Design Meeting Notes, 11/17/2017

[mhegazy]

Strict property initialization checks in classes

• PR: Strict property initialization checks in classes #20075
• Every non-initialized property that does not have undefined in it type will be an error if not definitely assigned in the constructor, (meaning all code paths needs to have an initializer).

• abstract properties should be excluded
• T extends number | underfed is an error, since we do not know if it is one or the other at instantiation time
• Enabled under new strict mode flag --strictPropertyInitialization
• Issues:

1. names in quotes are not checked:

   class B {
      a: number;  // Error
      "b" : number; // No error
   }

   • we do not do control flow analysis on x["foo"] currently, we only do it on x.foo.
   • we could do that if we address Discriminant property type guard not applied with bracket notation #10530
   • @sandersn did have a PR to address that, and the problem was perf
   • also applies to symbols, or any computed property names
2. Do we need a new explicit modifier?
   • How do you opt-out for a specific class
     • // @ts-ignore
       • seems dangerous to rely on this, you change the type, and now no errors
     • or put it in "b"
       • too cute..
       • this looks like a bug that we should fix in the future
     • new modifier:
       • kotlin uses lateinit, swift has lazy?
       • bang operator: a!: number
       • extend it to locals as well
         • not allowed on const
         • should be an error to have it on something with an initializer
         • readonly is ok to have ! to allow for external
         • does not amke sense for statics
         • @bterlson is not happy about !
           • but we have it in expression positions already
3. method calls in the constructor
   • this guards against observing un-initialized properties outside the class, but not inside the class
   • this allows for for init patterns
   • it is also easy to get around this using aliases, or even through super access

• @DanielRosenwasser, all angular projects, mobx, etc.. that use decorators to initialize properties will have to opt-out of these features

• you can bang these properties
• it is possible that in the future to allow decorators to change the type of the declaration, and in that world we would not need to bang
• this is opt-in at the end of the day, and you can keep the world you are used to

[DanielRosenwasser]

My notes:

---

Strict Property Initialization Checks in Classes

• Does adding | undefined to a declaration, do we still do initialization checking within the constructor?
  • No, it's in the domain, allowed not to initialize it.
  • What about with generics that have undefined as a constraint?
    • No, only signals it could be instantiated with undefined
• What about @decorated property declarations?
  • People can always opt out of this.
  • Code in the RWC doesn't break because they don't opt into strict.
• Does it make sense to have a modifier for this?
  • ! after the property name.
  • lateinit? Some sort of modifier to assure the compiler it doesn't need to write any checks.
  • Should we error on initialized declarations like a!: number = 10?
    • Probably should let linters take over this.
  • Make it !, it reflects expression level non-null assertion logic.
    • Allow it on locals.
    • But don't allow it on const.
    • Also, probably error when we give an initializer (nix that last point about linters taking this on)
    • But allow it on readonly.
• What about local methods that can be called from the constructor that might observe uninitialized values.
  • Only so far we can go.

[tinganho]

you can bang these properties

This sounds funny.

Nearly all of our classes are handled by some sort of IoC container.

Sometimes(99% in my case) classes, are just definitions and construction is "inversioned". The IoC container takes care of construction of the class. I don't want the compiler to complain in these type of classes. And providing a flag to opt-out or banging all the properties doesn't sound optimal either.

One suggestion, I can think of, at least for the decorator use case, is to have a modifier on the "definition" class decorator.

In the below example you can see a constructtarget modifier:

export function Entity(definitions?: EntityDefinitions) {
    return constructtarget function(target: any) {
        //...
    };
}

And applying it on entity definitions will not cause any errors:

@Entity()
class User {
    @Column()
    name: string; // no error
    @Column()
    age: number; // no error
}

Example from TypeOrm https://github.com/typeorm/typeorm.

[DanielRosenwasser]

We discussed decorators that can declare how they modify the target, so it's an option, but others in the meeting felt like doing so shouldn't be a blocker for the flag as it exists today.

[mihailik]

Should this flag hold off from graduating to --strict set?

If it needs to be disabled for a all/large subset of angular code out there, perception of TS as a stable platform might be affected.

[mhegazy]

Should this flag hold off from graduating to --strict set?

we have the infrastructure in place to opt-out of specific --strict flags as needed. --strict --strictPropertyInitialization false. --strict by definition means you are accepting breaking changes between releases. if this is not your intention, then do not use --strict and instead spell out the flags you are interested in, e.g. --noImplictAny --strictNullchecks

[mihailik]

Hope you're right — regardless of technical reasoning you still want gain > pain to proceed

[tinganho]

I think many people would rather bang all properties than setting the flag, to just keep the default "strict" mode. But I hope the TS team can hold this off a bit, certainly a contentious topic. As of right now, it feels like a choice between pest and colera.