Kraken Technologies: How we organise our large Python monolith
blog.europython.eu
blog.europython.eu
I wonder if this might be susceptible to the same issues that we had in java, specifically it doesn’t prevent accidental duplication, it makes refactoring harder/slower and it’s problematic for the team when it unexpectedly doesn’t catch a misuse - and this happens because the rules enforcement can only happen effectively if the rules themselves are perfect and complete and that gets harder to guarantee with scale and speed of iteration.
For my money today, there’s another model “polylith” which is similar in some ways and as easy to understand but maybe simpler to evolve/change your mind over time.
IMHO, project-level assertions about initialization-order feels a little bit dirty, but potentially practical as long as there's only a few rules. Surely even a bazillion line monolith can still often be thought of in terms of a few layers. But by the time you get to lots of rules it's getting hard to stomach the code-smell and keep justifying it as practical. At that point I'd want something dynamic and backed by a real solver.. not just a layer of static-analysis.
The deal framework for DBC doesn't quite do this afaik but is starting to expose some stuff along these lines, and at least can assert imports are "pure". https://deal.readthedocs.io/details/module_load.html
Really looking forward to trying verification ( https://deal.readthedocs.io/basic/verification.html ) but I think I need to hit some critical mass of annotation first.
Been learning some of the idioms of a polylith architecture using the `poly` tool + clojure. Really enjoying the idioms/practices it encourages.
sometimes sections of code can be syntactically duplicate at some point in time but aren't really semantically duplicate.
e.g. in terms of the example in the post, of having different client-specific logic. at times two different clients may end up with syntactically duplicate code in their corresponding client-specific layers. but unless this code reflects that some detail specific to both clients is actually subject to the same pressures and constraints, what is syntactically duplicate today may not be tomorrow when client A's code needs to change due to pressure P that does not also impact client B, so the syntactically duplicate code is no longer syntactically duplicate.
factoring out syntactically duplicate common code between things that need to change at different times for different reasons is unhelpful as it introduces unnecessary coupling.
not sure if i've understood the context of "prevent accidental duplication" but reasons for change and avoiding unnecessarily coupling is something that jumps into mind when considering multiple clients or multiple regions
Every client has to have duplicate modules for every territory that they operate in and if they begin to operate in another territory, that has to be copied in — or they have to keep a copy of all territory modules for every client. I think it is probably the former, since they talk about pruning and ignored imports.
I even have a lua codebase where I've linearized the sourcecode files into a particular order. The layout looks like this:
01_foo/
03_foo.lua
04_bar.lua
15_whatever.lua
02_bar/
22_whatever.lua
56_something.lua
So, my approach is to impose this linear order and only allow code that comes later to depend on code that comes earlier, never the other way around. This way, you can read through the code rather easily. As you look at any piece of code, it will only depend on code you've already looked at, so you don't have to be constantly jumping through the codebase as you're reading.
I've also found that failure to correspond to a linearizable dependency structure is a "code smell", and that, as I try to eradicate that code smell, I frequently end up with code that's better in all kinds of ways.
First, who has time to read a project from start to beginning. Second, as the time changes the dependencies between parts will change as well.
A well-written piece of code that actually solves a complex problem is a demanding mistress and doesn't appreciate the "who has time..." attitude.
Among the software engineering good practices that I engage in, I'm finding this to be one of the highest bang-for-the-buck ones.
For example, literate programming is a thing that many people believe is worth the effort, at least in certain scenarios and at least for certain parts of codebases: But, if you can make your codebase more readable without requiring additional prose to be written, then that's a comparatively high bang-for-the-buck alternative.
In practice, linear structure doesn't need all that much updating at all: You may have noticed that I don't use strictly sequential numbers. Early in the life of the codebase, I will generally skip lots of numbers, so that I can fill in the gaps with code that comes later. Mostly, all it really takes is to be intentional about where in the linear order to insert new code.
Also, note that it's still useful, even if no one ever goes on a mission to read the codebase in its entirety. If the codebase as a whole has a linear structure, that also imposes a linear structure on any subset of it. If you have five sourcecode files that are somehow pertinent to something you're doing, you immediately know the order in which to look at them.
I have 99 big problems developing software, and not being able to read it from start to end ain't one.
I'm a bit jealous for people advocating such things, being able to work on software tiny enough to even be able to think this is practical. Literate programming could work for tiny software that can be approximately flattened to a 1d "story": "beginning -> end" with only minor detours.
Most of projects I worked with was large enough that things are optimized precisely for not having to understand large parts that were abstracted away (and the problem is to try to do it without introducing too much complexity).
By "real" complexity, I mean to exclude complexity for complexity's sake, or complexity that's a side effect of a piece of software being concerned with itself and its own structure rather than with solving any problem that exists outside of it. -- We can surely all agree that the latter is bad.
Academic prototypes, for example, tend to solve highly complex problems with small codebases, but the problems are often contrived or they are tiny pieces of systems that would have to be much larger to be of any use.
Software systems you encounter in your run-of-the-mill software engineering business often have huge codebases, but the large size is not due to the fact that there's any problem there of any real complexity. The system just needs to solve trivial problems, and there are just a lot of them.
But you can certainly have systems that need to solve many problems and some of those problems have high intrinsic complexity. In such a scenario, it's important to recognize that there's a part of your system that's "extra special" due to its complexity, and needs to be approached in a way that's different from how you approach most of your engineering. You might need to write a library in a literate programming style, and then use that library from your other code, which isn't written in literate style. Engineering organizations are seldom able to have that kind of flexibility. The opinion you're likely to get is "we don't do literate programming here". The outcome is that such organizations are just not capable of writing such systems, because a system is only as good as the worst job you've done at solving any of the problems it raises.
I'd also like to point out that you're guilty of a bit of a strawman argument there, because all I did was point out a technique I sometimes find useful, and you responded as if I was advocating applying it "irrespective of costs and drawbacks" by pointing out that this often ends up as "crimes against software".
Reading and understanding an f# code base has always been vastly more simple to me than the incredible indirection of some other codebases (trying to find where the provider for the factory of the abstract class member that’s an interface implemented by 10 almost exactly similar c# classes? Good luck /s)
It's a nightmare. Very difficult to follow.
...see sibling comment, where the commenter does seem to imply they rather like it.
OP's technique, notably, avoids these problems by numbering the files but not actually mandating at the compiler/language level that they be ordered. Here they are acting as a hint to the programmer but the files are still, as far as the language is concerned, orderless. That would seem to be a nice compromise if you like the idea of an ordered codebase. More of a literate programming technique.
My personal experience has been that it's a pain to implement this as soon as you have strong typing and try to avoid "unsafe" casts. For example, in modern Java, if you have some kind of a Container<T> over elements of type T, and a matching ContainerIterator<T>, you will want some way in each of the type declarations to refer to the other type declaration, so you need cyclical imports.
If you look at it through the lens of a dynamically-typed language like Python 3.4 or Lua, the problem just disappears. This is the case to a lesser extent even for Java before it had generics, and collections were just collections of Object, and you used "unsafe" casts all the time. I am finding this practice not the slightest bit limiting in Lua, nor do I get the feeling that the readability of my code suffers in any way at all. (Rather the contrary, as I've pointed out before).
It's valuable to compare F# against C++ here. F# and C++ both require ordered declarations within a file (while C# and Java do not), but F# requires the files themselves to be ordered, too, while C++ does not. Both are statically typed languages. C++ avoids requiring a file-level order by defining file-level compilation units and a system for importing and exporting symbols between compilation units to allow each file to be processed separately with the imports/exports resolved during a final linking phase. There's complexity involved in accomplishing this. F# has not done it; an F# project is the same as concatenating all of the source files together, and thus declarations must be ordered across files just like they are within a file.
0: https://import-linter.readthedocs.io/en/stable/authors.html
For example, I much prefer this:
project/
feature_1/
view.py
service.py
model.py
feature_2/
view.py
service.py
model.py
project/
views/
feature_1_view.py
feature_2_view.py
services/
feature_1_service.py
feature_2_service.py
models/
feature_1_models.py
feature_2_models.py
What is mentioned in the article reminds of the Polylith Architecture, even if Polylith probably has a more simplistic view on code: a flat structure, made of something called “bricks” - that are small namespace packages. These ones are combined into features, the features are combined into apps or services.
You can choose how to deploy your artifacts - as a single Monolith, or several microservices. Since the code is not coupled to the built artifact, it isn’t a big thing to change it from one type to another.
Docs here if you want to know more: https://davidvujic.github.io/python-polylith-docs/
We use a similar tool for PHP - deptrac, it checks that our architectural decisions about layers and modules, and their interactions, are not violated. If they're violated, the build fails.
Interestingly, we don't use similar tools for our microservices. I'm not sure whether it's an oversight or it's just unnecessary in a microservice.
I tried clicking the bottom and the corner and it just doesn't work. Is it a dark pattern? Is it done deliberately to make people click on the newsletter button more?