Hackle's blog
between the abstractions we want and the abstractions we get.

Architecture? You mean spreadsheets?

3-layer architecture. 4-layer architecture. Onion architecture. Domain-driven architecture. Ports and adapters. Hexagonal architecture. etc. etc.

Software engineering is so diversified, excitedly they say, look at the styles of architecture for us to choose from!

Yes, they are indeed many, like flavours of cucumbers, although not nearly as tasteful. Also, maybe they are not that different - they are more like different schema designs, you know, as in relational databases, or really, like spreadsheets.

the expression problem

One of the best explanation for me about the "expression problem" is this by Calmarius using tables, you know, plain columns and rows. The gist is, it's easy to add data (rows), but not so changing schema (columns). Designing types is like designing schemas, they are essential activities in (typed) programming, as we have learned in the world of relational databases.

If types can be treated as data, then certainly code should be too? After all it's been a trendy thing to say, "code is data"!

So here we are again, looking at the perennial discussion of file structure - or roughly application architecture in a code-base. Given the following files, how should we organise them?

Never mind the naming convention, it's not what I would really have, but hopefully they are representative enough. and never mind the ordering for now although it does reveal biases.

First we observe each file has two aspects: the technical aspect - controller, model, service, repository, interface, implementation; and the domain aspect - payment, member(ship). That's two-dimensional, perfect for a table. Let's try to put them into one.

Of course, now we have to design a schema - which should the columns be? Let's pick the "features" first.

A. domain aspects as columns

Rigidity is an intended quality of schemas design; the point is to restrict and guide placement of data. In our case, it's placement of files. So, by making two columns "Payment" and "Member", this "architecture" prescribes all the folders must serve these two purposes ("features"), as they make up the entire domain.

Folder Payment Member
Controller PaymentController MemberController
Service Interface IPaymentService IMemberService
Service PaymentService MemberService
Repository Interface IPaymentRepository IMemberRepository
Repository PaymentRepository MemberRepository
Model PaymentModel MemberModel

You may be surprised, but this could be by far the most popular way to organise source code. Don't believe me? Check if your code-base has any of these folders,

If any of these folders exists, then it's a "yes", more of less. Of course it may not always be verbatim. We also call this style by different names: "n-layer architecture", or analogies like "service oriented", or "onion architecture".

There is a big advantage this to this structure: If we run a query: SELECT * FROM files WHERE Item = 'Interface', we get all the interfaces - imagine this can be quite handy for binding interfaces to implementations for a dependency injector.

Its disadvantages should have been made clear too: it's quite hard to add a new feature, say "Search"; we must backfill every single row with the respective SearchController, ISearchService, SearchService, etc.

Or, to find everything about "Payment", the query is SELECT Payment FROM files - this traverses the whole table!

If we put the cost of change in big O notations, all folder must change with any feature change; in complexity, this "architecture" style is of O(n) complexity.

One may argue it's O(1) to add a new technical aspect, for example, a new *ModelValidator for each feature. This may be true, but how often does that happen in a solution in production?

B. technical concerns as columns

What if we transpose the table, so the technical aspects are columns?

Folder Controller Service Interface Service IRepository Repository Model
Payment PaymentController IPaymentService PaymentService IPaymentRepository PaymentRepository PaymentModel
Member MemberController IMemberService MemberService IMemberRepository MemberRepository MemberModel

Now run SELECT * FROM files WHERE item='Payment', and we should get everything related to the story of "Payment".

If we want to add the "Search" feature, it's a matter of inserting another row. UPDATE or DELETE any feature should be equally straight-forward.

However, it's not so easy to get all the Controllers. Again, we can create a Controller_Index. This should be an acceptable cost; and we possibly won't be creating an index for every column.

In terms of complexity, adding or changing a feature is O(1).

In colloquial terms, this is what a "domain driven architecture" should look like.

Schema-less

But what if a new technical concern, like *ModelValidator is required? This again requires a change across the table (pun intended), as should be expected.

But it doesn't have to be so - this schema leads to very flat, lean and self-sufficient folders; as the folders are so autonomous and independent of each other, a form of freedom should grow; given time, programmers will find the constraints of the columns unnecessary, so they are removed, and we arrive at a schema-less, or "NoSQL" style - each folder contains whatever necessary to make up a feature; folders don't really share the same internal structure.

This may sound very liberating and appealing, and I personally think we should learn towards this style more, I believe it's worth noting there are times constraints and rigidity are equally, if not more valuable.

A combination

While I am one to be biased towards the latter (or schema-less) style, many from the industry had realised option A is pretty silly, so there have been methods of mitigation, for example, let's not have "Service Interface" and "Repository Interface"; they should be put in the same folder which is called "interfaces". Or, interfaces should live together with models, optionally services, and let's call them "core" or "domain".

If this sounded like a good idea, it's not - just look at the columns and rows in the table.

Folder Payment Member
Controller PaymentController MemberController
Domain
- Domain/Models PaymentModel MemberModel
- Domain/Interface/Service IPaymentService IMemberService
- Domain/Interface/Repository IPaymentRepository IMemberRepository
- Domain/Implementation/Service PaymentService MemberService
Repository PaymentRepository MemberRepository

Combining "layered" and "domain-driven" architecture gives rise to a confusing and awkward file structure: for some code (Controller or Repository), there are dedicated folders; for other code, I have to drill down a rabbit hole of folders organised around technical concerns.

SQL best practices offer the best explanation - this schema design breaks normalisation, and creates a hidden, nested sub-schema. Of course this is a bad idea - what if I want to further specialise PaymentRepository, for example, to have two strategies: one over Http and another over Postgres?

Folder Payment Member
Repository
- Repository/Postgres PaymentPgRepository MemberPgRepository
- Repository/Http PaymentHttpRepository

That's normalisation broken into pieces.

Another favourite example of mine is when all unit tests are put in a separate "test" folder (or package/namespace). This folder and its sub-folders will now reflect the file structure of the code under test, and become a duplication that will be a pain to keep in sync with its originals.

Question

This is the time I raise the question: in the essence, is application architecture the same problem as the expression problem?