I've seen a few large Rails codebases that included a state machine library like mentioned in the article. Every single one was worse because of it. It pushes the code down a path where hidden hooks run when magic methods are called. At this point in my career I'm just done with that type of "clever" code.
The article starts with examples that just define the state machine by hand. This is a much better approach and scales to larger code much better. You can grep "def start!" and just read what the method does. A state machine DSL is really not providing much value and eventually just gets in the way.
Funny, I have spent a ton of time around Rails code where I dearly wished we had a state machine, because the alternative was an unorganized cluster of home-grown "state transition" glue without any consistent way of handling it, with all the weird-ass edge cases and split-brain BS that come with it.
The quip I like best for this is: “Any sufficiently complicated model class contains an ad-hoc, informally-specified, bug-ridden, slow implementation of half of a state machine.”
Lots of models I've encountered eschew being organized as state machines in favour of having "if salad" strewn throughout their code. That being said, refactoring to a simple state machine and trying to maintain it as such in perpetuity isn't always the correct solution.
Sometimes, a hierarchal state machine is needed, and if it is expressed as a simple state machine, it's just as messy.
Sometimes, a portion of it needs to be a state machine, and the right thing to do is to delegate some of the methods to a strategy that reflects the current state, but not all of them.
Sometimes, the whole thing is just too fat, and a state machine won't save it, the right thing to do is to get very aggressive about refactoring to a composite object.
Any time you have a big, messy model, it's very easy to write a blog post espousing a single solution, like this one:
But the reality is that a big, messy model is always going to be some kind of problem, and unless you can break it down into parts, you're going to have a problem. A state machine is conceptually a way to break a big thing into parts based on its "state," but that's just one approach to factoring.
p.s. Another problem is that even if a simple state machine is the right answer, "rolling your own" usually isn't. Grab a well-tested and documented library already. This isn't your passion project, this is industrial programming. Rolling your own is one of the best ways to learn how state machines work. Once you've learned how, reach for a professional tool.
State machines aren’t exactly rocket science though, I don’t necessarily think writing one for your specific preferences/use cases is as bad of an idea as, say, writing your own ORM. It’s a pretty well understood concept.
They aren’t exactly rocket science, agreed, but there’s an interesting trap here:
Implementing classes on top of pre-ES2015 JS wasn’t exactly rocket science either, so there was a “Cambrian Explosion” of home-grown implementations and OSS libraries all over the place. Frameworks like Ember.js rolled their own too, and everything was incompatible with everything else, plus while the basic principles were the same, the details differed from implementation to implementation.
And when organizations are rolling their own, they tend to do just enough to serve their immediate needs. As their needs grow, different internal contributors add patches and bodges to the implementation, using different approaches.
Over the long haul, even though the concept is simple enough to roll your own state machine, it’s often a win to build on top of something which is well-supported and can grow with your organization.
I've been on the receiving and perpetuating end of state machines and it's painful to have to deal with one that started years before you got involved and easy to think that you can do better. I've got a modest one in a side project and due to it being a side project I touch it once every couple of weeks and already it is making me uncomfortable. And it's dirt simple and I wrote it!
There's a take on state machines in a railsconf talk further down in the thread and it seems like an awesome way to reap a lot of the benefits while keeping a grip on the complexity. Sometimes complexity can't be buried safely and you kind of need to frame it instead. I'd argue that state machines bury that complexity but breaking those states into ruby objects w/ their own validation frames and delineates them instead.
More generally, I would recommend to avoid any DSL. All DSLs are a downgrade over ruby language and there is a really narrow list of opportunities where they are an upgrade.
Everyone knows ruby when working on a Rails application. Everyone knows what a class, an instance is and what methods are.
Ruby is extendable, incredibly flexible and it doesn't need "special options" to be extended.
DSLs every time attempt to recreate the flexibility of ruby and require studying a new mini language.
When I'm in a dsl, I don't know if I can use an instance variable, if I need a method I don't know if I can have one.
More often then not, this question is not answered by the documentation and digging in the source code is needed.
Most libraries with a DSL violate the first rule of a DSL: a DSL method should delegate to a normal object performing the operation, so that DSL API has also a corresponding normal ruby API. If you do this, your DSL is now extendable and replaceable, it can also be integrated into other libraries.
As for state machines, the libraries recommended push toward objects violating the single responsibility principle. A state machine has already a super important responsibility: state-transitions. That's already a lot of work, if it has to trigger side effects, now it does way too much.
Some alternatives: pub/sub! let the state machine be observable, whatever needs to do work when a certain state is reached, can do it when the event is published
Alternatively as someone recommended below, service objects take care of: transition, persistance of the new state and side effects. If a new action is needed, a new service object can be written. The state machine can be used to validate the transition while the service object takes care of the rest.
Using explicit state action services, which are well tested and intentionally invoked are the ways to go. They encapsulate the code, and don't create these callback hells when things inevitably get complex.
Counterpoint: State machines in Ruby should be used sparingly, and only when necessary, because they start out quick and easy, but almost always grow into things that are big, ugly, and difficult to refactor or remove.
Here are somebody's notes from a Railsconf talk that made this point, plus a link to the video:
Of all the things I've seen in my time writing software I can qualify that some of the most fucked up things involved state machines. The worst of which was a 200 state hand written one in C for an embedded measurement system. There was no documentation, no comments and the states were numbered rather than named. The guy who wrote it originally was "anti RTOS" and had died after riding his motorbike recklessly leaving this shit for everyone to deal with.
It took me two weeks to do a trivial change because I had to reverse engineer a ton of it. And the only debugger I had was a couple of LEDs on the board I could wiggle on and off.
Sounds like a confused author or potentially an implementation-driven architectural decision (eg. regulatory or policy requirement, direct map from exposed states of a prior implementation that had been black box reverse engineered, etc.). AFAIK the normal way to implement state machines in embedded since ages yore is to use interlocking networks of state machines per isolate subsystem. At least, that's what I do and that's what seems natural in C and assembly where concerns such as manual memory management, code re-use and programmer cognitive load would all tend toward such an architecture. What was the hardware?
Was Coldfire + IAR. The design was driven entirely by idiocy and showboating. I was hired just to limp it along a bit further while they spun a new board and software as the MCU was obsolete anyway.
Thanks for sharing this. For other people who are hovering over this and unsure if they want to watch a 40 minute video, it's time stamped and the two parts I found super valuable were the 'Anti-Pattern: Custom Methods' and the 'Anti-Pattern: State Machines' sections. The TL;DR there is it's alright to create a controller that manipulates a single field on a single model. If it is enough of a resource to expose it by a custom url (think update password) it is enough of a resource to get its own controller.
And going beyond that, to the state machine, if each state (or really, value of a particular field) is important enough to have it's own contextual validation, transition points, etc, then it's important enough to expose each state as a resource with it's own controller.
Super cool stuff and I've got some thinking to do about some side projects now.
Indeed, a coworker at a previous job was fascinated/fixated with state machine patterns. After spending two weeks trying to force a process we were paired on into a state machine pattern and failing, I rewrote it in an afternoon without the state machine stuff.
It started out small and easy, when there were two states. A third state was manageable but to me already looked like it was going the wrong direction. The fourth state (out of IIRC 6) was never finished.
A lot of the comments here are rails oriented and anti state machine. In the context of rails that may make sense, but I do a lot of non rails ruby, and non object oriented Dev style, and state machines are a great way to write solid systems where the allowed state and transitions are explicit and easy to reason with.
A pretty big +1 to this. My team (at Crunchy Data building Crunchy Bridge[1]) and similar teams at prior Citus Data and Heroku did a lot of Ruby and a lot of state machines. In all cases it was actually less Rails, but that may be because the team really prefers sequel as the ORM to ActiveRecord, though Rails does a bunch that doesn't really apply to such a workflow. We wrote about some of how we used state machines to power our database as a service and the full control plane previously[2]. Another big piece of the approach and stack that helps today is leveraging Sorbet[3] and our approach to test coverage, we probably should blog a little more on some of this but it helps quite a bit.
Clearly some quite senior developers there. Makes me wonder if the people hating on state machines have been at the mercy of junior devs or people trying it out for the first time, instead of a well planned state machine by someone with experience.
Full disclosure, I think I wrote my first state machine in around 1996. So yea... Grey hair etc. Also it was a big part of my computer science degree, studied formally, so perhaps that's why I'm biased in favour of them.
Honestly state machines are fantastic in Rails too. My last company built Statesman (https://github.com/gocardless/statesman/) and being able to lean on it to prevent you getting into invalid states is fantastic. You also get the bonus of tracking the history of states your resources went through (which is especially useful when you're dealing with payments).
At some point you'll have to think about query performance on the state transition table, but it'll go further than you think and is firmly in the realm of problems of success.
In a previous life I implemented a business system in rails with a core state machine. The staff who used it could fundamentally understand the states and transitions, bringing them closer to understanding how the whole system worked. Not all benefits or costs of a technique are limited to the developer
Sure. I am working on the firmware for a camera. Believe it or not it's a bunch of ruby services coordinated via redis pubsub. But at the core is the main system state which is a service in itself which is the main state machine.
This pattern is something which I've evolved over the years and it's fantastic (if I do say so myself).
I worked at a business (in ruby) where the entire core logic of the app was a state machine, as an idea the thought was good, a state machine modeled order flows pretty well, but in practice it was very slow and the amount of side effects that had to be considered as the business grew just compounded this.
It got to the point where this state machine approach was a real bottleneck and though I left before it was resolved last I heard they had to put real engineering time in to move off of it.
On the flip side, I've built several production apps with AASM or Workflow that made adapting to changing business logic a breeze. If you're given a flow-chart you can translate that 1-1 and when that chart changes it's easy to adjust.
For sure, this isn't an indictment against state machines, my message was exactly to be careful when making the decision to use them. They fit a lot of workflows and I've also worked in environments where they were not at all a bottlenecks, and regardless of the performance bottleneck it wasn't hard to reason about or adjust which for a long time meant more than performance for that business.
I guess I could amend my statement to "be careful about which library or implementation of a state machine you go with"
This has been my experience as well with AWS Step Functions. Obviously your use case is critical to the architecture decision. For us it's made automating workflows a breeze.
Seems to be a lot of anti-state machine comments here. Are these people trying to use state machines to model their entire application state? The example used in the article is a very practical finite state machine centered around order processing. It's very unlikely these steps in an order lifecycle are going to change significantly. I suspect a lot of organizations jumped on all-in with state machines when they got a resurgence in popularity with x-state library etc. Trying to model your entire application as a state machine just so you can show off an incomprehensible state chart or something is a recipe for disaster.
Creator of XState here - I agree. State machines and statecharts should be used to model parts of your application that need it, and they can potentially communicate with each other. Fitting the entire app in a single state machine is just like fitting all of the app logic in a single function; it's going to be incomprehensible.
I get how these state machine gems can sound like a great idea if what you need is taming some state changes that are just a little too much to handle by hand, but they can also delay the realization that you're doing something wrong.
State changes are hard and most involve more than object, there are usually associations that need to be taken care of, validated or maybe also transitioned to a different state, there can be a myriad of conditions and things can go wrong at any moment. And I feel that being able to define state transitions this easily, almost as an afterthought, with a simple DSL, doesn't do all this justice (just like those friendly before/after callbacks can turn into massive spaghetti abominations).
The way I do it is that any state transition more complex than checking two attributes and updating a third one is an object. Transitioning a Campaign object from :reservation to :order? That's a new PORO called CampaignOrder, with a #perform! public method or something. All the conditions, assumptions and side effects have their private methods (check_showings!, update_occupancy!). Passed a dirty record into something delicate? That's a raise, I want something clean and valid to work with. Is this something where we expect a dirty record? Go right ahead, we'll save it later in the transition, plus it will be documented right in the constructor. Does it need to raise? It will. Custom error messages on any record? Right here in this private method. Does any model need to know anything about the state transition or own any related code? Nope.
The article starts with examples that just define the state machine by hand. This is a much better approach and scales to larger code much better. You can grep "def start!" and just read what the method does. A state machine DSL is really not providing much value and eventually just gets in the way.