Comment by weavejester | Hacker News Reader

Comment by weavejester | Hacker News Reader

Parent Full thread

> In fact the opposite is true: raw values have no valid operations and types allow you to add operations that make sense for those values.

I think you're using 'value' to mean something slightly different to the way I meant it. So to get us onto the same page, by 'value', I mean something independent of how its stored; that is, the number 1 is a value, whether it's stored as as 00000001 or 0000000000000001.

A type limits both which values it is possible to represent (i.e. a u8 limits us to representing the integers 0 to 255), and also determines how it is encoded in memory (in this case 8 bits).

> Most statically-typed languages are actually very loose about how values are structured at runtime

Yes, but ultimately the compiler needs to be able to map a sequence of bits to the value it represents, even if there's not a strict one-to-one mapping.

> I'll have to do some guesswork here, but I imagine when people make arguments like this they are significantly imagining a situation in which, say, all the values are expected to be in the range [5, 100] and some befuddled experimenter or piece of machinery gives you the value 2

That's a common assumption, but the reality can be much more messy. It might be that we expect an integer between 5 and 100, but receive a string of UTF-8 characters instead.

For example, suppose you want to record a patient's date of birth. In Clojure, we might represent that as a map that connects a :patient/birthdate key with an encoded date object:

    {:patient/birthdate #date "1972-04-08"}

But what if the patient doesn't know their exact birthdate? Perhaps they immigrated when they were a young child from a less developed country and don't know the exact year they were born in. However, they tell the doctor that they do know they were born before 1980, because that's when they first arrived in the country.

In this case, the doctor might record the data as:

    {:patient/birthdate "before 1980"}

Even though the data doesn't match the type we expect (a date), it's important that it still be recorded as it could affect the medicine that the patient is given. These sorts of messy entries are not uncommon in areas where it's more important to accurately record the data than precisely type it.

> I think you're using 'value' to mean something slightly different to the way I meant it. So to get us onto the same page, by 'value', I mean something independent of how its stored; that is, the number 1 is a value, whether it's stored as as 00000001 or 0000000000000001.

I'm using it in its most general sense: as an object of discourse in a programming language, independent of representation or semantics. 1 is a value in most languages, to be sure, but it's also specific type of value, viz. a number: you can do number things to it, like add it or divide it, or check it for equality with 2.

> A type limits both which values it is possible to represent (i.e. a u8 limits us to representing the integers 0 to 255), and also determines how it is encoded in memory (in this case 8 bits).

To reiterate: a type doesn't limit the values but specifies the values (or rather, more generally, the meaningful operations on a value of that type). Values are not numbers by default; only by being typed as a number does a value take on number semantics. Without the knowledge that a value is a number it is meaningless to treat it as a number.

> Yes, but ultimately the compiler needs to be able to map a sequence of bits to the value it represents

Sure; in any language implementation you have to represent the values somehow, nobody could disagree. My point is that the type doesn't (necessarily) specify that representation in any language I can think of.

> That's a common assumption, but the reality can be much more messy. It might be that we expect an integer between 5 and 100, but receive a string of UTF-8 characters instead.

Sure: that's not fundamentally different from the first example I gave. The real type of `:patient/birthdate` there is just the (discriminated) union of the date type and the string type. It still has a type, and if it didn't you wouldn't be able to process it (definitionally, because a type tells you what kind of processing makes sense for the value). And the string values aren't ‘outside’ the type: even if you choose to write the wrong type down in your Clojure, the fact that you also process strings means that you know the real type (and you embed that knowledge into the code).

> To reiterate: a type doesn't limit the values but specifies the values (or rather, more generally, the meaningful operations on a value of that type).

But a value can have more than one possible type. The number 1 could come from an unsigned byte, or a signed long, for example. These are different types that support the same numerical operations, but differ in cardinality. So we can't say that a type's only purpose is to specify the meaningful operations on a value, as we might have two types that are identical in that regard.

> The real type of `:patient/birthdate` there is just the (discriminated) union of the date type and the string type.

Yes, in this particular instance that would be the case, but that's not necessarily something you know ahead of time. The point is that you may not have anticipated that not everyone would know their date of birth, and the data you receive is invalid according to your earlier assumptions.

In Clojure this results in a more graceful failure condition. Functions that don't require the date of birth will continue to work with no change required. If I want the average white blood cell count of a patient, I don't care what the date of birth is, and therefore the output for that particular operation isn't affected.

> But a value can have more than one possible type. The number 1 could come from an unsigned byte, or a signed long, for example.

Here we disagree. Unsigned 8-bit integers† and signed 64-bit integers, while both conveniently notated with Arabic numerals, are actually different values that support different operations, for example negation. They share quite a few similarities in how their operations interact with one another, for example each (assuming wrapping) is a monoid with 0 and +, but the semantics of the actual operations differs if looked at more closely. Mathematics agrees: the element ‘1’ of N/2⁸ and the element ‘1’ of Z/2⁶⁴ are not the same thing (their encodings coincide sometimes, but it's poor form to make assumptions about it).

† Bytes are data but not numbers, and so support only data operations like duplication and discarding, not number operations like adding and multiplication: as an artefact of representation you can usually ask the hardware or programming language to manipulate them as if they were numbers, but the result remains meaningless.

> The point is that you may not have anticipated that not everyone would know their date of birth, and the data you receive is invalid according to your earlier assumptions.

But that's exactly what I'm saying: you must have made some assumptions about that missing data, otherwise there is no safe thing you can do to it (including discarding it, which is a popular choice). This works in Clojure only because Clojure couples some semantics (data semantics plus operations on runtime type information) into every value, i.e. it restricts what values are even representable in the language in order to ensure that this function will always be safe to write.

In more strictly typed languages you can still talk about values that support these behaviours, but you are required to be explicit about it, because there are some values that can be represented that don't support these operations.

> Functions that don't require the date of birth will continue to work with no change required.

Any function that takes the date of birth ’requires’ the date of birth (or more generally a possibly-empty set of ‘leftover’ values). The only thing that differs is what's required from it: some functions might require that it be a date while other functions only require that it be data, for example. The universally imposed limitation that all values must be coupled to data semantics and runtime type information is convenient for ergonomics if you write a lot of these functions (since you don't have to remember to write that assumption down), but it's important to remember that it is a coupling — the resulting values are more complex than values without those things bundled on, and the trade-off is that you can no longer talk about values for which they don't hold.

> Here we disagree. Unsigned 8-bit integers† and signed 64-bit integers, while both conveniently notated with Arabic numerals, are actually different values that support different operations, for example negation.

What about a 32 bit unsigned integer and a 64 bit unsigned integer? Are they still separate values?

You appear to be saying that the values of a type cannot be a subset of another type; that is, there is no '1', only a '1' that is an integer, a '1' that is a short, and so forth, and every '1' is distinct.

Fine, that's a possible way of looking at it, but why is that more valid or consistent than a model that allows subtypes? That, for example, the value '1' could be both a Number, an Integer, and a NaturalNumber?

Further, what's the practical difference between a type defined as consisting of the numbers [1 2 3], and an integer that's restricted to those values?

> This works in Clojure only because Clojure couples some semantics (data semantics plus operations on runtime type information) into every value, i.e. it restricts what values are even representable in the language in order to ensure that this function will always be safe to write.

Even if we view a Clojure value as a coupling between type and data, that's only a coupling between two things, and it ensures we can avoid further coupling caused by large record types. On net, we reduce the amount of coupling a statically typed language that uses closed record types would require.

For instance:

    (defn average-wbc-count [{:patient/keys [wbc-counts]}]
      (/ (apply + wbc-counts) (count wbc-counts)))

This function is coupled to only one key/value pair. Any other information in the map is irrelevant, which is why an invalid :patient/birthdate doesn't cause the function to fail. There's no coupling between :patient/birthdate and average-wbc-count.

Conversely:

    fn average_wbc_count(patient: &Patient) -> f64 {
      patient.wbc_counts.iter().sum::<f64>() / patient.wbc_counts.len() as f64
    }

This function requires patient to be a Patient struct, and therefore the function is implicitly coupled to every field in the struct, regardless of whether that field is ever actually used. I need to ensure that a patient has some birthdate that's an anticipated type (even if that's an error type) before I can call the function.

> You appear to be saying that the values of a type cannot be a subset of another type; that is, there is no '1', only a '1' that is an integer, a '1' that is a short, and so forth, and every '1' is distinct.

That's true, and you can prove it simply by observing that they react differently to ‘the same’ operations, e.g. (2¹⁶ - 1) + 1 has a different value (that responds differently to tests like `≥ 0`) depending on which type we're talking about. There's an injection into the larger type (a type coercion) that is very well-behaved, but it's not an identity map — it changes the operations on the value.

> Further, what's the practical difference between a type defined as consisting of the numbers [1 2 3], and an integer that's restricted to those values?

Nothing (ish: you have to be careful to describe what happens to all the integer operations when restricted to your type) — that's a totally valid way to define a type. But it's not the only way to define a type, because the world of types is much bigger than the world of restricted sets of integers.

> I need to ensure that a patient has some birthdate that's an anticipated type

No, you just have to be explicit about the possibility of a birthdate (or other fields) being of a wide type, e.g.

    struct Patient {
      wbc_count: Vec<u64>,
      birth_date: Box<dyn Any>,
      other_fields: HashMap<String, Box<dyn Any>>,
    }

Clojure just attaches that by default to every value.

> Nothing (ish: you have to be careful to describe what happens to all the integer operations when restricted to your type) — that's a totally valid way to define a type.

Then at which point does a type become equivalent to a restriction that could be decoupled from the underlying type by choosing a broader type?

> No, you just have to be explicit about the possibility of a birthdate (or other fields) being of a wide type

Yes, you could recreate Clojure's semantics in Rust. More accurately, I'd say you'd be looking at defining it like so:

    struct WbcCount {
      wbc_count: Vec<u64>
    }

    struct BirthDate {
      birth_date: Box<dyn Any>
    }

Since we want to be able to reason about these keypairs individually. You might have a structure that has a WbcCount but not a BirthDate, or one with a BirthDate but not a WbcCount, or one with neither.

We'd then be faced with the challenge of creating a type that could contain an arbitrary number of unique structs, and to be able to pull a struct out of that set by its type. Realistically we'd probably just use a HashMap at that point and discard all static typing.

Alternatively, we could create a mega struct that contains every possible field we could want to use, whether or not we know they are related, and use this data structure to represent all structured data in the application. The "if everything is coupled, nothing is" approach.

But both of those options are difficult or unidiomatic to write.

Languages like Rust, Java, etc. encourage coupling of data because it's more convenient and space efficient to group data together in records/structs. In Clojure, there's no need to do so; we can couple only when necessary. If we don't need to know the birthdate to calculate the white blood cell count, then we can exclude that field from the input type checking. Most statically typed languages find this difficult, with TypeScript being one of the few exceptions in this regard.

> Then at which point does a type become equivalent to a restriction that could be decoupled from the underlying type by choosing a broader type?

I'm not totally sure how to interpret this question. If I already have a type that includes all the values I want as a subset, I can restrict that type by limiting the values it can take and restricting or removing the operations on it to guarantee they never produce any of the forbidden values. That's the basis of refinement type systems like Liquid Haskell etc. For any type I can describe this way I can also build it ‘from the ground up’ by starting with the empty type and adding operations, though it might not be as convenient. But the converse isn't true: not every type I can build additively can be refined subtractively from another type. For a start, you need to have a broader type to begin with, so that has to already be built somehow: you can't refine an 8-bit integer into an HTTP server.

> You might have a structure that has a WbcCount but not a BirthDate, or one with a BirthDate but not a WbcCount, or one with neither.

Remember that a `Box<dyn Any>` could also be a ‘no value’ type like `()` or a ‘maybe no value’ type like `Option<Date>`. But yes, there are many equivalent ways to write it depending on how likely the values are to exist; the precise formulation is just a question of ergonomics.

> We'd then be faced with the challenge of creating a type that could contain an arbitrary number of unique structs, and to be able to pull a struct out of that set by its type.

By its name, but yes, that's exactly what you get with the `other_fields` field in my example above.

> and discard all static typing

Well, that's not quite true: we discard (syntactically, but not conceptually!) the type information about the fields of this struct, but we can regain it later (from the conceptual information we retain) by trying to downcast the `Any`. Then you get your static type information back and the compiler can help you again.

> But both of those options are difficult or unidiomatic to write.

Sorry, I hope I haven't come across too harsh to Clojure here. As I thought I was clear about above, Clojure's ergonomics for writing dynamically-typed code are vastly superior to Rust's. By coupling runtime type information to values and restricting itself to only being able to deal with values that are data and have runtime type data attached, Clojure gets to make a whole bunch of simplifying assumptions that reduce the work the programmer has to do when dealing with such values. That's the trade-off: Rust chooses to be able to express a much wider range of types by not requiring that coupling, but in exchange you have to be much more explicit when you do want to couple runtime type information to your values. Some languages like C# aim to be able to do both, i.e. start strongly typed and drop down ergonomically into dynamic typing as desired, but Rust is not one of those languages.

> If we don't need to know the birthdate to calculate the white blood cell count, then we can exclude that field from the input type checking.

But that very exclusion requires that you know enough about the value to know that it can be safely ‘excluded’ without having any additional information about it. Specifically you need to know:

- how to access a subfield of the type regardless of the other fields on it — this requires that it contain RTTI

- how to discard the value — this requires that it have a destructor with a known calling convention

- how to move/copy the value into the function (and maybe even move out of the function if you return it or imperatively add it to some global state) — this requires that you know a ‘moving constructor’ for the value that can be used to safely move it to another location

In the Rust case you need to write down these things explicitly so the compiler can make sure you don't pass a value that doesn't satisfy them. In Clojure, the reason you don't need to write them down is that every value in the language is restricted to only be able to represent such values, so those constraints are implicitly on every function that you write.

> I'm not totally sure how to interpret this question. If I already have a type that includes all the values I want as a subset, I can restrict that type by limiting the values it can take and restricting or removing the operations on it to guarantee they never produce any of the forbidden values.

Sorry, perhaps I'm being a little obtuse. What I'm ultimately trying to get at is that sometimes its useful to have different types for the same data under different circumstances. A subset of data might benefit from a narrower type.

For example, suppose we have some CSV file that contains a bunch of patient information, including average white blood cell count and birthdate. This CSV file might contain bad data! So perhaps we type it as:

    struct Patient {
        avg_wbc_count: Either<f64, String>,
        birthdate: Either<Date, String>,
    }

This covers all our bases, but it's also somewhat annoying to work with. Perhaps we only want to find patients with WBC counts outside a certain range, and discard those lines where the data is invalid. In which case, we could write a narrower type, and simply not parse the CSV rows that are invalid:

    struct PatientWithKnownWbcCount {
       avg_wbc_count: <f64>,
    }

So in this case we have a couple of options for types. The latter type decouples the birthdate (it's never even parsed), but adds the restriction that the WBC count needs to be numerical; while the former type is a more accurate representation of the CSV file overall, but has greater coupling.

Obviously which we use depends on the nature of our program, but what I'm trying to get at is that a type isn't set in stone, but something we choose. Ideally we choose the most restrictive and decoupled type for the circumstances, but that might result in having many hundreds of different variations of the same type, so there's a tension between the number of distinct types and how specific or narrow they are.

> In the Rust case you need to write down these things explicitly so the compiler can make sure you don't pass a value that doesn't satisfy them. In Clojure, the reason you don't need to write them down is that every value in the language is restricted to only be able to represent such values, so those constraints are implicitly on every function that you write.

Granted, but Clojure's approach can result in greater decoupling with less effort, particularly when dealing with imperfect data.

In the previous examples I presented a scenario where we might want to think carefully about how exactly we store data that might contain invalid fields. But in Clojure we don't care - we can quite easily have a data structure that's partially invalid or unparsed - the entire problem of how we represent this data in memory is sidestepped.

We can still define what constitutes valid fields:

    (s/def :patient/birthdate     inst?)
    (s/def :patient/avg-wbc-count float?)

But these schema definitions are independent of the type system (i.e. decoupled), so we can apply them selectively or not at all. We can say "this function requires patient data with valid birthdates and WBC counts, but we don't care about any other field".

Could you do the same in Rust or other similar languages? Sure, it's ultimately just maps and predicates. But many languages aren't geared up to make that pleasant to use.

> Perhaps we only want to find patients with WBC counts outside a certain range, and discard those lines where the data is invalid. In which case, we could write a narrower type, and simply not parse the CSV rows that are invalid:

These two types aren't typing the same data, though: in the latter case (in Rust) you've actually thrown away some of the data, and the type reflects that. In some languages like Clojure or TypeScript you can't distinguish that from the case where you haven't actually thrown away the data and are secretly carrying it around as well as whatever data is explicitly typed there, but if you want to get it back you have to know that it was once there, i.e. have additional (conceptual) type information that you chose not to write down.

> Granted, but Clojure's approach can result in greater decoupling with less effort, particularly when dealing with imperfect data.

I think this isn't a great comparison: the usual vehicle for this thing in Rust is the trait, which manages to abstract over the representation statically without introducing a bunch of runtime machinery for it (though you can opt in to the machinery by using a trait object!). That's very idiomatic in Rust, although the trait syntax is a bit noisier (because it's more general).

> the entire problem of how we represent this data in memory is sidestepped.

It's not sidestepped — the language just makes an opinionated choice for you, which you can't avoid.

> these schema definitions are independent of the type system (i.e. decoupled), so we can apply them selectively or not at all

These are just predicates; I don't think they have much to do with this discussion? They don't help you to type your data: you (the programmer) still have to carry that type information around in your head in order to work with the values even after they are validated.

> Could you do the same in Rust or other similar languages? Sure, it's ultimately just maps and predicates. But many languages aren't geared up to make that pleasant to use.

There are good reasons for that, though. Nominal typing allows you to express and enforce constraints that aren't necessarily enforced by the structure of the type, at the cost of some extra ceremony. And systems languages often want to be able to express types that cannot be treated this way: types in which there is no sensible way to ‘ignore’ a value. Clojure makes some things easier to write by forcing you to carry around a bunch of extra stuff (both runtime data and static semantics) with all your values; that's a perfectly valid choice that makes an important subset of programs nicer to write, but it also excludes values that don't fit into that category or can't reasonably be coupled to their RTTI.

As we get better at writing compilers we're increasingly seeing a move towards mechanisms like Rust's traits or C++'s concepts that allow the programmer to actually abstract over the representation of a type (as opposed to forcing a uniform representation as is common in dynamically typed languages). If you do that well you can get the best of both worlds semantically, but these kinds of systems languages will always be a bit heavier syntactically because the syntax needs to support a wider range of types.

> These two types aren't typing the same data, though: in the latter case (in Rust) you've actually thrown away some of the data, and the type reflects that.

In Clojure the data is thrown away too, just at a slightly later point. The data is parsed then destructured then processed, and it is at the destructuring stage that irrelevant data is discarded and marked for GC.

Just as the Rust function avoids coupling by using a more narrow type, the Clojure function avoids coupling by using a more narrow binding.

> I think this isn't a great comparison: the usual vehicle for this thing in Rust is the trait, which manages to abstract over the representation statically without introducing a bunch of runtime machinery for it

But you have to create and implement those traits. I'm not saying this is impossible in Rust; just a lot more onerous because you don't get all the machinery Clojure has that makes it trivial.

> These are just predicates; I don't think they have much to do with this discussion?

This might be why we're talking past each other. When I say "type", I mean a set of possible values. I don't think this is an uncommon definition; the first sentence of the "data type" Wikipedia page pretty much defines it the same way.

Given this definition, you perhaps see why we can narrow a runtime type by composing it with a predicate.

My understanding is that you view types differently, as more of an interpretation of some sequence of bits, rather than a set of data values. Is that correct?

> As we get better at writing compilers we're increasingly seeing a move towards mechanisms like Rust's traits or C++'s concepts that allow the programmer to actually abstract over the representation of a type (as opposed to forcing a uniform representation as is common in dynamically typed languages).

I agree that's there's no reason in principle that you couldn't statically type it with a sufficiently advanced compiler.

> In Clojure the data is thrown away too, just at a slightly later point. The data is parsed then destructured then processed, and it is at the destructuring stage that irrelevant data is discarded and marked for GC.

Right; the difference is that by the destructing time you have lost the type information about that ‘discarded’ data, meaning you no longer know how to discard it. In Clojure this is not a problem, because the only values you're allowed to have are those that can be garbage collected, and (furthermore) garbage collected by calling a known function with nothing more than a universal set of GC metadata that is attached at a standard location on every value. But this is only possible because of a large restriction Clojure places on allowable values: the base ‘unitype’ of all Clojure values already contains all of this extra data, and any values for which it can't be attached or it doesn't make sense are simply forbidden.

> But you have to create and implement those traits. I'm not saying this is impossible in Rust; just a lot more onerous because you don't get all the machinery Clojure has that makes it trivial.

On this I absolutely agree. The only thing I'd be careful to distinguish is not simply that Rust doesn't have the machinery but that the machinery makes sense only given a fairly onerous set of additional assumptions about your values that Rust doesn't make (and that are usually held to be incompatible with a systems programming language, but IMO are also helpful for many — though not all — ‘userland’ programs).

> When I say "type", I mean a set of possible values. I don't think this is an uncommon definition; the first sentence of the "data type" Wikipedia page pretty much defines it the same way.

It's not an uncommon one, indeed, and while there are more sophisticated definitions (e.g. including equality) I don't think they need to be invoked for this discussion.

> Given this definition, you perhaps see why we can narrow a runtime type by composing it with a predicate.

No, you're right: I think I was thrown off by your assertion that these are uncoupled from the type system when we were talking originally of ‘conceptual’ types.

> My understanding is that you view types differently, as more of an interpretation of some sequence of bits, rather than a set of data values. Is that correct?

Quite the opposite. These words aren't entirely standardized so I set out my definitions above:

- I am using ‘value’ to refer to any object of discussion in a programming language that nominally exists at runtime; effectively, a value can be defined by the set of operations available on it - I am using ‘data’ to refer to values that have ‘data-like’ semantics, i.e. has ‘default’ move and discard mechanism like bitwise copy/overwrite or GC

> Right; the difference is that by the destructing time you have lost the type information about that ‘discarded’ data, meaning you no longer know how to discard it. In Clojure this is not a problem, because the only values you're allowed to have are those that can be garbage collected

Yes, this is very true. Languages like Rust can certainly manage memory and data structures more precisely. Not only is Clojure garbage collected, by virtue of being based on the JVM and being dynamically typed, but in idiomatic usage it encourages data to be built from a small number of basic types.

> On this I absolutely agree. The only thing I'd be careful to distinguish is not simply that Rust doesn't have the machinery but that the machinery makes sense only given a fairly onerous set of additional assumptions about your values that Rust doesn't make

I guess only I'd disagree here that the additional assumptions are "fairly onerous" - but that's just an opinion. You may be doing work that butts into those assumptions far more often than I do (which is almost never).

> No, you're right: I think I was thrown off by your assertion that these are uncoupled from the type system when we were talking originally of ‘conceptual’ types.

I don't think I expressed myself very clearly, and indeed Clojure's type system isn't very clear either, since it's bolted onto the JVM and so can't entirely do its own thing. So we have "types" which refer to Java classes/interfaces, and "schema" or "specs" which refer to Clojure's own runtime validation ideas.

> I guess only I'd disagree here that the additional assumptions are "fairly onerous" - but that's just an opinion. You may be doing work that butts into those assumptions far more often than I do (which is almost never).

I originally meant this mostly in the sense that it forbids a lot of values. But I do also think this is a bit of a Blub situation. There's a good class of pure data algorithms where basically everything you want to interact with is data, and if you're working in that space it's pretty great. But having spent some time in resource-typed languages, for any situation where I have to work with external resources or effects that can't be undone working in a language that doesn't have resource types and deterministic constructors really feels like a big step backwards. There are so many invariants you can't capture if you can't reason about value invalidation.

It certainly could be a blub situation, as aside from some toy applications, I don't have much experience with Rust. Obviously in Clojure, like most languages, you can restrict a resource to a particular scope - but it sounds like you mean something more than that.

Reply on news.ycombinator.com