2 releases
0.1.1 | Jul 12, 2023 |
---|---|
0.1.0 | Jul 12, 2023 |
#150 in Simulation
52KB
357 lines
caet
: Cause and effect tester
What is it?
caet
simulates a universe, operated by a judge, in which you can
test an object's behavior in response to a stream of events, while
that object, in turn, influences the universe.
Why would I want to do that?
caet
is primarily intended to prototype and validate object-oriented
analysis and design models because of the deficiencies in earlier
approaches:
- Any sort of theoretical argumentation: quick and easy (compared to actually writing code), but may be inaccurate or incomplete. Surprises.
- Writing code: slow, expensive, lock-in.
- Writing unit tests: rigid, brittle, lock-in.
- Writing integration tests: not generally applicable; scaffolding effort.
- Many elaborate approaches: at some point, you get the impression that all they care about is selling you their book.
And, when is simulation a better approach?
- When you want to test a design model, not an implementation.
- You have a proposal to break down a complex system to manageable parts. Alright, but are you sure that it's good design? How do you know?
- Well, simulate it and see what happens.
Then, how would simulation be better?
- Not too expensive: you do have to write code, but you don't have to deal with actual I/O or performance. Just define the data types that represent events and reactions, and add some mock code for the object under test (and the judge, as well; more on that later). Once you're done, you can review it, discuss it, throw it away, or keep it.
- Not locked-in: you don't to connect to existing code at all. If you know your domain, you can write the simulation code from scratch, and then throw it away or keep it as documentation. So, even though this crate uses Rust, and you must use Rust to write the simulation code, you don't have to use Rust for your actual implementation. I figured you'd appreciate this if you're already working on a big project.
- Not rigid: most testing suites simply recall a sequence of
events and check the results.
caet
allows you to define a customJudge
machine that can be as low effort as to simply be a unit tester, or as high effort as to be a fully autonomous adversarial "god" that will strategically try to break your system-under-test. You do have to do this yourself. You have to write theJudge
yourself, but it's not hard. - Not brittle: it's not about recalling a sequence of events, it's checking it using an actual computer program. So, it depends on you. If you spend the effort, you'll be rewarded with a robust test suite. If you don't, you'll get a fantastic demonstration of what real code will look like if you were to actually start writing it in under an hour (depending on the complexity, of course).
- Feedback: you can get feedback on your design model in a matter of hours, as I said! When I do this simulations thing myself, I'm always surprised how much my original design was lacking, and I liked the fast feedback cycle so I can actually build something quickly in a problem domain I'm completely unfamiliar with. Large projects are much less overwhelming than before. And, I hope you like it, too.
- Not a book: I'm not selling any books here. I'm just giving you a tool that you can use to test your ideas. Ok, there is some custom philosophy going on, but it's mostly a paraphrase of the good-old concept of object-orientation (OO), nothing proprietary.
Explain to me the philosophy behind it.
caet
is based on the following principles:
- Protocol not algorithm: We're solving an interactive problem, not a computational problem. In other words, we're building bots, not cracking LeetCode.
- Object, Universe, Time: We look at any interactive problem in three ways:
- Object: this is the system that's being modeled. All objects are specialized for each of their universes. An object can also be called a thing, entity, agent, whatever. There's many names for it, but it's all the same thing.
- Universe: this term is a bit ambiguous, but it's essentially
a view of the outside world stated in a language so simple that the
object can understand. You can, if you want, incorporate a notion
of some absolute laws this universe obeys during its evolution,
but
caet
assumes nothing by itself. It's also called the environment, surroundings, context, whatever. Again, many names, same thing. - Time: again, ambiguous concept and there's a lot you
can add to this concept for your own gain, but for our purposes, it's just a
definition of the order in which things happen. Maybe you
might want to incorporate physical timing, simultaneity, etc.
But
caet
assumes nothing except "X happened before Y."
- Singleton principle: A universe contains exactly one object. So if you wanted to model
communication, you'd have to do it roundaboutedly: first convert
their collective actions to changes in the universe, and then
let each object react to those changes. But if you really want to
model concurrency and communication,
caet
is not the tool for you, though it can be used to prototype such a system, if you hack it. - Divide and Conquer: An object may be composed internally of distinct, simpler, shorter-lived universes, akin to how the human body is composed of multiple distinct cavities, each isolated from the outside, where different organs are specialized to rest in. And, the collective evolution of these universes determines the larger compound object's behavior. This breakdown can be done recursively until the programmer decides that a prepared dependency or module can handle it, or the universe no longer requires interaction, so it can be replaced with an infallible function call.
- Lifetimes: A universe has a lifetime, and it can be born and die. Independent instances of the same type of universe can exist at the same time, each inhabited by exactly one object.
- Specialization of objects: An object can be specialized for a type of universe. This is why you can't put a liver in stomach acid and hope it'll survive. Liver works in the liver universe, and stomach works in the stomach universe.
- Observation is passive, instantaneous, and reliable: The object can "sense" or "observe" things from the universe, and this is cheap. In fact, the object can sense anything in the universe reliably and instantaneously, but only passively; it can't actively choose what to sense and when because random access amounts to violating causality by reading the future.
- Reaction is active but neither instantaneous nor reliable: Still, the object does have some agency.
The object can "act" on the universe, which is neither instantaneous nor reliable.
caet
tests exactly this part of the object's behavior: whether an object's reaction is timely, valid and appropriate. An object's action caused by an event is called a reaction. - Causality: The future can't affect the past. Causality, as
caet
understands it, is exactly the same as any of the following (and mutually so):- Encapsulation: The universe can hide information from the object, and the object can hide information from the universe.
- Inherent interactivity: Some processes may, inherently, be interactive; they could also be perhaps, but not necessarily, adversarial.
- State: Both the universe and the object have state, which is not only about keeping information but, also, about being sensitive to time.
The last bit can be a bit surprising. But, yes, all four concepts are the same thing, one single concept.
How do I use it?
First, this is more like a prototyping tool than an actual testing tool. Or, this is a tool to "test" (validate) your ideas before you actually write code. So, keep that in mind.
Also, this is a Rust crate, so you need to know Rust to use it.
Now, you need to perform some conceptual work before using this tool.
- You need to define the language of universe. (Let's call it
MyEvent
for now.) That sounds daunting, but remember,caet
actually has a pretty narrow and particular definition of a universe. So, specifically, you need to define:
- A data type that represents an event. This is the language of the universe.
(Or, pedantically, the "alphabet" part of the language.)
Note that
caet
conflates the concepts of events, messages, and commands, and doesn't distinguish between them. What this means in practice is that the return values of the object's reactions are described using the same data type as the method calls.
- Now, you need to implement the
Judge
trait for yourself. This trait takes in the set of events that the object has last emitted, judges these events for acceptance, maintains and evolves the internal universe, and then---if it decides to keep going--- generates a new event (a challenge) for the object. Or, halts the simulation and reports its findings. - There's no dedicated "Object" trait, but in the
judge
function, which is responsible for running the simulation, you can pass in a closure of this type:FnMut(MyEvent) -> Vec<MyEvent>
. The object will be reacting to a single event at a time, generating an ordered list of events in response. What happens when any of these events interleave with the outside world is up to the judge. Perhaps surprisingly, this is all you need to do to define an object. - Run the test using the
judge
function. This function takes in your Judge implementation, and the closure that represents the object. It will run the simulation until the judge decides to halt it, after which anOutcome
structure is returned. You can inspect theOutcome
to see whether the simulation was successful or not and how many iterations it took, and if not, what went wrong.
On why I decided to use a closure for the object:
I designed it that way to hide any state that the object might have because, maybe, the object you're testing is actually a real-world system, so maybe it isn't a computer program at all. So, it's not reasonable to expect the object to be a structure that you can just instantiate and pass around. It might be a physical system that you can't just copy and paste, after all.
So, let's summarize this:
- Define all events in a data type. (Let's call it
MyEvent
.) Commands, method calls, signalling, return values, whatever. Cram them all into this data type. - Implement the
Judge
trait for yourself. The judge is like a god: It runs an adversarial universe that tries to break your object by strategically generating events that the object will have to react to. (Or, it can be an indifferent universe that just recalls a pre-recorded sequence of events. You decide.) - Define the object as a closure of type
FnMut(MyEvent) -> Vec<MyEvent>
. The return value is the set of events---if any---emitted by the object listed in the order they were emitted. The goal of the object is to emit the events, in the correct order, that the judge accepts. - Run the simulation using the
judge
function by passing ownership of the judge and the object closure. This function will let them fight and return anOutcome
structure that you can inspect.
One thing you should note is that the Judge
trait is simply unaware of object types, and objects also don't know about the Judge. The only thing that connects the two is the judge
function.
That means, any judge can test any object if they agree to use the same event type,
MyEvent
in this case. Hopefully, this is a powerful feature because it means
you can test your object against multiple judges, and you can also test
multiple objects against the same judge.
Where to find examples
Perhaps looking at examples will make this clearer.
As for examples, there's the main page of the documentation, and there's also the unit tests for the crate itself.
The documentation on the main page is simple but doesn't cover all the features. The unit tests are more comprehensive, but they're also more complicated. I recommend you start with the documentation and then move on to the unit tests.