8 months ago

Once a project leaves the rapid development phase and enters the calm waters of maintanenance mode, the programming paradigm changes. Fancy hyped frameworks favored by project creators no longer play important role. What matters most now is the code readability and understandability. The code is already written - now it is going to be read, understood and fixed. After all, that's usually what happens to the code - it is written once but read by many. This is true both for open-source, but also for closed-source: people in the team come and go, and new team members need to read the code, to familiarize themselves with it.

So far I have came up with the following main principle of project maintenance:

  • You must be able to understand what the program does, without having to run it

When fixing bugs originated from production, you often don't have the possibility to reproduce the bug in your testing environment - the production usually contains a different set of users, different data, the workflow to reproduce the bug may not be clear, etc. The only things you actually have are: a log file, a stacktrace, a vague description from the user on what she did and how she interacted with the app, and perhaps a screenshot. Usually this alone is not enough to reproduce the bug, so you start examining the code, looking for possible code execution flows that might have been taken by the CPU to trigger the bug.

In other words, we maintainers need to invest the mental power to read the code surrounding the stacktrace in hopes to discover the proper code execution flow which may lead to the bug. All that so that we could reproduce/trigger the bug ourselves, in our happy contained debug environment. Thus:

  • The execution flow of the code must be obvious: you can't understand the algorithm if you can't follow the instructions in the algorithm. Too many misnamed utility methods, too many asynchronous calls with callbacks (especially if they continue the algorithm in a completely different class/actor/observable) seriously hinder the readability.

To find the code execution flow that led to the bug, you usually need two things:

  • A stacktrace, with possibly good error message, to point you at the end of the code execution flow, and
  • The code, obviously.

The better the stacktrace, the stacktrace's message and the code, the faster we can discover the code execution flow. In order to discover the code execution flow faster, the code should have the following properties:

  • Code Readability - it's obvious on the first sight what the code does. By no means I'm implying that the code is the documentation; however to fix a bug, one first needs to read the code and understand it. Badly named variables, too many clever filters in your List-filtering code, and you'll need comments to explain that particular spaghetti.
  • Code Locality - the code doing similar functionality should be co-located: in one class, in one file, in one package. That's how we human beings like things that relate - nicely wrapped in one box.
  • Ability to Navigate - to understand the algorithm one must understand the functions which the algorithm calls. You must be able to Ctrl+click to anything in the code and IDE must navigate to code which implements that thing, be it a function, class, interface, operator, property, anything.

I have drafted a sketch list of things that tend to hinder these properties.

Overzealous protection against copy-paste

Affects: Code Readability

So now you've extracted all repetitive code to utility methods. Congratulations! Now your function calls a function, which calls a function, which calls a function... If those functions are not well-defined or they are misnamed, the code readability will suffer.

Often there are multiple methods which share only a few lines of the code. One usually extracts the moving parts to an abstract class (generally a bad idea, see below) or introduces some builder, with moving parts injected as blocks or Runnables (ugly, but better than using inheritance).

Remember: every rule has an exception and sometimes it's better to copy-paste. Especially in the UI world: either turn part of the UI into a reusable component, or copy-paste mercilessly.

OOP Inheritance

Affects: Code Readability

The main principle of OOP is still awesome: cram related code to a single class, which improves code locality. Hiding stuff into private produces sane autocompletion results and makes your component really usable. Trouble starts when you start using functionality from those classes. You generally have to choose whether to use functionality (that is, call methods) from other class either by reference, or by inheritance.

Composition promotes components: reusable objects with well-defined API that do not leak their internals and do stuff. This is the good approach.

Inheritance promotes adding functionality by extending classes. Peachy, but it's not easy to do it properly. Quoting Josh Bloch: Design and document for inheritance or else prohibit it:

  1. Your protected methods are now part of the API as well
  2. Constructor must not invoke overridable methods. This is because the overrided method in the constructor of super class will always be invoked before the constructor of the sub class. This may cause invocation of an object before that object has been initialised properly.

Ever tried to read a class which extended a deep hierarchy, and methods were magically called out of nowhere? You're effectively combining your class with all of the code from all of the subclasses, and you need to understand those as well. This hurts code locality (since superclasses are located completely elsewhere), and also code navigability (since it may only be clear at runtime what the method will actually do, thanks to polymorphism).

I once had a colleague who blindly followed the newest hip rule in OOP, quoting: "every class must have an interface". There were lots of interfaces in that project. Ctrl+Click would bring you to an interface. Luckily nowadays with Intellij you just press Navigate - Implementations, but it hurted the code navigability quite badly back then.

Dynamically-Typed Languages

Affects: Ability to Navigate

The very foundations of dynamic languages (say, JavaScript, Ruby, Python) are based on duck-typing - if a class has quack() method then it can quack. The problem with duck-typing starts to show once you Ctrl+click on the quack() function call and realize that the IDE has no way of knowing of:

  • which class you are trying to call (since you often do not specify the type of the class)
  • whether the callee class actually has the quack() method or not (since the method_missing() catch-all may be used)

Thus, on every Ctrl+click, the IDE usually presents all classes that define quack() method (unless it can guess the callee class type); and this list may actually not list all possible target code points because of method_missing().

Incorrect Code Grouping

Affects: Code Locality

Sometimes people tend to put all JPA entities into one package, all Controllers into the controller package, all Views into the view package. This causes the code to be grouped by type instead of by the functionality. This is the same fallacy as grouping racing cars by color (not the best parable but it's the best I have ;).

The fix is actually quite easy: all Views, Controllers and JPA entities dealing with the customer should instead be placed into the customer package (unless of course there is a glaring reason to do otherwise, for example when the JPA entity resides in a completely another layer, invoked remotely via REST).

Overuse of Annotations (Annotatiomania)

Affects: Ability to Navigate, Code Readability, Code Locality

An example:

public void manage(@Qualifier('time')int time) {

Ctrl-clicking on @Transaction will take you to the Transaction annotation definition which contains no code. To actually see the code, one needs to search for any code touching that annotation in all libraries (and this is actually impossible with JavaEE). And then one usually encounters an uber-complex function dealing with all of the annotation parameters (say, code dealing with Required/RequiresNew/Mandatory/NotSupported/Supports/Never transaction attributes, XA transaction handling, bean/container managed transactions) which is virtually impossible to grok in a sane time span.

The idea behind annotations is noble: to standardize a particular functionality and make it configurable via annotations. Using the convention-over-configuration principle one can often simplify the annotated class so that the intent is immediately clear (say, @RequiresRole("ROLE_ADMIN")). However, often that is not enough, and in the fervous attempt of avoiding code writing, an annotation-based DSL language is born. Usually with morbid results, for example @Secure({@And(@RequiresRole(@Or("ROLE_ADMIN", "ROLE_MANAGEMENT")), @RequiresLoginPlace("ACME_HQ")}). The problem is that annotations do not actually execute anything; the actual implementation handling all aspects of the configuration is thus someplace else, not easily reachable (making the possibility to navigate hard). This can be remedied for example by specifying access verifier class in the annotation; however the code that actually calls the verifier class is still hardly reachable.

However, sometimes this may be acceptable. Say, Retrofit: the library is simple enough so that the functionality is clear; also one rarely needs to actually dig into its core and check how exactly that GET is executed. And since Retrofit forms a simple detained island which does just one thing (as opposed to Spring with all its plugins which tend to permeate all of your code base), the complexity is detained within an acceptable black box.

The interceptor annotations such as @Transactional can be replaced with custom functions with blocks, for example:

db { em.persist(person) }

The db is just a function which starts and commits (or rollbacks) the transaction. It has some 15-20 lines, is easy to understand, and you can just Ctrl+Click on the db function name to navigate directly to the source code. The idea behind the db {} function is described in mode depth in Writing Vaadin Apps In Kotlin part 3. And you can just call this function from anywhere, as opposing of having an injector everywhere (by making all classes managed beans) and injecting @Transactional class which finally can call the database.

Should the db{} function stop serving your needs (no XA or other reason), just fork and write your own, or use XA vendors. It's actually not that hard: the dirty little secret of XA is that it is still not 100% error proof; naive approach may work here quite well.


Affects: Ability to Navigate

JPA is a tough beast. Its sheer amount of annotations makes you want to run and hide. There is absolutely no way one can understand what's going on behind the scenes by a simple Ctrl+click approach. However, once one follows a good tutorial and learns how to navigate himself around pitfalls such as detached beans, then it does its job and is again an acceptable black box.

Yet, I grew weary of constant surge of exceptions, N+1 selection problems and now I just select exactly what I need with plain SQL selects, using sql2o plus a bit of code to add simple CRUD capabilities in the Vaadin On Kotlin framework.

Loose Coupling

Affects: Ability to Navigate

Loose Coupling of large modules makes sense - it's not easy to understand the modular structure of the app if all of the modules leaks internals like hell. However, it is easy to go all loose-coupling-crazy, stuff everything behind interfaces and thus turning Ctrl+click into searching for implementors of that particular function (and God forbid should there be more than one!). The granularity here really matters. Loose Couple on module level, tightly couple on package level.

Argumenting by loose coupling is often fallacious. Say, "using Rx makes your code loosely coupled". This actually sounds like a good thing, until you realize that you can no longer see the code flow! To rebute this fallacy, just ask:

  1. Do we need loose coupling in this particular place?
  2. Does it outweight the cost of not seeing the code flow anymore? (Usually it doesn't).
  3. Are you just fallaciously promoting your favourite framework?

Remember: Loose coupling directly reduces the ability to navigate and the possibility to read the code flow.

Interceptors / Aspects

Affects: Ability to Navigate

The interceptors/aspects are immensely popular with the Dependency Injection frameworks. However, when overused, this can become a maintenance nightmare. Usually interceptors are employed by the means of annotating class/function, which hurts the ability to navigate, but is still doable - by searching all accessors of that particular annotation one can eventually learn the code that "executes" the annotation (even though this gets harder as the number of annotated components increase). However, interceptors can also be configured (in Spring XML or Guice Module) so that theyapply to certain class types (and subclasses) only, and that makes them virtually undiscoverable for any coder who just joined the project. This directly boosts the coder's frustration and the desire to leave that project, which is not what you want.

One of the solutions may be to use functions with blocks instead, as shown with the db {} example.

Java Stream API

Affects: Code Readability

A simple filtering with Java Stream API turns into this:

list.stream().filter(it -> it.age > 0).collect(Collectors.toList());

The stream() and collect(Collectors) is just a noise, however in Java this noise takes almost half of the entire expression, which decreases readability. Compare that with the following Kotlin expression:

list.filter { it.age > 0 }   // the result is already a list.

Going more stream-crazy allows you to write unreadable code like this:

Stream<Pair> allFactorials = Stream.iterate(
  new Pair(BigInteger.ONE, BigInteger.ONE),
  x -> new Pair(
return allFactorials.filter(
  (x) -> x.num.equals(num)).findAny().get().value;

Please read more at This Is Not Functional Programming. Streams are to be used cautiously: Java's nature is not a functional one, and using Java with functional approach is going to produce weird code with lots of syntactic sugar, which will make little sense to traditional Java programmers.

To avoid the noise introduced by Java you can consider switching to Kotlin.

Java Optional API

Affects: Code Readability

Optional is the Java's way to deal with nulls. Unfortunately Optional introduces methods which tend to be used instead of the built-in if statements. Consider this code:

optionalString.flatMap(s->something).ifPresent(s->{System.out.println(s.toString());}).orElseThrow(()->new RuntimeException("something"));

Using the traditional approach of ifs and null-checks would have worked better here.

Optional is also capable of breaking of the Bean API. Consider the following code:

public void setFoo(String foo) {}
public Optional<String> getFoo() {}

This is actually not a property since the type for setter and getter differs. This also breaks the Kotlin support for propertyising Java beans. The code could be fixed by changing the setter to setFoo(Optional<String> foo) but that makes it really annoying to call the setFoo method.

Instead, it is better to use Intellij's @NotNull annotation and traditional language statements. The resulting code tends to be way more readable than any Optional black magic.

Spring, JavaEE, Dependency Injection

Affects: Ability to Navigate, Code Readability, Code Locality

Spring was founded as an alternative for J2EE, which in the old days contained a lot of XMLs. At that time it was lean; nowadays it has tons of modules, with varying quality of the documentation. It seems that Spring permeates everything and uses all of the above methods:

  • In a lots of projects, the wiring granularity is too high. Instead of wiring modules, it often wires classes, often with synthetic interfaces in front of those. This is basically Loose Coupling on a way too granular level, which makes the navigation in such code harder.
  • Uses interceptors/annotations heavily. Spring Data, Spring Security for example. Spring Security has became so huge that coders are avoiding it on purpose.
  • Uses clever tricks to employ interceptors. To employ interceptors to a class, Spring uses the CGLIB library to create a dynamic class which extends the original one, overrides all methods and invokes interceptor chain. This has the unfortunate side-effect of having lots of $$$PROXY fields with null values which clutter the state of the original class.

JavaEE suffers from the very same problems, and adds the following issues on top:

  • It is impossible to see the actual implementation of, say, @Transactional, since it's implemented somewhere in the application server, which may not even be OSS. And even if it was possible, the code is highly complex and hard to read.
  • If JavaEE can't employ annotation, it simply will ignore it. For example, calling @Asynchronous method on the same bean instance will silently result in a synchronous call. This directly violates the main principle since the program silently does something else than it should, from the look of the code. To actually call the method asynchronously, the class needs to inject itself and call the method on that.

The Dependency Injection paradigm is highly invasive. For example, the need of injecting classes into Vaadin (or Swing, JavaFX) components: since all DI frameworks usually tend to hide Injector from the programmer, the programmer usually makes every Vaadin component a managed bean, to have depending services injected. This makes it harder to use java-debug class-reloading techniques.

I've seen people using Spring only because of Spring Boot, since it is a pain in the ass to launch a war project in Eclipse. Let me rephrase: one's code base hurts just because the tooling of his choice sucks.

Overuse of async: Reactive, EventBus, AKKA

Affects: Code Locality, Ability to Navigate

Reactive, Rx, React or whatever the asynchronous programming is called today. On larger granularity, and with proper task, it makes total sense to employ AKKA Actors: when they are big enough in order for them to form a grokkable box of synchronous functionality in which you can navigate. Also, in large loads, having too many thread being blocked will cause their stacks to eat the memory, so async is always the way to go, right? It depends.

It is true that it is extremely handy to have only a single thread in node.js, or a single thread in AKKA actor consuming messages, as opposed to traditional multi-threaded approach. With a single thread, you don't need to worry yourself with guarding your shared state with locks. And with advanced techniques like promises and/or coroutines, neither code locality nor ability to navigate suffers.

Shit tends to hit the fan when you have a hammer labelled Rx and every problem is a nail. Then, usually Observable gets slammed on every button listener, splitting the previously readable synchronous listener method into chunks of three-line asynchronous objects. There is no way to navigate there - this is like a too granular Dependency Injection all over again, since the code which configures (creates a pipelines of those asynchronous small blocks) is usually completely elsewhere than the code itself. This horribly affects the code locality.

Even worse when you have lots of actors with no promises/coroutines to handle request/response type of messages. If all you have is just an onMessage() method, then it's usually very hard or impossible to figure out who sent the message and thus who the caller is, and thus the ability to navigate suffers. And even more worse when the actor needs to send and receive messages in certain order. Then you need to mark down in the actor state which message you already received and which you still need. You are effectively creating a defensive state machine. State machines are hard to maintain, since by adding more state raises the possible code execution flows exponentionally. This price is rarely worth more than any dubious benefits the async approach may have.

Debugging suffers with async as well - it is no longer possible to debug and check the stacktrace for the information who called you, since the caller is usually a background message-processing thread. To actually get the caller, you need to track the code which created the message, as mentioned above.

The motto of EventBus - fire [the event] and forget - can be read as I have no idea what is going on when I fire this event since it's nearly impossible in design time to discover all event receivers who gets notified. This hurts the ability to navigate.

Java 8 Default Methods on Interfaces

Affects: Code Locality

Everything is fine when the default methods are used purely as utility methods (or in place of extension methods since Java is lacking those). The hell breaks loose if the interfaces are used as traits, adding completely new functionality, demanding to store the state as in the following example:

interface Bark {
    public abstract int getBarkedCount();
    public abstract int increaseBarkedCount();
    default public String bark() {
        return "barked " + getBarkedCount() + " times";

Of course this example is simple, silly and self-explanatory. Yet, imagine having 4-5 of such interfaces implemented on an object, each bringing its tiny little functionality inside the object, but every piece of functionality stored in a different file. The code locality suffers: it can no longer be seen what the object is responsible for, what's the actual state and what are the state mutation paths. The object becomes a God object with a lots of responsibilities.

Missing Comments

Affects: Code Readability

Quoting Uncle Bob's Clean Code book, chapter 4:

Comments are, at best, a necessary evil. If our programming languages were expressive enough [...], we would not need comments very much. The proper use of comments is to compensate for our failure to express ourselves in the code.

That's totally wrong, stupid, and very dangerous:

  1. People who take that literally will end up with having no comments in their code, and they will be proud of that. I can't even begin to describe how stupid that is.
  2. Uncle Bob actually contradicts himself - at first he says that all comments are evil, and then he lists examples of good comments: legal comments (!!! yeah, truly the most important type of comments. Uncle Bob, certainly you do care about code clarity, right?), informative comments, explanation of intent...
  3. The "expressive language" myth is easy to debunk: there is no such language. Scala tried and failed to produce not even code with intents, but even readable code. Computers don't execute intents, they execute code and they don't care about intents. Computer code is not about what we want the computer to do, it is about what the computer will do, and those are two different things. Thus, we need comments to explain our intents.

In short, having no comments because of fear that the comments may get obsolete, is just plain stupid.


Affects: Code Readability

There are languages which say what they do in English, and then there's Scala:

← Debugging GWT superdevmode remotely from your Android Phone The Fallacy of Future-proof →