Better Software Design with Clean Architecture
fullstackmark.com
fullstackmark.com
After a couple of decades of experience, I've come to the conclusion that this isn't right. Most business rules involve processes, which are inherently procedural, or, from another perspective, functional - functions of the whole state of the system to a new state of the system. Most processes don't logically belong on an object, and when you stick them to one end, you create lots of problems for yourself.
Just take that example Student entity class; what stops you from writing:
student.RegisteredCourses.Add(course);
Invariants and validations for fields are trivial with an OO approach; for entities, they're reasonably easy, but sometimes you need partially invalid versions while editing or constructing an entity; but once you bring in multiple entities, and relations, everything starts getting hairy pretty quickly. Assertions and validations that would be trivial to write [1] in a relational language like SQL aren't possible in an object-oriented language without a lot of system-building.
So I've come around to the idea that the database is a better thing to put at the centre; that encapsulation and hiding of the main fact store is harmful to the architecture of a system, especially in a heterogeneous environment where your entities are represented in different languages, all backed by the same fact store.
[1] Trivial to write, but not necessarily cheap to evaluate. I'm not advocating writing all global validations in SQL.
Course - Aggregate Root
Student - Aggregate Root
RegisteredCourses[RegisteredCourse[]) - Sub Entity Collection - with id reference to course. With metadata like date time when the registration occurred.
RegisterForCouse Method
CourseRegistration.ViewCoursesFor(student);
I think giving either entity "ownership" of the relation is a disaster waiting to happen. I just have this vision of someone wanting data about students accidentally loading all the course information. In the DAL of course you end up modelling however but seeing courses in a student just makes me ill.
You normally need load an entire aggregate, so that you can maintain your invariants. If you separate into 3 aggregates such as Course, Student, CourseRegisteration then starts to become harder to maintain business rules. For example if you wanted the RegisterForCourse method to limit student course registration to 10, it would now have perform another query.
If start doing queries that don't match up your domain model, then should probably do CQRS. Which would allow you to build a model optimized for queries.
One day someone is going to want slightly different rules for student course registration and then they'll realise this rule is baked into the core and that's sad.
This is a desirable trait, and one of the main points of DDD. If you want to change registration, you change the core business logic. It then applies to all applications using that business logic. If you have a real business case for different registration methods, then you simply model that on your entities.
If you don't do this, your going to get business rules applied inconsistently as programmers will interpret requirements slightly differently.
So what's easier to change? Giving customers versions that give them all different types of the core base type Student or JUST changing the type of CourseRegistration that they visit? (CourseRegistration is a service as opposed to a currency).
You keep your currency CLEAN.
If you want something radically different, then are you even developing the same application anymore?
By the way i've never heard anyone call the core entities of an application "Currency" before, where did you even learn that? I imagine that could be confusing.
I also imagine having an application where you have build custom methods for each customer who wants something slightly different is terrible for codebase maintainability.
I make a career out coming to companies that should have (according to you) made radically different systems but didn't. Most companies don't find it economically justifiable to re-write based on slightly different customer requirements and most companies have many deployments, not just "one perfect one" that most modellers appear to imagine. If your design doesn't have the flexibility of change designed into it you're going to struggle when you get a second client.
> I also imagine having an application where you have build custom methods for each customer who wants something slightly different is terrible for codebase maintainability.
Yeah of course its a massive pain in the ass but it enables you to sell more.
Then don't. Make the domain model customisable with various options.
Having specific methods for specific companies seems a lot worse.
I build SAAS applications which have various customization options customers can do. I don't really have a problem with it. I've designed a clean way to customise.
I made the same mistakes as OP and learnt the hard way. I encourage you to explore a greater purity in your models in the future because I genuinely believe it leads to better code.
It's always a good idea in these situations to appeal to real life. The actual business will point the way of the business simulation. In the real world we don't ask students to register for a course. Instead students ask the Registrar to register for a course. When the Registrar makes her decision she considers far more than just the internal state of the Student; she considers (1) does the course have any available seats? (2) has the Student met all the course pre-requisites and, most importantly, (3) has the student paid his tuition and is he even a valid member of the university community.
What the Student does have is a history and a context -- that is, a state -- that must be considered when registering for courses. The student may also have preferences -- courses he wants to register for.
The language of the business should guide these decisions always. A student submits a request tfor a course and it is the office of that accepts or denies this request.
Global.do_some_thing(foo, bar)
do_some_thing(foo, bar)
- For tests at a smaller level, decompose the application such that most parts are easily testable in isolation (without external "hard" dependencies).
- For mid-sized tests with external dependencies but mostly unidirectional dataflow, setup a global virtual table with all the mock methods (and instances) that are needed. Alternatively, traditional linking methods.
- For larger "integration" tests there is no substitute to testing the real thing. At some scale and level of interactivity with the database, you just have to talk to the true filesystem, the true database etc. You can still setup a test instance for most cases where there is no interaction with external services.
In the former, access hiding cleans up the API and prevents unsafe usages of the container. In the latter, OO enforces a protocol to keep the state machine sealed off and only aware of key inputs and outputs.
And that's it. I've found nothing else. Data itself is better off when strategized to fit in a database, whether off-the-shelf or a custom-tuned, in-memory design. The state machines may need to query a part or all of the database, as well, so their ability to restrict scope only goes so far.
When #1 is missed I usually see that theres a design that doesn't mimic its domain and thus lose the ability for developers and users to have clear, concise communication. At that point OO is a disservice.
When #2 is missed you end up with IFilteredCourseAdapterProcessor as people attempt to bolt on components to solve future needs.
The addition of the "Registrar" to the domain immediately demonstrates how the naive interpretation is missing core components and I bet the users and devs fundamentally aren't speaking the same language.
This, imo, leads to conversations like, "Of course so and so approves all the registrations! Otherwise it would be madness!"
I use some OO to make abstract datatypes in languages that are inherently OO, but I think the explicit virtual table approach in C or the type classes approach in Haskell are much cleaner.
Lately I had to make a REST API which is basically a distributed Object-oriented interface. I think I managed to get it done with compromises, but I'm not happy. Another idea would be to make a procedural interface first and then make a REST API on top if needed. But I have some doubts it can work out practically.
An API is there because someone in the other end wants to do something. If you don't design it but just slap REST on top of your data, then you're not doing that person a favour.
Actually, after having tried a few times, I'm pretty sure I don't want to put REST ideas at the center of my architecture. You just happen to need a network transport, and not even in all cases (debugging for example). And an existing model is never going to be able to represent all the concepts of your application domain. This means that you need to build your own representations.
In theory there is a point of using a standardized object protocol which can represent some CRUD use cases. But from my limited experience I think it breaks pretty quickly, to the point where I can use only GET and POST.
Using protocol layers of increasing specificity may make it easier to use existing tooling with data exchanges. For example, many APIs use HTTP statuscodes as a more coarse-grained version of the return codes in the body (in an ad-hoc format). Also caching is often brought up as an example where you want to buy in to a standardized protocol.
But some APIs don't buy in, like Facebook, which reportedly returns 200 OK always. It seems like a lot of best-effort work with little returns to me as well. But I don't know - I'm not a professional business software guy.
You could defiantly model it as the Register method on the course. But this example shows how you can avoid people messing around with internals once you've modeled it. There is only one way of registering for a course.
In other words, I don't see how it is a domain question, and it seems likely that the domain practitioners just think "there are courses, and there are students, and students can be registered for courses".
My suspicion is that DDD would basically be better without the focus on single dispatch OO.
school.register(student, course)
I really think single dispatch OO is a huge distraction and I haven't seen any strong arguments for its use as a general paradigm.
That's why I like Django. You just tell it that you want a many-to-many between students and courses and it does they for you, symmetrically, without forcing you to introduce a class such as "school".
If your design is some data, which you can mutate however you want to get the job done, you don't really have model of the application.
Your business rules can become inconsistent because different developers are implementing different but similar methods in your business logic layer all applying slightly different rules. If the entities themselves enforce these rules, it become a lot less likely.
Practically speaking, I find that abstracting a system into a set of commands and a set of queries (i.e. CQRS) is often the "right" solution. Each command/query encapsulates an individual use-case, and all of the business rules and database access required.
Command side has a domain model, which would be your business objects in this case. A command simply loads the entity, performs a method on it, saves it.
Query has a query model which designed for fast reads, normally built by events emitted by the domain model.
Where systems breakdown is when other viewpoints get added. For this system it might be the following:
- Pricing/Invoicing
- Prerequisites
- Academic Status
- Professor assignment
- Course Reviews / Social Media
The problem is that at a high level two viewpoints look 95% similar, but in the code it equates to creating interdependencies between all of the viewpoints.
Different teams might be successful in keeping the separation in the system, but in most systems I've seen the entanglement starts in the database with the entities. Columns that become nothing more than status flags for different viewpoints. Columns with near identical names that mean almost the same thing, but are handled differently because the viewpoints treat them differently.
When the system gets big enough, a developer cannot mentally map the whole thing. When implementing a feature they will look for what is available vs what the architect had in mind.
This is how you wind up with 10 different getCourse calls all of which are building off one another with various parameters. The code will have a lot of if/thens checking the parameters to make it work for a particular viewpoint and avoids bugs for the others.
The more separation you have between the viewpoints the better. I now prefer separate entities/databases for every viewpoint. There is a set of entities common to all, but these are fact level entities. A course, A student, A professor, An admin, A TA. The entities should contain no status.
Where the viewpoints need information from each other they should just query the appropriate viewpoint, or have the source viewpoint send out updates (great place for event sourcing).
It might sound like I'm describing micro services. I wouldn't argue that, but I would say that a viewpoint is a higher level concept than a service. A viewpoint could be a collection of micro-services, or a single system (using this Clean Architecture).
Sandi Metz did a pretty good talk about this, basically saying we need to think more before we abstract things away because "code duplication". https://www.sandimetz.com/blog/2016/1/20/the-wrong-abstracti...
I've seen systems that take this to the extreme of not allowing a single piece of logic into the domain. Domain objects were essentially data containers only. So, if you had a Person object with firstName and lastName properties that represented those DB columns, then even a getFullName() that concatenated the two was verboten.
Instead, all logic had to be in a service. This led to lots of duplication and a super-massive service layer in a system that was decidedly procedural, even if it was implemented in an OO-language.
In my admittedly limited experience, applications come and go, but data lasts forever. So I find it unusual when a database is not at the heart of business software.
When you want to ask questions about the state of the app, you make queries. These queries could be sql - in which case, your app is directly dependent on the type of storage, and is almost un-abstractable. it could be a custom/bespoke query language (where a set of hardcoded api calls to the Db/datastore counts as an api).
The event source system is responsible only for storing facts. Therefore, the "problem" of where the validation of students against courses doesn't exist, because that relationship is a "fact" in the event store (there must've been a registration happening at some point for this fact to exist). Therefore, a programmer _cannot_ make the mistake of accidentally adding a course to a student who didn't register, unless they maliciously do it.
Then apply a command to the domain object, then save the emitted events.
They are still entity centric.
I've come around to the idea that the database
is a better thing to put at the centre
I'm not advocating writing all global validations
in SQL.
Turns out that the answer is a data access component might know where to find a database, and a client app might know where to find a web service. These things were rather obvious to me, so I failed by trying to find loftier answers. Which probably says a lot of about how I think.
Also, the concentric circles did nothing for my way of thinking, as messages are linear - they start from a place, and they end in a place. Thus a stack was a lot easier for me to digest than these concentric circles.
And finally, the ambiguous language. Why should I care that there are enterprise business rules, and application business rules? They're both rules. Rules go into a component in the middle, or business layer.
And thus I found it a lot more useful to think of app design in terms of vertical tiers (hardware abstractions), and layers (software abstractions). Tiers include perimeter, DMZ and corpnet, and layers include (but not limited to) user interface, façade/service gateway, business layer, and data layer. With some cross-cutting concerns like comms, security and operational management (logging, exceptions and so on).
I don't think either is better than the other. It was, however, the first time I really understood that two people can be very knowledgeable about the same thing, and yet speak using a completely different vocabulary.
Messages might be linear but scopes are encompassing inner scopes. Encapsulation is not usually depicted as a stack.
>And finally, the ambiguous language. Why should I care that there are enterprise business rules, and application business rules? They're both rules.
Obviously because different scopes apply to the former than to the latter. TFA even clarifies that: "The key is that they ("enterprise business rules") contain rules that are not application specific - so basically any global or shareable logic that could be reused in other applications should be encapsulated in an entity".
In general, the similarities ("they're both rules") between two things don't say much (if anything at all) without considering the differences. Shiitake and Amanita Muscaria are "just mushrooms", but one can kill you.
It can be, as concentric circles and stacks are isomorphic. What can be represented by one can be represented by other. Examples of dealing with scopes as stacks would probably be familiar to anyone working with C, C++ and other languages with stack-based local variables - your scopes there literally correspond to what's on a stack. Similarly, lexical scoping can be seen as a sequence (i.e. a stack) of associative containers.
(Hell, you can see concentric circless as a Tower of Hanoi viewed from top - i.e. a stack (albeit inverted in this case - I'd put the innermost circle at the bottom of the stack).)
I understand GP's pain because I too would prefer the same diagram as a simple stack of abstraction layers. But then again, such diagrams almost never work without corresponding textual explanation anyways, so it doesn't matter much how the diagram looks as long as the companion text is OK :).
Well, the Wikipedia lemma puts it this way:
Although classified as poisonous, reports of human deaths
resulting from its ingestion are extremely rare. A fatal
dose has been calculated as 15 caps.[55] Deaths from this
fungus A. muscaria have been reported in historical journal
articles and newspaper reports,[56][57][58] but with modern
medical treatment, fatal poisoning from ingesting this
mushroom is extremely rare.[59] Many older books list
Amanita muscaria as "deadly", but this is an error that
implies the mushroom is more toxic than it is.[60] The
North American Mycological Association has stated that
there were: "no reliably documented cases of death from
toxins in these mushrooms in the past 100 years".When are people going to learn that if you just start with the entry point with granular components, letting each define the interface for its dependencies, you get a much nicer, looser, more flexible structure than these enterprise "patterns" that ultimately all just turn into a big ball of mud?
Stop pretending you can design codebases.
My code is typically some dumb data objects, static functions to apply rules to that data, and a bunch of interfaces for abstracting external dependencies. The code that ties everything together lives at the entry points, and is hopefully written like one is reading a requirement.
To take the course registration example, the "register" entry point might look something like:
register(studentID, courseID):
// Load student from data store interface.
Student student = this.datastore.getStudent(studentID)
// Business rules for whether a student can still register for classes.
// This would check things like course overload and duplicate registrations.
if (!Students.canRegister(student, courseID)):
return CannotRegisterError
// Record registration.
this.datastore.addRegistration(studentID, courseID)
* Data objects are dumb. No need to even test them.
* Rules are all static functions, so you can call them anywhere, anytime, including for testing. Extremely flexible and allows easy remixing of rules for new requirements.
* All external dependencies live behind an interface, so they can all be mocked away for testing.
IMO, this general idea is very powerful for most code. I believe algebraic effect systems will popularize it further by making it a natural pattern. While OO should be written in this way, we are currently in this weird period of programming history where we don't talk about practices and design as seriously as we should.
But, I'm glad you mentioned this approach. It is not hard to understand, it is elegant, and it is as simple as can be. Just requires a tiny amount of glue and forethought.
* Architecture / design pattern lasagna, where breaking through abstraction layers felt like Inception. And adding a data field meant going through all 5 layers of code...
* Inheritance for functionality instead of typing.
* Also, breaking Liskov substitution willy nilly then wondering why things are hard to reason about. (Because your code becomes entirely reliant on the types actually present at runtime, since you can't rely on semantic equivalence.)
* Tacking methods into data objects because that's where the data is. See other comments on this post advocating a "register" function on the Student object, or should it be on the Course class... Answer is neither. Encapsulation is about maintaining invariants. It's not about putting all functions related to students in the Student class.
* Zero unit tests because external dependencies weren't isolated and all the logic is hard coded into exact use cases, which eventually end in hard coded external dependencies...
There can be a lot of code repetition. To me, that's not a bad thing. I've seen people to themselves in knots trying to DRY some common code, just to later have to undo it all because a requirement changed for only one of the code paths. And, as I said, I like my procedures to read like use cases.
I'll see if I can think of anything else...
In my case that looks like: Write code for 6h, review and refactor code 2h. Result is that less code is produced (due to refactoring/removing/etc.) and codebase keeps being simple. On the other hand it's easier to write tests.
No need to use complex enterprise patterns. Most of time simple facades and delegators are enough. Consider writing small simple components instead of using heavy patterns with a lot of boilerplate code.
So, yes, each granular component should have its own perspective of "the Enterprise". But, contexts can be very different. For a simple example, a "car" may have completely different interface depending on if its user is a "car designer", "car assembler", "car salesperson", "car driver", etc. So, someone trying to define a universal interface for a car will make themselves and their teams crazy. Worse, the idea of "a car" will change through time, so even perfect interfaces will have to change.
This isn't limited to object oriented design, by the way. Functional programming paradigms have the same issues.
What you describe, at least in my experience, more often than not leads to a big ball of mud as easily as the methodology in the article.
Software can be designed. It's just there's little appetite for an actual rigorous design process in this industry. Instead there is a lot of bandwagoning and looking for one size fits all solutions (like the article's), mixed thoroughly with people who haven't ever really grown beyond thinking of textbook CS as the solution to every problem.
I get that they're useful for taming a certain amount and kind of complexity, but it's not clear to me that it's always going to be apparent at the beginning that it's going to need taming in that specific way. I've found that starting with the concrete cases, I only sometimes have to go up a level of abstraction and indirection. Conversely, I've built the wrong abstraction many times by starting too high.
I don't read or write Java so I'm sure I'm missing a lot of context. That's what I'm looking for. Uncle Bob's an eloquent speaker and his talks make a lot of sense, but I have trouble reconciling that with the code samples I see.
Then a few years ago I had to maintain a software stack written by contractors from Pivotal (https://pivotal.io/) in a Clean Architecture style - it was truly a revelation for me; akin to that "aha" moment of fully grokking homoiconicity in LISPs.
Now, up until this point, all code I'd encountered at Twitch heavily reflected the domain of the problem it was solving and the technology in which it was written. That is, if you were looking at a certain piece of architecture you had to really fully understand not only exactly the intent of that code base but also the details of the framework in which it was written. As codebases increased in number and size, it became much harder to scale as an engineer. Jumping from a Rails monolith, to a highly concurrent Golang HTTP CRUD-API, to a highly asynchronous Twisted-Python request routing system (broadly the main three backend techs) had a very large cognitive load. The eng org cleaved along these lines and maintaining velocity in that world was very hard; attempts to introduce new tech or join chunks of the org took on a "religious" tone.
So initially coming across this clean architecture stack felt very similar to that. It had a lot of "weird new things" in it, but once I understood that much of it was routing (tagging inbound requests in a manner that a deeper layer could understand the intent of the request and pass it to the correct interactor, which would then work on the appropriate entities) it suddenly became incredibly easy to hop around the code base and update the important aspects of it.
I asked the authors who had been contracted to build this system where they got their inspiration and they cited many lunchtime discussions and pair programming sessions influenced heavily by Uncle Bob's clean architecture.
I would have really enjoyed seeing more systems built like this because, to the maintenance programmer, it was very clear where things had to go. However only encountering one Clean Architected system didn't really give me a solid idea of how well it would scale across various domains.
There's a lot of engineers at Pivotal, and we have a lot of projects, so maybe I can prove I worked on that codebase by referencing obscure details from it? I remember writing a LOT OF TESTS that used quotes from Wutang Clan and 36 Chambers as strings for names of entities, and a lot of references to 90s hip hop. Was this the same project?
Also, I took some time after our project (and several others) to start a small example of applying Clean Architecture in a toy Go Project, which is open sourced. Maybe check it out?
This can be a kind of organizational solution much like microservices to separate concerns. This is important when collaborating with a large amount of people. Clear boundaries and all that.
You pay for those boundaries with a convoluted mess of classes that describe the design pattern rather than the business logic. You end up with AbstractRequestInterfaceFactory type stuff.
Trade offs. Decide when they're worth it.
Clean Architecture™ has been around for a while now; I wouldn't be surprised if it was more than a decade already. This post is a guide to an existing concept, not a presentation of something new.
EDIT: This is Uncle Bob, 5 years ago: https://8thlight.com/blog/uncle-bob/2012/08/13/the-clean-arc.... So maybe it's just 5 years old - but then there were similar things (like Hexagonal Architecture) before.
I disagree. If a design (call it an architecture) is clean and coherent, it's obvious.
Because obviously a client must send the order to a server, which is protected by a gateway. And of course that gatway validates the client request, and the data coming in. And of course the server must have business rules to validate and process the order. And clearly the business components must call a data access component to persist the orders to a database. Simple, no?
what happens when the business rule clashes with the database's constraint implementation, because different people either misinterpreted, or due to incompetence or miscommunication?
I think splitting up sounds great in theory, but in practise, has lots of pitfalls. That's not to say it's not a good idea, but one musn't look at it with rose-tinted glasses.
Either way this is just part of building a system. Refector, re-test and release (or re-release). It's hardly going to be the only bug you find.
I think it's hard to really explain why software should be built like this. This is why architects and senior devs and the like often end up ruling by dictat and simply laying down rules like "messages only in the ACL, no messaging in the domain." The problem that Clean Architecture and DDD are trying to solve are architectural and ultimately organizational problems and these problems are not clear to most developers who are just given a story and told to implement some new feature or story. The only hope is to make these these architectural problems everybody's problem. Spread the pain.
Microservices btw are no escape hatch. Microservices that don't get the essential system boundaries right are in fact going to become a horrible mess. If anything microservices make it more important to think carefully about the boundaries and the interfaces of the system.
Because good architecture has somehow evolved?
>This can be a kind of organizational solution much like microservices to separate concerns.
This is orthogonal to microservices. And microservices aint but one model of creating applications, not "THE 2017 model".
Of course. Every year gives us greater insight into what approaches work and don't work. Which is why we have seen Java technologies like EJB, JCA, JTA, OSGI, JSP etc generally fall by the wayside in favour of less convoluted and less architecturally heavy approaches.
> And microservices aint but one model of creating applications, not "THE 2017 model".
Microservices are unquestionably the defacto standard for most larger software development projects. Especially since they are intrinsically linked to the containerisation deployment approach which is a modern day concept.
There's a wide world of software out there that you have yet to explore.
In the HN echo chamber maybe -- but not sure even there.
Can't remember ever seeing them in plain or modern Java EE code.
Scott Wlaschin has recently released an [ebook]https://pragprog.com/book/swdddf/domain-modeling-made-functi... on the topic which goes into more detail.
Overall architecture in development is still early on in maturity.
There should not be a lot of room for creativity when implementing specific business rules. The goal should be clarity and readability.
These convoluted patterns obscure the logic and confuse the reader with their pointless abstractions.
"Enforcing separation of concerns" is a good one. But so would be "identify the various runtime threads of your system easily", or "minimize code surface responsible for mutation of shared data", or "decouple configuration primitives from the components themselves".
Those concerns are more or less important depending on your use case. MMORPG server code or regular desktop app, or mobile webapps sold as templates, all have very very different concerns, so i dont think it would make sense to use a single architecture as a template for every problem.
[1] https://en.wikipedia.org/wiki/THE_multiprogramming_system
This is a massive fuck up and it won't scale or be modular. The idea is fine but you need the entities/currency into a smaller core with your DAL and relationship objects above that layer. The only functions your currency should have are accessors or read-only convenience calculations. Other operations should be handled by another parent because let me tell you that that Course collection is going to end up being null or empty A LOT when it "shouldn't" be.
You'll want your use cases to return shallow object graphs (e.g. GetCourses => courses.Enrolled => StudentName, StudentId) from the data layer/services and then if you want to look up a Students details you make that another call. The alternative is to run some sort of "scope" object that defines the depth of graph you want when accessing a general api. The thing you're looking to avoid though is returning more information than is necessary.
Also who taught this guy to model? Students are in courses in this model the courses are inside students and that's silly.
If it is cheap to pull the additional data and it makes sense to expect that data in using the object, I'd simply embed the representation. If it was expensive or not commonly used, I'd include a reference to the data (usually ID(s) of the data).
Then in the repository, I'd just get everything necessary to return a full object as designated.
An onion architecture gets funky if you don't know if the object you are working with has all its data, especially because the object itself shouldn't know how to fetch more data about itself.
In this particular case, the two objects probably shouldn't have ANY domain connection, and there should just be a method on the course repository to "GetCoursesForStudent" that accepts a student ID. It's perfectly ok to model relationships in the database that don't have parallel relationships in the domain objects.
Or, one could create a "SemesterEnrollment" object that contains a student object and a collection of courses. Which would probably make more sense, as that's the object that should be referenced to generate bills, report cards, etc.
Came across this collection of articles and talks, for anyone interested: https://gist.github.com/kbilsted/abdc017858cad68c3e7926b0364...
I became familiar with the names of people like Booch, Rumbaugh, Jacobson and others. Over the years I learned various object oriented programming languages and patterns. Each new thing was like discovering a horcrux, at first magical and powerful, but ultimately evil. Later I began to learn functional programming and that's what I try to use most these days but it too has its promises and lies. I have built and helped others reason about many many complex systems and all I can say is this:
The only system that is well ordered internally is the one that accreted complexity in increments, and was continually refactored along the way. Best practices be damned.
Those systems might not look how you'd design them were you Uncle Bob. They work, can be understood, and can be modified without much consternation. We all can recite examples of masterpiece turned morass. Conversely some of us can recite an example of a frog, a weird complex beast but ultimately very well adapted to its environment and quite resilient. Frogs are not beautiful but great at eating bugs.
One fact of complex systems is: humans cannot know the "right" design until AFTER he has arrived at it.
If a human can know the design a priori then the problem is NOT COMPLEX and if it is not complex then it is not modern software.
This is humbling knowledge for someone who wants to believe there can be a language and pattern of order for all systems. One that can be expressed in anything less precise than code itself.
Given a fanciful machine that could assemble subatomic particles in any fashion the user desires none of us could, having never seen one, design a frog.
Basically, for me, it boils down to the following:
1. Create domain package(s) with the basic business entities (users, vendors, products, etc) - can't import anything. 2. Create usecases package(s) that acts on these objects and can import anything in the domain package, but nothing else (though can accept repository store interfaces). 3. Create infrastructure packages for taking to databases, caches, etc - can't import from rest of application. 4. Build repository stores to translate between repositories and domain objects (can import domain and usecases, and accepts interfaces for infrastructure). 5. Add controllers/main packages that build repository dependencies with infrastructure config, and pass those interfaces into usecases that do the necessary work and spot back results (imports everything).
The end result is a bit topsy turvy to get used to at first, but then it's a dream. Extremely easy to test (very little dependency coupling in each package), and the most complex parts of the application logic are totally isolated from the complex parts of the infrastructure, so you end up with less cognitive load when dealing with either.
After my experiences so far, I can't see ever switching back to "top down" architecture instead of "inner out" when given a choice.
I've seen enormously complex domains implemented as DTOs that are acted duplicately in ORM queries, in PDF rendering, in CSV exporting and then again in CSV importing.
It looks like this only deals with in-memory data structures. How does stuff get read from / written to the DB?
You would have some kind adapter/library that handles the persistence.
It would take these models, and covert them into your datastore.
RegisteredCourses.Add(course);
Course - Aggregate Root
Student - Aggregate Root
RegisteredCourse[] - RegisteredCourses - An array objects with date time of registration, and id reference to course.
_studentRepository.save(student) would be a simple matter converting this in memory representation into sql.
I'm not big fan of using lazy loading, and direct references to other aggregates inside of other aggregates. This makes mapping these models without ORM difficult.
I find it helps a lot to think carefully about (1) message validation -- is this a valid message? (2) entity validation -- is this entity in a valid state? and (3) enterprise validation -- is the business system as a whole now in a valid state? These three questions map on to what is often an ACL, domain, and domain services layer. At the end of the day there's going to be corner cases though in which case as a rule of thumb there's something to be said for doing validation at the edges and moving it in closing to the center as needed. In my experience when it's not clear where validation belongs that often means nobody really understands that validation (or that validation is even wrong and is not what the business wants -- it happens!) and so there's something to be said for keeping it out of the core domain.
That's a good approach unless a validation requires a db call with inner join (for example). This becomes too costly to do it 5 times (for example) just to get something written into db.
A interface to define the validation aspect. 3 separate implementations of said interface : 1 for db lookup, 1 for cache lookup and last for unit tests. 1 factory method which return the instance to use depending of context. Another temporal cache that holds the instance of said interface (after all, you can never have enough caches).
And all of a sudden, validating some single shit takes +300 lines of code (plus tests). Welcome to enterprise , clean, better designed, greatly arhitected software everyone....
Range acceptance - could defiantly be a business rule. Cross-field consistency - Could also be a business rule.
I implement these rules both sides. Ideally i don't want an invalid command sent off in the first place. I also don't want badly implemented UI corrupting my business data.
This way i can provide instant feedback to the user, and also protect the data in case the UI is badly implemented.
The main point of these architectures is that you can use them from many user interfaces.
Adhering to dry in all cases, is not always the best solution. As the software community has discovered over the past years. Especially in relation to microservices.
In my software there's a bit of a disconnect between the application and interactors. As there is a message queue between them.
Business rules does not mean what the business (bosses) dictates that the app should have.
Business rules the rules for the business domain.
Whether you should escape sql (or even use sql) is a technical/application concern, not a business rule.
A business rule would be something like "customers must get 10% discount if they bought more than 100 units" or "no employee should be allowed to have a salary bigger than their manager", etc.
Business rule validation(E.g student can't be assigned more than 10 courses) is done on the entities themselves.
You put it on both sides to get the best of both. Instant feedback to the user, and on entities as last resort protection against a badly implemented interface.
In two places in UI: as close to the user as possible (which improves ergonomics) and at the system's border (which prevents entering invalid data to the system at all). While the former is somewhat optional, the latter is absolutely necessary and cannot be left to client-side JavaScript.
But yeah, regardless of the lame appropriation, a solid concept.