Declarative, explicit, tool-friendly finite state machines in Python

# friendly_states

This is a Python library for writing finite state machines in a way that's easy to read and prevents mistakes with the help of standard linters and IDEs.

You can try this README in an interactive notebook in binder.

## Introduction

Here's a simple example of declaring a state machine:

from __future__ import annotations
from friendly_states import AttributeState

class TrafficLightMachine(AttributeState):
is_machine = True

class Summary:
Green: [Yellow]
Yellow: [Red]
Red: [Green]

class Green(TrafficLightMachine):
def slow_down(self) -> [Yellow]:
print("Slowing down...")

class Yellow(TrafficLightMachine):
def stop(self) -> [Red]:
print("Stop.")

class Red(TrafficLightMachine):
def go(self) -> [Green]:
print("Go!")

TrafficLightMachine.complete()


That looks like a lot of code, but it's pretty simple:

1. TrafficLightMachine declares the root of a state machine.
2. There are three states, which are subclasses of the machine: Green, Yellow, and Red.
3. Each state declares which transitions it allows to other states using functions with a return annotation (->). In this case, each state has one transition to one other state. For example, Green can transition to Yellow via the slow_down method.
4. TrafficLightMachine.complete() makes the machine ready for use and checks that everything is correct. In particular it verifies that the Summary class inside TrafficLightMachine matches the actual states and transitions declared. Each line of the summary shows all the possible output states from each state.

As we'll see later, a lot of the boilerplate can be generated for you, so don't worry about the effort of writing all those classes.

To use this machine, first we need an object that can store its own state. When using this library, state can be stored however you want, but the way we've declared our machine means it expects an attribute called state:

class TrafficLight:
def __init__(self, state):
self.state = state

light = TrafficLight(Green)
assert light.state is Green


The magic happens when making transitions:

Green(light).slow_down()
Yellow(light).stop()
assert light.state is Red

Slowing down...
Stop.


Making an instance of a state class such as Green(light) automatically checks that light is actually in the Green state and raises an exception if it's not. Then you can call methods just like any other class. The state will automatically be changed based on the method's return annotation.

Why is this exciting? Because it's obvious to tools which transitions are available to each state. Your IDE can autocomplete Green(light).slo for you, and Green(light).stop() will be highlighted as an error before you even run the code.

Compare this with the popular library transitions. States, transitions, and callbacks are all declared using strings, so it's easy to make typos and tools can't help you. In fact you have to stop your tools from warning you about all the attributes it magically generates which you have to use. There's no easy way to see all the transitions or output states from a particular state. Callbacks can be declared far away from transitions.

By contrast, when using friendly_states, there are no strings or magic attributes anywhere. Code is always naturally grouped together: all the transitions specific to a state appear inside that class, and logic related to a transition is in that function where you'd expect it.

Here is the equivalent of the main example in the transitions doc implemented using friendly_states.

## Basic usage steps

1. Ensure that you are using Python 3.7+.

2. Install this library with pip install friendly_states.

3. Add the line from __future__ import annotations at the top of the file where you define your state machine.

4. Declare the root of your state machine by inheriting from the appropriate class (usually AttributeState, see the BaseState section) and setting is_machine = True in the body, like so:

from __future__ import annotations
from friendly_states import AttributeState

class MyMachine(AttributeState):
is_machine = True

class Summary:
Waiting: [Doing, Finished]
Doing: [Checking, Finished]
Checking: [Finished]
Finished: []


Declaring a summary is optional, but highly recommended. The class must be named Summary. It should declare the name of every state your machine will have, each annotated with a list of every state that can be reached directly from that state by any transition.

When you call MyMachine.complete(), this summary will be checked, and an exception will be raised explaining the problem if it doesn't match your state classes. In particular if a class is missing the exception will contain generated source code for that class so you can spend less time writing boilerplate. Let's try it now:

try:
MyMachine.complete()
except Exception as e:
print(e)

Missing states:

class Waiting(MyMachine):
def to_doing(self) -> [Doing]:
pass

def to_finished(self) -> [Finished]:
pass

class Doing(MyMachine):
def to_checking(self) -> [Checking]:
pass

def to_finished(self) -> [Finished]:
pass

class Checking(MyMachine):
def to_finished(self) -> [Finished]:
pass

class Finished(MyMachine):
pass


You can copy the code above and will have a working state machine matching the summary. It's usually not exactly what you need, but it should save you a lot of time for the next step.

1. Write a class for each state. Make sure you call .complete() at the end.
class Waiting(MyMachine):
def start_doing(self) -> [Doing]:
print('Starting now!')

def skip(self) -> [Finished]:
pass

class Doing(MyMachine):
def done(self, result) -> [Checking, Finished]:
self.obj.result = result
if self.obj.needs_checking():
return Checking
else:
return Finished

class Checking(MyMachine):
def check(self) -> [Finished]:
print('Looks good!')

class Finished(MyMachine):
pass

MyMachine.complete()


State classes must inherit (directly or indirectly) from the machine class, e.g. class Waiting(MyMachine):. A class can have any number of transitions. Waiting has two separate transitions, while Finished has none, meaning you can't leave that state. It can also have any other methods or attributes which are not transitions, like a normal class.

A transition is any method which has a return annotation (the bit after the -> in the function definition) which is a list of one or more states that will be the result of this transition. For example, this code:

class Waiting(MyMachine):
def start_doing(self) -> [Doing]:


means that start_doing is a transition from the state Waiting to the state Doing, and calling that method will change the state.

The transition Doing.done demonstrates several interesting things:

• A transition can have multiple possible output states. In that case the method must return one of those states to indicate which one to switch to.
• Transitions are just like normal methods and can accept whatever arguments you want.
• States have an attribute obj which is the object that was passed to the class when it was constructed. This lets you interact with the object whose state is changing.
1. The machine doesn't store the state itself, make a different class for that:
class Task:
def __init__(self):
self.state = Waiting
self.result = None

def needs_checking(self):
return self.result < 5



Our example machine expects to find an attribute called state on its objects, as we've provided here. If you have different needs, see the BaseState section.

1. To change the state of your object, you first need to know what state it's in right now. Sometimes you'll need to check, but often it'll be obvious in the context of your application. For example, if we have a queue of fresh tasks, any task we pop from that queue will be in state Waiting.

Construct an instance of the correct state class and pass your object:

Waiting(task)

Waiting(obj=<__main__.Task object at 0x1173677f0>)


If you get the current state of a task wrong, that means there's a bug in your code! It will throw an exception before you can even call any transitions:

try:
except Exception as e:
print(e)

<__main__.Task object at 0x1173677f0> should be in state Doing but is actually in state Waiting

1. Once you have an instance of the correct state, call whatever methods you want on it as usual. If the method is a transition, the state will automatically be changed afterwards:
Waiting(task).skip()


If you try calling a transition that doesn't exist for a state, the library doesn't even need to do anything. You'll just get the plain Python error you always get when calling a non-existent method, and your IDE/linter will warn you in advance:

task = Task()
try:
except AttributeError as e:
print(e)

'Waiting' object has no attribute 'done'


Here are the other 2 possible paths for a task from waiting to finished:

task = Task()

Starting now!
Looks good!

task = Task()
# The result '7' doesn't need checking

Starting now!


## Abstract state classes

Sometimes you will have common behaviour that you need to share between state classes. In that case you can use class inheritance. Here's what you need to be aware of:

1. You can't inherit from actual state classes.
2. Transitions must live in classes that inherit (directly or indirectly) from the machine class.
3. Classes that inherit from the machine (and thus can have transitions) but do not represent actual states (and thus can be inherited from) should have is_abstract = True in their body.

Here's an example:

class TaskMachine2(AttributeState):
is_machine = True

class Summary:
Waiting: [Doing, Finished]
Doing: [Finished]
Finished: []

is_abstract = True

def finish(self) -> [Finished]:
pass

class Waiting(Unfinished):
def start_doing(self) -> [Doing]:
pass

class Doing(Unfinished):
pass

pass



Here the Waiting and Doing states are both subclasses of Unfinished so they get the finish transition for free, and you can see the result of this in the summary.

If you have an object which is in one of these states but you're not sure which, and you'd like to call the finish transition, just use the Unfinished abstract class:

import random

for i in range(100):

if random.random() < 0.5:

# Now the task might be either Waiting or Doing


This will still fail if you try it on a finished task:

try:
except Exception as e:
print(e)

<__main__.Task object at 0x11401cb00> should be in state Unfinished but is actually in state Finished


However, you may want to use methods and attributes in the class representing the actual current state. In particular you might have overridden methods in the concrete state classes and want the correct implementation to run. To allow this, instances of a state automatically change the class to the actual state of the object:

task = Task()


## BaseState - configuring state storage and changes

At the root of all the classes in the library is BaseState, which has two abstract methods get_state and set_state. Subclasses determine how the object stores state and what should happen when it changes.

For example, here is the start of AttributeState which we've been using as the base of our machines so far:

class AttributeState(BaseState):
attr_name = "state"

def get_state(self):
return getattr(self.obj, self.attr_name)


You can declare a different attr_name in your machine class to store the state in that attribute of the object.

If you're storing your state in a dict or similar object, you can instead use MappingKeyState, which starts like this:

class MappingKeyState(BaseState):
key_name = "state"

def get_state(self):
return self.obj[self.key_name]


It can often be useful to override set_state to add extra common behaviour when the state changes, e.g:

class PrintStateChange(AttributeState):
def set_state(self, previous_state, new_state):
print(f"Changing {self.obj} from {previous_state} to {new_state}")
super().set_state(previous_state, new_state)


set_state is called after the transition method completes.

So overall, your class hierarchy typically looks something like this:

BaseState <- AttributeState <- Machine <- Abstract States <- Actual states

Machines, states, and transitions have a bunch of attributes that you can inspect:

# All concrete (not abstract) states in the machine
assert TaskMachine2.states == {Doing, Finished, Waiting}

# All the transition functions available for this state
assert Waiting.transitions == {Waiting.finish, Waiting.start_doing}

# The transition functions defined directly on this class, i.e. not inherited
assert Waiting.direct_transitions == {Waiting.start_doing}

# Possible output states from this transition
assert Waiting.start_doing.output_states == {Doing}

# All possible output states from this state via any transition
assert Waiting.output_states == {Doing, Finished}

# Root of the state machine

# Booleans about the type of class
assert Waiting.is_state and not Unfinished.is_state
assert Unfinished.is_abstract and not Waiting.is_abstract


### Slugs and labels

Classes also have slug and label attributes which are mostly for use by Django but may be useful elsewhere. slug is for data storage and label is for human display.

By default, slug is just the class name, while label is the class name with spaces inserted. Both can be overridden by declaring them in the class.

assert Waiting.slug == "Waiting"


## Troubleshooting

If things are not working as expected, here are some things to check:

• Check the attributes machine.states, state.transitions, and transition.output_states to see if they look as expected.
• If you override set_state, remember to call super().set_state(...), unless you want to prevent the state from changing or you're subclassing BaseState directly.
• Check that the annotation on your transition is a list, i.e. it starts and ends with []. For example this will not be recognised as a transition:
    def start_doing(self) -> Doing:

• If your transition has any decorators, make sure that the decorated function still has the original __annotations__ attribute. This is usually done by using functools.wraps when implementing the decorator.
• Make sure that the object stores state the way the machine expects. Typically you'll be using AttributeState and you should make sure that attr_name ("state" by default) is correct. Note that a typical machine expects objects to have just one way of storing state - you can't use the same machine to change state stored in different attributes. To overcome this, see the recipe 'Dynamically changing the attribute name'.
• Check that you've inherited your classes correctly. All states need to inherit from the machine.

## Recipes

The API of friendly_states is intentionally minimal. Here is how you can do some more complicated things.

### Construct and draw a graph

Here is how to create a graph with the popular library networkx:

import networkx as nx
machine = MyMachine

G = nx.DiGraph()
for state in machine.states:
for output_state in state.output_states:

print(G.nodes)
print(G.edges)

[Waiting, Finished, Doing, Checking]
[(Waiting, Finished), (Waiting, Doing), (Doing, Finished), (Doing, Checking), (Checking, Finished)]


To draw the graph with matplotlib:

%matplotlib inline
nx.draw(G, with_labels=True, node_color='pink')


To label each edge requires some more work:

edge_labels = {}
G = nx.DiGraph()
for state in machine.states:
for transition in state.transitions:
for output_state in transition.output_states:
edge = (state, output_state)
edge_labels[edge] = transition.__name__

pos = nx.spring_layout(G)
nx.draw(G, pos, with_labels=True, node_color='pink')
nx.draw_networkx_edge_labels(G, pos, edge_labels=edge_labels);


### Creating multiple similar machines

Suppose you want to create several machines with similar states and transitions with duplicating code. You may think you can use inheritance somehow, but that won't work, and in fact a machine can't subclass another machine. Instead you must make a function which creates the classes locally. There are many ways this could be one depending on your needs. Here's an example with identical machines except that one has an additional state:

from types import SimpleNamespace

def machine_factory():
class Machine(AttributeState):
is_machine = True

class CommonState1(Machine):
def to_common_state_2(self) -> [CommonState2]:
pass

class CommonState2(Machine):
pass

return SimpleNamespace(
Machine=Machine,
CommonState1=CommonState1,
CommonState2=CommonState2,
)

machine1 = machine_factory()

class DifferentState(machine1.Machine):
def to_common_state_2(self) -> [machine1.CommonState2]:
pass

machine1.Machine.complete()

@machine1.Machine.check_summary
class Summary:
CommonState1: [CommonState2]
CommonState2: []
DifferentState: [CommonState2]

machine2 = machine_factory()
machine2.Machine.complete()

@machine2.Machine.check_summary
class Summary:
CommonState1: [CommonState2]
CommonState2: []


### Dynamically changing the attribute name

A typical AttributeState machine can only work with one attribute name. You may need to use the same machine with different object classes that use different attributes. One way is to follow a similar pattern to above with a configurable attribute name:

def machine_factory(name):
class Machine(AttributeState):
is_machine = True
attr_name = name

...


Another option, which may be useful for more complicated situations, is to subclass AttributeState to accept the attribute name on construction:

class DynamicAttributeState(AttributeState):
def __init__(self, obj, attr_name):
# override the class attribute
self.attr_name = attr_name

# must call super *after* because it checks the state
# in the attribute with the given name
super().__init__(obj)

class Machine(DynamicAttributeState):
is_machine = True

class Start(Machine):
def to_end(self) -> [End]:
pass

class End(Machine):
pass

Machine.complete()

thing = SimpleNamespace(state=Start, other_state=Start)

assert thing.state is Start
assert thing.other_state is Start

Start(thing, "state").to_end()
assert thing.state is End
assert thing.other_state is Start

Start(thing, "other_state").to_end()
assert thing.state is End
assert thing.other_state is End


### On enter/exit state callbacks

Whenever you want to execute some generic logic on every state transition, you should override set_state. But if you put all your code in there it may get long and confusing. If you want to group this logic into your state classes whenever a transition enters or exits that state, here's a mixin that you can apply to any machine:

class OnEnterExitMixin:
def set_state(self, previous_state, new_state):
previous_state(self.obj).on_exit(new_state)
super().set_state(previous_state, new_state)
new_state(self.obj).on_enter(previous_state)

def on_exit(self, new_state):
pass

def on_enter(self, previous_state):
pass


Then use it as follows:

class Machine(OnEnterExitMixin, AttributeState):
is_machine = True

class Start(Machine):
def end(self) -> [End]:
pass

class End(Machine):
def on_enter(self, previous_state):
print(f"Ending from {previous_state}")

Machine.complete()

thing = SimpleNamespace(state=Start)
Start(thing).end()

Ending from Start


## Django integration

friendly_states can easily be used out of the box with Django. Basic usage looks like this:

from django.db import models
from friendly_states.django import StateField, DjangoState

class MyMachine(DjangoState):
is_machine = True

# ...

class MyModel(models.Model):
state = StateField(MyMachine)


StateField is a CharField that stores the slug of the current state in the database while letting you use the actual state class objects in all your code, e.g:

obj = MyModel.objects.create(state=MyState)
assert obj.state is MyState
objects = MyModel.objects.filter(state=MyState)


All keyword arguments are passed straight to CharField, except for max_length and choices which are ignored, see below.

DjangoState will automatically save your model after state transitions. To disable this, set auto_save = False on your machine or state classes.

StateField will automatically discover its name in the model and set that attr_name on the machine, so you don't need to set it. But as usual, beware that you can't use different attribute names for the same machine.

Because the database stores slugs and the slug is the class name by default, if you rename your classes in code and you want existing data to remain valid, you should set the slug to the old class name:

class MyRenamedState(MyMachine):
slug = "MyState"
...


Similarly you mustn't delete a state class if you stop using it as long as your database contains objects in that state, or your code will fail when it tries to work with such an object.

max_length is automatically set to the maximum length of all the slugs in the machine. If you want to save space in your database, override the slugs to something shorter.

choices is constructed from the slug and label of every state. To customise how states are displayed in forms etc, override the label attribute on the class.

## Project details

Uploaded source