Ongoing Exploration

Towards class-less JavaScript

I am getting more and more convinced that classes in JavaScript or TypeScript are an anti-pattern. At least in the way I use them.

I guess this will be a familiar example to you:

class PetStoreManager {
  constructor(petRepository: PetRepository, account: CheckingAccount) {
    this.petRepository = petRepository;
    this.account = account;
  }

  buyPet(name, price) {
    const pet = this.petRepository.findByName(name);
    if (!pet) throw new Error(`Pet ${name} not found!`);
    this.petRepository.remove(pet);
    this.account.addFunds(price);
    return pet;
  }
}

PetStoreManager is a class, which when instantiated receives dependencies, and upon calling the butPet method, will interact with these collaborators according to the business logic (When a pet is bought, add the paid price to the funds.)

petRepository is an instance of another class:

interface PetRepository {
  create(name: string): void
  findByName(name: string): Pet | undefined
  remove(pet: Pet): void
}

class FileSystemPetRepository extends PetRepository {
  constructor(storageDir) {
    this.storageDir = storageDir
  }
  create(name) {
    writeFile(
      path.resolve(storageDir, `${name}.json`),
      JSON.stringify({
        name
      })
    )
  }
  findByName(name) {
    return JSON.parse(readFile(path.resolve(storageDir, `${name}.json`)))
  }
  remove(pet) {
    unlink(readFile(path.resolve(storageDir, `${pet.name}.json`)))
  }
}

We use these classes to achieve the S, L and D of the S.O.L.I.D. principles:

  • Single responsibility principle: we separate the concerns of handling the transaction and storing pets in two separate classes, which makes it easier to understand what they do and also it simplifies testing this code.
  • Liskov substitution principle: the PetStoreManager class does not care how the pets are stored, one can swap out the instance of the FileSystemPetRepository with a MySQLPerRepository without needing to modify the managers implementation. The PetStoreManager is getting passed an instance of the repository, it does not use new ... to create an instance. This is important so we can test the implementation and can effectively configure our software depending on the environment.
  • Dependency inversion principle: The PetStoreManager depends on a repository interface, not on the concrete implementation.

My observation however has been that by following this pattern we eventually end up violation the Single responsibility principle because classes give us and arbitrary boundary which is hard to pierce because there is no easy and obvious way to slice a class. We end up adding more and more methods to classes who roughly interact with the same dependencies. I have seen multiple instances where class methods were added which introduced new dependencies, and they were added to the constructor:

class FileSystemPetRepository extends PetRepository {
-  constructor(storageDir) {
+  constructor(storageDir, storageQuota) {
    this.storageDir = storageDir
+   this.storageQuota = storageQuota
  }
  create(name) {
+   this.storageQuota.hasQuota(storageDir)
    writeFile(
      path.resolve(storageDir, `${name}.json`),
      JSON.stringify({
        name
      })
    )
  }
}

Now create needs to check if there is enough quota in the storage before adding a new entry. But findByName and remove can be kept unchanged, since for them the quote information is not relevant or respectively implicit.

We keep growing the class further and further and this bugs me for multiple reasons:

  • for every dependency we are also adding an import statement, a constructor argument, and a class property. This adds additional noise in the beginning of the file.
  • it gets harder to reason which method needs which dependency
  • class files become longer and longer, easily stretching above 500 lines

What if we had no classes in the first place?

The most obvious reasons for using classes is as collaboration holders, meaning they provide the dependencies which the business logic needs to interact. There are basically no classes in my code that have state in the classically sense: storing computation results to be reused later. Because this is hard to test I avoid having state in my classes and methods will only have local state and return results. They are already functional.

Classes were not meant to be used for that, that's why using Classes for the purpose of dependency management is an anti-pattern.

Now, where do we go from here? How does a better version of that code looks like, that addresses the concerns listed above:

  • only provide the dependencies for a method that are needed
  • shorter files

My current practice is to convert interfaces to TypeScript types for each method:

// petRepository.ts
export type create = (name: string) => void;
export type findByName = (name: string) => Pet | undefined;
export type remove = (pet: Pet) => void;

The concrete implementation for e.g. the file-system backed pet repository looks like this:

// petRepository/fileSystem/remove.ts
export const remove: petRepository.remove = (pet: Pet) => {
  unlink(readFile(path.resolve(storageDir, `${pet.name}.json`)));
};

and convert all class methods to functions which receive the implementations of these types.

// petstoreManager/buyPet.ts
export const buyPet: (
  findPetByName: petRepository.findByName,
  removePet: petRepository.remove,
  addFunds: checkingAccount.addFunds
  name: string,
  price: number
) => {
    const pet = findPetByName(name);
    if (!pet) throw new Error(`Pet ${name} not found!`);
    removePet(pet);
    addFunds(price);
    return pet
  }

In cases were an implementation has a dependency but the consumer does not know about this dependency I return the implementation with the bound dependency.

Here is the example of the file-system backed pet repository method create which needs to ask for the quota.

// storageQuota.ts
export type hasQuota = (location: string) => boolean
// petRepository/fileSystem/remove.ts
export const create = (hasQuota: storageQuota.hasQuota) => (name: string) => {
  writeFile(...)
}

When I need a petStore.create method somewhere in the code it is constructed like this:

import * as petStore from "petStore";
import { create } from "petRepository/fileSystem/create";
import { hasQuota } from "storageQuota/fileSystem/hasQuota";
const createPet: petStore.create = create(hasQuota);

// The code that creats a new pet does not know
// about the quota dependency
createPet("Kitty");

This approach solves the issues I listed, but feels a little clunky. Especially passing dependencies to methods is, at least for me, a thing I would need to get used to.

I'd love to hear your feedback and thoughts on this topic, this is an ongoing exploration of mine, so keep checking back here or on my Twitter for updates.