Comments on the New R OOP System, R7

Object-Oriented Programming (OOP) is more than just a programming style; it’s a philosophy. R has offered various forms of OOP, starting with S3, then (among others) S4, reference classes, and R6, and now R7. The latter has been under development by a team broadly drawn from the R community leadership, not only the “directors” of R development, the R Core Team, but also the prominent R services firm RStudio and so on.

I’ll start this report with a summary, followed by details (definition of OOP, my “safety” concerns etc.). The reader need not have an OOP background for this material; an overview will be given here (though I dare say some readers who have this background may learn something too).

This will not be a tutorial on how to use R7, nor an evaluation of its specific features. Instead, I’ll first discuss the goals of the S3 and S4 OOP systems, which R7 replaces, especially in terms of whether OOP is the best way to meet those goals. These comments then apply to R7 as well.

SUMMARY

Simply put, R7 does a very nice job of implementing something I’ve never liked very much. I do like two of the main OOP tenets, encapsulation and polymorphism, but S3 offers those and it’s good enough for me. And though I agree in principle with another point of OOP, “safety,” I fear that it often results in a net LOSS of safety. . R7 does a good job of combining S3 and S4 (3 + 4 = 7), but my concerns about complexity and a net loss in safety regarding S4 remain in full.

OOP OVERVIEW

The first OOP language in wide use was C++, an extension of C that was originally called C with Classes. The first widely used language designed to be OOP “from the ground up” was Python. R’s OOP offerings have been limited.

Encapsulation:

This simply means organizing several related variables into one convenient package. R’s list structure has always done that. S3 classes then tack on a class name as attribute.

Polymorphism:

The term here, meaning “many forms,” simply means that the same function will take different actions when it is applied to different kinds of objects.

For example, consider a sorting operation. We would like this function to do a numeric sort if it is a applied to a vector of (real) numbers, but do an alphabetical sort on character vectors. Meanwhile, we would like to use the same function name, say ‘sort’.

S3 accomplishes this via generic functions. Even beginning R users have likely made calls to generic functions without knowing it. For instance, consider the (seemingly) ordinary plot() function. Say we call this function on a vector x; a graph of x will be displayed. But if we call lm() on some data, then call plot() on the output lmout R will display some graphs depicting that output:

mtc <- mtcars
plot(mtc$mpg) # plots mpg[i] against i
lmout <- lm(mpg ~ .,data=mtc)
plot(lmout)  # plots several graphs, e.g. residuals

The “magic” behind this is dispatch. The R interpreter will route a nominal call to plot() to a class-specific function. In the lm() example, for instance, lm() returns an S3 object of class ‘lm’, so the call plot(lmout) will actually be passed on to another function, plot.lmout().

Other well-known generics are print(), summary(), predict() and coef().

Note that the fact that R and Python are not strongly-typed languages made polymorphism easy to implement. C++ on the other hand is strongly-typed, and the programmer will likely need to use templates, very painful.

By the way, I always tell beginning and intermediate R users that a good way to learn about functions written by others (including in R itself) is to run the function through R’s debug() function. In our case here, they may find it instructive to run debug(plot) and then plot(lmout) to see dispatch up close.

Inheritance:

Say the domain is pets. We might have dogs named Norm, Norma, Frank and Hadley, cats named JJ, Joe, Yihui and Susan, and more anticipated in the future.

To keep track of them, we might construct a class ‘pets’, with fields for name and birthdate. But we could then also construct subclasses ‘dogs’ and ‘cats’. Each subclass would have all the fields of the top class, plus others specific to dogs or cats. We might then also construct a sub-subclass, ‘gender.’

“Safety”:

Say you have a generic function defined for the class, with two numeric arguments, returning TRUE if the first is less than the second:

f <- function(x,y) x < y

But you accidentally call the function with two character strings as arguments. This should produce an error, but won’t

In a mission-critical setting, this could be costly. If the app processes incoming sales orders, say, there would be downtime while restarting the app, possibly lost orders, etc.

If you are worried about this, you could add error-checking code, e.g.

> f
function(x,y) {
   if (!is.numeric(x) || !is.numeric(y))
      stop('non-numeric arguments')
   x < y
}

More sophisticated OOP systems such as S4 can catch such errors for you. There is no free lunch, though–the machinery to even set up your function becomes more complex and then you still have to tell S4 that x and y above must be numeric–but arguably S4 is cleaner-looking than having a stop() call etc.

Consider another type of calamity: As noted, S3 objects are R lists. Say one of the list elements has the name partNumber, but later in assigning a new value to that element, you misspell at as partnumber:

myS3object <- partnumber

Here we would seem to have no function within which to check for misspelling etc. Thus S4 or some other “safe” OOP system would seem to be a must–unless we create functions to read or write elements of our object. And it turns out that that is exactly what OOP people advocate anyway (e.g. even in S4 etc.), in the form of getters and setters.

In the above example, for instance, say we have a class ‘Orders’, one of whose fields is partNumber. In S3, the getter in the above example would be named partNumber, and for a particular sales order thisOrder, one would fetch the part number via

get_partNumber(thisOrder)

rather than the more direct way of accessing an R list:

pn <- thisOrder$partNumber

The reader may think it’s silly to write special functions for list read and write, and many would agree. But the OOP philosophy is that we don’t touch objects directly, and instead have functions to act as intermediaries. At any rate, we could place our error-checking code in the getters and setters. (Although there still would be no way under S3 to prevent direct access.)

ANALYSIS

I use OOP rather sparingly in R, S3 in my own code, and S4, reference classes or R6 when needed for a package that I obtain from CRAN (e.g. ebimage for S4), In Python, I use OOP again for library access, e.g. threading, and to some degree, just for fun, as I like Python’s class structure.

But mostly, I have never been a fan of OOP. In particular, I never have been impressed by the “safety” argument. Here’s why:

Safety vs. Complexity

Of course, OOP does not do anything to prevent code logic errors, which are far more prevalent than, say, misspellings. And, most important:

• There is a direct relation between safety and code complexity.
• There is a direct relation between code logic errors and code complexity.

One of my favorite R people is John Chambers, “Father of the S Language” and thus the “Grandfather of R.” In his book, Software for Data Analysis, p.335, he warns that “Defining [an S4] class is a more serious piece of programming …than in previous chapters…[even though] the number of lines is not large…” He warns that things are even more difficult for the user of a class than it was for the author in designing it, with “advance contemplation” of what problems users may encounter. And, “You may want to try several different versions [of the class] before committing to one.”

In other words, safety in terms of misspellings etc. comes at possibly major expense in logic errors. There is no avoiding this.

There Are Other Ways to Achieve Safety:

As noted above, we do have alternatives to OOP in this regard, in the form of inserting our own error-checking code. (Note too that error-checking may be important in the middle of your code, using stopifnot().) Indeed, this can be superior to using OOP, as one has much more flexibility, allowing for more sophisticated checks.

Why the Push for R7 Now?

Very few of the most prominent developers of R packages use S4 as of now. One must conclude either that there is not a general urgency for safety and/or authors find that safety is more easily and effectively achieved through alternative means, per the above discussion.

As to encapsulation and inheritance, S3 already does a reasonably good job there. Why, then, push for R7?

The impetus seems to be a desire to modernize/professionalize R, moving closer to status as a general-purpose language. Arguably, OOP had been a weak point of R in that sense, and now R can hold its head high in the community of languages.

That’s great, but as usual I am concerned about the impact on the teaching of R to learners without prior programming experience. I’ve been a major critic of the tidyverse in that regard, as Tidy emphasizes “modern” functional programming/loop avoidance to students who barely know what a function is. Will R beginners be taught R7? That would be a big mistake, and I hope those who tend to be enthralled with The New, New Thing resist such a temptation.

Me, well as mentioned, I’m not much of an OOP fan, and don’t anticipate using R7. But the development team has done a bang-up job in creating R7, and for those who feel the need for a strong OOP paradigm, I strongly recommend it.