Hacker News Reader

Comment by ivan_gammel | Hacker News Reader

I hope they will find a better solution, as this proposal will break encapsulation.

Let’s say you model allowed transitions via business methods and thus restrict certain state changes. You can either serialize or deserialize state or perform an allowed transition. If wither becomes a language feature, it will allow forbidden state changes that cannot be caught by validation in constructor, which will allow deserialization of states A, B, C, but will not know that A->C is forbidden, thus making the following possible:

   var r = new R(A);
   r = r with { state = C; } // passes, resulting in untraceable error much later

where alternative could be:

   r = r.transitionTo(C); // throws exception

View on HN

The proposal is meant for records, and records are used for "plain old data", explicitly without any encapsulation.

Nope. If records were structures, they would not be allowed to have methods or overloaded constructors or even default constructor. The most obvious example of encapsulation in records is validation of state.

They still don’t have encapsulation, all members are readable on purpose.

Withers will likely go through the same default constructor so constraints can be upheld, though.

>They still don’t have encapsulation, all members are readable on purpose.

Public state does not mean no encapsulation. The purpose of encapsulation is to bundle state and behavior and hide them both behind an interface, but that interface can offer read access to state. The key here is behavior.

>constraints can be upheld

My example above demonstrates a constraint that cannot be implemented in a constructor.

It can't but that's not due to the proposed solution for withers.

The cause is Records themselves. Record are immutable on purpose and you want to add a mutation constraint whish will not play nicely.

The proposed constraint will be triggered in this case :

    var r = retrieveFromDb(); (value is A)
    r = r.with(C) -> throw an exception

But this only work if B is generated by the wither. If I deconstruct the record manually and reconstruct with C manually, it'll work. So this offer no garantee that this transition will never occur.

I would even argue that this constraint can't be implemented at the level of the class, at least if the class is only a data carrier without external dependencies.

>>My example above demonstrates a constraint that cannot be implemented in a constructor.

>It can't but that's not due to the proposed solution for withers.

This has nothing to do with withers. It can't be implemented simply because it is a constraint on a specific transition of state. There's no transition of state in constructor.

>Record are immutable on purpose and you want to add a mutation constraint whish will not play nicely.

The purpose of immutability is not to create constant objects, but to prevent side effects from sharing mutable objects: state modifications are possible, they are simply reflected in a modified copy of object. That also means that my argument stands also for entities modeled as classes with final fields, it has nothing to do with specifics of records.

Indeed, the fact that we have a constructor from which we can build any valid state and that we can deconstruct an immutable object means that we can bypass transition validations, but that will require some extra effort from developer and explicit demonstration of intent compared to simply using `with` block. Compare this:

  var rA = new R(I1, I2, I3, A); // deserialization, e.g. from persistent state
  
  // verbose, explicit intent to create a copy in state C
  var rC = new R(rA.i1(), rA.i2(), rA.i3(), C); // error occurs later

this:

  // no semantics, no validation
  var rC = rA with { state = C; } // error occurs later

and this:

  // clear semantics, validation of transition
  var rC = rA.onSomethingHappened(C); // exception thrown now

I can't find anything in the linked eg-draft that would indicate that the error would "occur later". In fact, it explicitly says:

    Note too that if the canonical constructor checks invariants, then a with expression will check them too. For example:
        record Rational(int num, int denom) {
            Rational {
                if (denom == 0)
                    throw new IllegalArgumentException("denom must not be zero");
            }
        }
    If we have a rational, and say
        r with { denom = 0; }
    we will get the same exception, since what this will do is unpack the numerator and denominator into mutable locals, mutate the denominator to zero, and then feed them back to the canonical constructor -- who will throw.

Please re-read the thread, the problem was described multiple times in it.

Withers invoke the constructor. E.g.

    r = r with { state = state + C; };

compiles to:

    r = new R(r.state() + C);

All state transitions of ADTs can be checked in constructors.

State transition cannot be checked by constructor because it doesn’t know previous state.

Ah, I see what you mean. It's not about that proposal, but about the fact that record constructors cannot be private (at least not for a public record). That's because records are meant as product types, or nominal tuples, which don't break encapsulation but exist to represent the notion of an unencapsulated tuple. That's their job: to be unencapsulated data with everything that enables.

Now, you may ask why they don't also do other things, and I guess its possible that in the future we'll allow private record constructors, but there's less need for that, because Java already has a construct for encapsulated state -- ordinary classes.

Regardless of what records were meant to be, they are designed in a way where they do offer encapsulation (which by definition means not only the state, but also the behavior of an object, and it does not mean that we always hide the state - only that we control the interface to it). Records can have business methods and can have non-trivial constructors, so they are basically a syntax sugar for special form of classes, not tuple structures. If you can use them to encapsulate certain forms of behavior, e.g. by declaring methods for state transitions, then with{} block will be a change in their interface. Do I expect that new version of language will change interfaces of my data structures? Hell, no! This feature must be designed as opt-in solution, following the example of Iterable and foreach loop and requiring explicit declaration of wither interface. What kind of declaration could that be?

Imagine uniform declaration of intent for records and classes like in this example:

    public record Point with (int x, int y) {}
    
    public class Order {
       public Order() with (OrderStatus status, Instant timestamp) { … }
       public @(OrderStatus status, Instant timestamp) { … }
    }

Here we explicitly say that Point generally supports „with“ block for all fields and Order supports deconstruction to status and timestamp and construction of a new object with the same fields. This way existing code retains the interface but can be easily modified to support the new syntax.

> they are designed in a way where they do offer encapsulation

They are very intentionally designed to represent unencapsulated data. Records can have non-trivial constructors, but they all have a public canonical constructor, and while you can do strange things in your constructor and accessors (we needed that for technical reasons), the JEP/Javadoc/tutorials warn you against doing so, and that the reasonable assumption is that you don't.

The invariant is that if you have a record and deconstruct it using a deconstructing pattern, then you can also reconstruct it to get an object that's equal to the first by using the public canonical constructor. You can break that invariant, but libraries are allowed to assume that you don't.

> If you can use them to encapsulate certain forms of behavior, e.g. by declaring methods for state transitions, then with{} block will be a change in their interface.

But you can't and so it won't. All (public) records have a public canonical constructor that you can use regardless of "state transition" methods, with or without withers. The relevant point, again, is not the "with" feature, but the publicness of the canonical constructor. You cannot limit the construction of a record to a state transition method even today, because you can't hide the canonical constructor.

There are certainly classes that do need private constructors, but if they do, then those classes are not records (we may expand the role of records in the future, but so far they're specifically designed to not allow that so that the reconstruction invariant is maintained).

> Imagine uniform declaration of intent for records and classes ...

There's no need to do that for records, because they have a public canonical constructor, and that is the constructor that's used by the feature.

I don’t believe this is true. From my reading of the document, all with expressions go through the primary constructor.

Well, that is exactly the problem, as I already explained a few times in this thread. Constructor cannot validate state transitions, so the “with” statement or the block surrounding it will have to do it, breaking encapsulation and likely requiring multiple copies of this code.

But if that’s the case, then a record is the wrong data type for what you’re doing anyway, since you can call the primary constructor at any point with those same un-validatable values.

It’s not “with” that’s problematic. If a record doesn’t work for what you’re trying to do, just use a class.

Though I will say that this is why I generally think ML’s (i.e. OCaml, Standard ML) approach to encapsulation with modules is generally superior to using classes for encapsulation.

Have a look at the proposal. Half of it is dedicated to regular classes, which makes sense, because record is just a syntactic sugar in Java at the moment.

Also this: https://news.ycombinator.com/item?id=34398982

Reply on news.ycombinator.com