Vapour: A typed superset of the R programming language
vapour.run
vapour.run
1) How does this work with function parameters that are intended to be captured unevaluated with substitute()? Do you type the input as "any" and document separately that the parameter is kept "unevaluated" as a symbol/name or call?
2) How does this work with existing untyped R code? Does it at least include types for the standard library (or some subset thereof?)
3) Is there any type inference, or does it require explicit type annotation everywhere?
4) How do you propose to handle NA (which can appear "within" any typed vector)? Does the compiler support refinement types? If not, how does checking for and preventing nullability work, when checking for NA values requires a runtime check?
5) How do data frames work? Are they typed like structs?
6) Which object systems does it support, if any? S3, S4, Reference Classes, or the 3rd-party R6?
As much as I like static types, I feel like R is maybe the language where I need or want them the _least_. How often do you really run into a situation where you pass a character vector to a function that requires a numeric vector and it crashes your program?
99% of the time what you really want is known-valid data frames for data processing, and statically-sized arrays for math stuff.
I really disagree with this.
I think one of the whole reason there is a whole Tidyverse ecosystem that the behavior of (some) R code is unintuitive in a way that adding typing would absolutely improve.
It seems like you're deeply familiar with the R ecosystem, but as a user what I want is a safe subset of R that I can use.
> How often do you really run into a situation where you pass a character vector to a function that requires a numeric vector and it crashes your program?
In R the more likely situation is that you pass in the wrong typed thing and it silently continues with very unexpected values being passed, causing trouble or errors much later in the program. Which is very much a problem that typing helps with.
Can you name one practical example of this happening as a result of passing in a vector of the wrong class()/mode()? Not a data frame with the wrong column types, but an actual standalone vector. Can you name an example in the Tidyverse ecosystem that specifically improves on the type-safety ("class/mode-safety") of the standard library? I can't, but maybe that's just because it's been too long since I did anything serious with the language.
I can definitely think of complicated interfaces where you can silently get strange results by passing in the wrong thing. sweep() and apply() are obvious examples, where you can accidentally swap the argument order and silently get a nonsensical result. But that's a matter of array shape, not of type. Try passing an argument of the wrong type (again, where "type" in this case means the class or mode of the vector) to sweep() or apply(), and watch what happens: you get an error message informing you that you passed a value of the wrong type. At worst, you get an obtuse error message informing you that you passed a value of the wrong type, but bubbled up from some internal code. But you get an error all the same.
R is actually very strongly-typed, and abstracts over some details that would otherwise cut into that type-strength. For example, R doesn't have the Numpy problem of exposing different physical storage sizes for integers and floats! It just has abstract numeric arrays, backed by whatever the hell storage type the R language implementers decided to back them with, with no opportunity for the user to accidentally mix things up and lose precision, overflow, or crash on contact with some pre-compiled Numba function.
I maintain that a much, much more pertinent problem is that array shape is not part of the type system, and moreover that a lot of R code is (by design) highly polymorphic with respect to array shape, precisely because there is no such thing as a "scalar" number or string but we still want to let people use numbers and strings in scalar-like fashion.
NULL I think falls into this category as well. NULL in R is a bit like nil in Lua or undefined in Javascript, in that it has a kind of dual function as a "value that is not any other value" and a "non-value that cannot be inserted into a collection, instead deleting whatever was previously there". But when is the last time someone got a NULL and a numeric vector mixed up, and wasn't able to figure out what happened? Is all the complexity of a static compiler really necessary to catch that relatively rare mistake?
Maybe the one exception here is the factor class. But there's no mention of factors here, and (as with array shape), validating factor levels is probably more important as validating that the thing is a factor in the first place, as opposed to character.
The NA checking proposed is another story. Now that would be useful, but so would checking things like min/max ranges, the presence of certain columns in a data frame, etc. For example Python has its data frame input validation framework Pandera that offers at least some of these guarantees at the type level.
As for classes, I noticed that they implement what looks like a nice concise syntax for creating S3 class objects with structure(). That's great, but you could have just written a helper library to do that.
Anyway, here's a project where someone designed a whole language and wrote a compiler for it, and I'm just one cantankerous former R user doubting whether that project is ever going to be useful. If this is just a hobby project to scratch someone's itch: ignore me. But if this is intended to be a serious thing for serious use in production, then I'd encourage the creators to reconsider how they portray their value proposition, and to maybe reconsider whether the goal of their project aligns with the needs and desires of actual R users in industry, of whom there are still many, but definitely not as many as there used to be.
let x: int = 1
let x: int = (1, 2, 3)
"R" is the only programming language I know and I can't find a job that uses a R because job search engines don't allow you to sort by skill
"R language" is the closest substitute on linkedin but the results are still a jumbled mess of jobs, some looking moreso for other skills (SQL/Python)
I know R-heavy jobs exist but finding them on LinkedIn is virtually impossible
E.g. tidyverse or dplyr is like 20-40 jobs. ggplot is 88. There's definitely way more than 100+ companies looking for R-heavy users.
Looking for a similar job where my desire/interest to spend all day in Rstudio is a value add to a business
For example, I used R (data.table) when I was a solo data scientist working on a consulting project where I needed to work with a dataset on the order of a few billion rows. I had nobody around to constrain my choice of tools, so I went with whatever felt convenient, familiar, and ergonomic for getting the job done.
Today, I am on a team of 5 other people, none of which know a lick of R, and my code needs to run in production pipelines that need to at least in theory be debuggable, auditable, fixable, etc. by people other than me. Therefore I use Python, because we are a Python team and that's the language that we use, end of story. (Python also happens to be a good choice on our team for other reasons, but that's not the point here).
Maybe the best industry where you are likely to find people doing "production" work in R is some form of insurance. But even back in 2017-2020, things were shifting towards Python at the one P&C company I worked for.
Insurance is also still using a lot of R. Actuaries I know still use it, and they talk of Python, but I don’t see anyone actually moving to using Python.
I know this is totally bike shedding, semantics, vi vs Emacs, BigEndian vs LittleEndian and it's too late now to affect anything, but to me using a colon after the variable is just wrong!
let x : int = 1
func add(x: int, y: int): int { return x + y }
I see that and it looks like int = 1 and the function's return type is totally lost.
This seems completely backwards to me. Maybe I'm just used to the way C did it, but the variable modifiers should come first.
let int x = 1
func int add(int x, int y) { return x + y }
Why we reversed it and added in the colon just doesn't make much sense to me.
To be clear, I love R, it excels in prototyping but I have seen too many real world struggles of folks trying to move to prod that I would say save it for EDA projects and one time analyses.
Do you think the culture of the package ecosystem could possibly change in the future?
IME with both Python and JS/TS, it helps a lot (which is different than completely solving the problem), for reasons which should generalize to other typing add-ons/supersets for untyped languages. Typing your code forces validations at the boundaries, which obviously doesn't stop upstream sources from messing with formats but it does mean that you are much more likely to catch it at the boundary rather than having weird breakages deep in your code that you have to trace back to bad upstream data.
eg. `type DateString = ${number}/${number}/${number}`
A super naïve check for using "/" instead of "-" as the separator character for a date formatted as a string. If a date is provided with some other separator character it will throw an error. If my function takes a DateString the string must be formatted correctly to pass the type check. Obviously this isn't enough (YYYY/MM/DD is different than DD/MM/YYYY) but the intention was to show a way to enforce something via types rather than validating a string to check that your have a DateString you can simply enforce that you have one.
The idea is that checking should be the only way of making a value of the type. That prevents you from forgetting to check when you turn some broader type (say, string) into the more narrow one (date, in this case).
Yeah, of course you can cheat the typechecking in the code at the boundary in several ways, or convert from wire format to internal types in a way which plugs in type-valid defaults for bad data rather than erroring, or just use too-broad internal types to start with (you can have "stringly-typed code"), and fail to help the problems. But if you use the types that make sense internally for what the code is doing, than conversion including validation at the boundary becomes the path of least resistance in most cases. "Forces" is not strictly true, but my experience is that adding types does create a strong push for boundary validation.
reticulate works for going in the other direction: https://rstudio.github.io/reticulate/
With the good interoperability these days, let's stop rewriting functionality in other languages. If the interoperability is no good, work on fixing that, please.
Vapour is an interesting choice. Hope it’s in name only :)
Good luck to the authors of this. I believe it solves an important problem for R package authors and others wanting to write bigger programs. It's hard to argue with the benefits of static typing for this type of work.
Bugs are normal software development.
Changing syntax and breaking things make work for everyone else for the convenience of developers. Reliability is what makes a tool a tool.
How else might one explore a new language (vapour) in the open among interested like-minded developers seeking to iterate on a tool found lacking (R)?
Changing and iterating things makes.
What part of this is giving any sense of stability? It's clearly an experimental language, so I find it hard to understand why you are discussing stability and compatibility at all.