1. I can’t say I have found any real downsides yet. We have had to make improvements to support new capabilities, but it has been surprisingly easy to do so. Our most recent improvement was to formalize the concept of MVVM so that our ViewModels essentially acted in the same way as our Feature types. It allows us to continue using almost the same pattern at individual screen that we have been using at the feature level.
Our next big effort will be fitting the architecture into SwiftUI, which I think will actually be also straightforward since SwiftUI is extremely functional.
2. Onboarding is probably the biggest “downside”. I would say there is absolutely a learning curve that requires some investment but everyone I speak to about it has been pretty positive after getting over that hurdle. And it feels like our productivity is very high.
3. Android does not follow the same pattern. I am planning on sharing your article with the team though. :-)
1. Create a struct (value type in Swift) that implements a protocol (interface) to serve as your ViewModel
2. The main methods to implement are both called “update”. One is called automatically by our infrastructure when Feature state (state that exists across screens) changes. The other is called by the developer in their view controllers to pass along user actions (or other UIKit events).
3. The developer subscribed to all VM changes and updates the view in response. This subscription is done via Combine which is Apple’s FRP framework.
I am planning on writing more about this and the rest of our architecture in the future, hopefully.
Hey, I can't say I have ever worked on such a project, it sounds like an interesting problem space.
I am assuming when a user interacts with one control button / dial / interaction it has a cascade effect on other controls / displays etc?
I am also assuming you already need to have a data structure representing the complete state of the system no?
There are multiple interesting aspects here, like are you / do you need to utilise complex state machines here? Just the subject of imperative vs functional state-machines is interesting and a small part of the wider approach.
Another interesting area is it depends on the language and execution environment you are operating in and how expensive operations over immutable data structures are.
>I am assuming when a user interacts with one control button / dial / interaction it has a cascade effect on other controls / displays etc?
Sometimes yes, it depends on the instrument. Some have user interaction, but most don't and are just controlled by the software I write. When I said "knobs" I didn't mean literal knobs, just all the different things these gizmos can do .. they come with thousand-page manuals.
OTOH sometimes shit breaks and you have to go in and send commands manually over telnet to unfuck it, or to gracefully shut down the operation without some delicate and expensive piece of equipment getting torn to pieces.
>I am also assuming you already need to have a data structure representing the complete state of the system no?
Nope, it's only feasible to keep track of the most salient variables. The code is very imperative: machine_a.do_this(), machine_b.do_that(), etc, sometimes for hundreds of lines. So the state changes in the background but it's not legible to us. The machines themselves probably have some representation of their own state in their firmware, but there isn't a single omniscient data structure for the whole system.
>There are multiple interesting aspects here, like are you / do you need to utilise complex state machines here? Just the subject of imperative vs functional state-machines is interesting and a small part of the wider approach.
Yeah we use finite state machines to organize our code, but it's very coarse grained. Stuff like: LoadingState, ArmatureIsMovingState, TakingMeasurementState, etc. Each state in itself will issue many different commands to different machines. It isn't really formally defined where the boundaries are, it's just an informal way to orient ourselves so we can draw a flowchart and not lose our minds from the complexity, and so that there's some way of comparing similar-but-different programs (e.g. a run that measures current and a run that measures Young's modulus will both go in a TakingMeasurementState, but obviously the commands sent will be different).
It's not very amenable to automated unit testing ... I'm not sure what that would even mean in this context. The correct behavior you're testing against would just be ... a sequence of commands to the machines. Which is what the code does already. A unit test would be tautological.
>Another interesting area is it depends on the language and execution environment you are operating in and how expensive operations over immutable data structures are.
We're using Python. I know that pyresistent exists, but honestly we could probably get away with doing naive deepcopies for every single operation and it wouldn't matter for performance. The runtime is completely dominated by the latency of waiting for the machines to do physical stuff (no concurrency for the machine-controller code thankfully; we're not insane). Immutable data structures don't really matter if we can't write down the entire state in the first place.
----
I guess the problem is that pretty much everything interesting that the code does is a side-effect. There are a few things here and there that can be factored out to pure functions and tested, but the overwhelming bulk is the "imperative shell". It's like one of those planets where the core has cooled and the crust has frozen down to comprise most of the total mass.
Ha, I love that you mind went to that final analogy.
I think you have correctly identified the important difference between the kind of applications I have been writing (mobile applications, which are relatively small and often have a good chunk of both side-effectful code and biz logic) and what your doing, which does not sound like it would have much to gain by spitting into a core & shell.
I have not much to offer you, having not worked on code such as you describe.
I guess it depends what exactly you mean by "reactive paradigm".
In my experience the commonality found between client side mobile applications and reactive principles is essentially everything is a stream, and the UI subscribes to those streams.
The approach outlined in the post is similar in terms of the UI observes a (single) stream of data, which can just be a simple (State) -> Unit) as opposed to some reactive library stream primitive.
The approach outlines is different from all the "reactive" client side applications I have seen as there is no proliferation of stream primitives present throughout the codebase, which often seems to create complex code with very little if any user benefit.
I'm not sure what other elements you see in the post which also fall into the "reactive paradigm"?
Not targeted at the API crowd really no, more client side applications, but it seems similar principles have been applied in various places / frameworks as spoken about in other comments here.
> Maybe this isn’t targeted for the API crowd, but anybody who has spent even a few months learning modern client-server programming understands this, and web dev is pretty common these days
Yes, I'm sure your right, but there are a huge number of people who are not web devs who have not been exposed to these fundamental principles and how they may apply in other operating environments :)
Thanks. Others have said similar, but more of a request for an additional post expanding on that aspect. Until then, there is a little more information in a previous related post https://doridori.github.io/Android-Architecture-Runtime/
Your totally right, it's pretty similar conceptually to Elm in many ways, amazingly I only discovered Elm in the last 8 weeks or so after writing applications in this way for the last few years, maybe not surprisingly as I have been a mobile developer for 13 years and Elm is not big in the mobile space!
> Application is all about transitions of state, yes, but in context.
Like the other commentor says, it would be great to hear more about what you mean exactly by "in context" here.
I blame myself for this misunderstanding as I wrote the article. I certainly am not suggesting that an application is a function, rather an application can be represented by single function as its core.
There is an important distinction here as the production applications I have utilised this architecture in certainly do not literally behave as functions to the external user in the manner you describe, the user interaction is in no way affected by the architectural choice.
As others have eluded to this core pure function is really just a boundary between the external side-effect laden world (e.g. UI / network / sensors / OS etc) and the applications business / domain logic and NOT the boundary between the OS and the application which seems to be what you are describing.
I appreciate your input and the discussion it is provoking.
And apologies for the misunderstanding due to my terseness, which is partly due to me having written about this in great depth on a number of occasions, here are two examples:
- Can Programmers Escape the Gentle Tyranny of call/return?[2]
The basic point is that we are so used to programming simply being call/return (and functional is a subset of that), that it is very hard for us to conceive of programming being anything else.
And one consequence of that is that we try to map every problem onto a call/return (incl. FP) solution, no matter how inappropriate the mapping, and the mapping very often is inappropriate, leading to architectural mismatch. [3][4]
We do this partly because we really don't know better, but also partly because there are tangible benefits, primarily that once we have done that mapping, we can then express whatever we arrived at pretty directly in our languages, because our languages effectively only allow call/return based abstraction.
But this is the problem with basically every functional programming article that makes it's way to HN. As a thought experiment it makes sense but it's not transferrable to the real world.
Like take the log out action as an example. Mid request we are going to have to change the state from a user logged in to a user logged out. Otherwise when we render the home screen, as an example, they'll see the logged in version instead of the anonymous user one.
And realistically we want need a transition state to show a successfully logged out message since we do not want that message to show the next time they come. So the whole sequence for this simple action is:
Old State -> Intermediate State -> Render Response -> New State for the next request.
So we have to change state a couple times and pass that along.
And then how do we deal with stuff that's actually state. Like a database or whatever. Are we really passing that as a function parameter so that 10 calls deep can use it?
And what happens 6 months in when we want a caching layer? Do we go back and edit every single function to accept a Redis argument now?
Realistically, no. We have some sort of config or server object that we pass along. So we have an object that contains all of our server state that gets passed to most functions. How is that better than a single global variable holding that state accessible everywhere in the application?
> ... as a thought experiment it makes sense but it's not transferrable to the real world.
For the record, it's certainly not a thought experiment, I have been using this in the real world for a few years.
> Like take the log out action as an example...
The problem you outline with regards to intermediate UI states around the logout flow example is one of assumed responsibilities. If an application needs to show some specific transient UI when the application has moved between two states well, that's a UI concern and does not have to be modelled as part of the core application state. For example if you have a UI for an application which is required to show some jazzy animation when a user transitions from logged-in to logged-out that in no way needs to be part of the core application logic. The UI can just inspect the incoming state and conditionally when detecting a move from LoggedIn to LoggedOut trigger some UI animation / transient dialog / whatever.
One nice thought experiment / thought pump I find of value when dividing core application (~domain) and UI responsibilities is thinking about if you were to switch the UI from say a mobile framework (say Android/Compose) to a desktop terminal, what stuff stays the same and what stuff is display specific. In the above example you most likely would not want to show a jazzy animation on the terminal and therefore is display specific and should no be modelled in the core application logic. This is a nice simple thought process for dividing these responsibilities.
> And then how do we deal with stuff that's actually state. Like a database or whatever. Are we really passing that as a function parameter so that 10 calls deep can use it?
No, a database would fall under IO and would make more sense to access behind a Command interface in this pattern, like most other side-effect interactions. This fits into the "functional core imperative shell" mindset referred to in the post.
> And what happens 6 months in when we want a caching layer? Do we go back and edit every single function to accept a Redis argument now? Realistically...
I am in no way experienced with BE caching layers, my experience is mostly client side mobile applications. However, my first approach for an in-memory DB cache would be to chuck that in front of the DB access behind the same abstraction, and also this would be behind the Command interface, again like all other side-effect interactions.
> This kind of specification doesn't need to specify `State` at all, as we can simply refer to previous inputs. E.g. imagine a counter system with the inputs `Increment`, `Reset`, and `Get`. The functional spec for `Get` is "find the latest reset and then count the number of `Increment`s since then".
I could be misunderstanding you, but this does not sound functional to me as `spec` seems to have access to previous inputs to `spec` invocations. What am I missing here?
> I think this is neat, because if you are developing the functional spec with a domain expert that doesn't know programming you don't wanna bother them with, for them, unnecessary bookkeeping details.
This sounds very similar to DDD (like as described as the DDD book at the end of the blog post).
> I learnt this from the Cleanroom software engineering people
At IBM direct, or are there any particular resources you can point to which you found useful?
> I could be misunderstanding you, but this does not sound functional to me as `spec` seems to have access to previous inputs to `spec` invocations. What am I missing here?
In a stateful system (where outputs depend on previous inputs) you need to have access to all previous inputs in order to determine the output of some input, right? Also keep in mind that this is a specification, not an efficient implementation. If you'd implement it like this you'd have to recompute the output based on the inputs over and over again for each new input. And yeah, you can think of it as some kind of functional composition as nine_k pointed out.
> This sounds very similar to DDD (like as described as the DDD book at the end of the blog post).
Yeah, I think DDD gets this right: some lightweight methods (e.g. event storming) around inputs and outputs. But yeah, Cleanroom did this in the 80s.
> At IBM direct, or are there any particular resources you can point to which you found useful?
