gutter.client.testutils import switches
from gutter.client.default import gutter

with switches(cool_feature=True):
    gutter.active('cool_feature')  # True

And when using switches as a decorator:

from gutter.client.testutils import switches
from gutter.client.default import gutter

@switches(cool_feature=True)
def run(self):
    gutter.active('cool_feature')  # True

Additionally, you may pass an alternate Manager instance to switches to use that manager instead of the default one:

from gutter.client.testutils import switches
from gutter.client.models import Manager

my_manager = Manager({})

@switches(my_manager, cool_feature=True)
def run(self):
    gutter.active('cool_feature')  # True