Thursday, November 5, 2015

The mystics are coming out

Discussion on the Scala collections revamp is starting to get mystical. It really bugs me: good language design makes a huge difference, but it's hard to see unless you actually deal with thousands or more developers on millions or more lines of code. Casual observers of language design discussions don't see it themselves, so they don't realize what these problems look like to the other people in the discussion. So they think everyone is just goofing around and throwing out random ideas just because they are possible.

I started to follow up on the issue itself, but I fear making the problem worse. So I'll post a couple of rebuttals here. I don't really think the people actually involved in the redesign will be distracted by the mystics, anyway. They are like gold-sellers in MMOs or beggars in San Francisco; unless you are specifically trying to engage with them, you just learn to tune them out. Well, okay, they are like really nerdy gold sellers who like to talk about higher math. Okay I better just drop this attempt at an analogy.

First, Python's indexing operator has a lot of practical problems. Here are a few concrete examples: . Thinking backward from those puzzlers, I think the root of the problem is that the [a:b] slicing operator means something different depending on whether a and b are positive, and whether a is smaller or larger than b. This is foundational syntax used everywhere, and if you want code to be readable, people need to know whether it's doing a forward or reverse slice without having to do mental data flow analysis on the parameters. Java avoids this trap, and Scala should, too. The sublist operation should only work on non-negative arguments, and only when the first argument is smaller than the second.

The other thing I will say is that exceptions, null, and -1 are also all fine, when used by a tasteful library designer. We tried at Semmle to get people using more Options, and we found that it harmed our reputation as a quality lint tool. I can only speak publicly about open-source code, but to give an example, Apache Spark has over a thousand occurrences where they use null but, with only local changes, they could use Option instead. It's too many. It means that the basic premise of the programming advice has some sort of problem.

As one stab at it--though it's really a huge topic--you have to think about what you want the callers to do to defend against a missing value. If Scala's basic get/apply methods starts returning an Option, then people will just litter their code with calls to .get, so the net result is that the code is more bloated but otherwise behaves exactly the same. Even in pure math, people will write notation like f'(x) as the derivative of f, but you know, derivative isn't always defined. So should smart mathematicians instead write get(f')(x)? Or (f') match { ... }?

That's my try, but you don't even have to understand this if you are willing to ape mature libraries in areas that they are working okay. It's not a practical problem in Java that the various .get() methods return null or throw exceptions; even if you say it's not perfect, it's certainly not bad. It is, however, a very big problem that Java collections are overly mutable. For example, see this Stack Overflow question: Scala will be more attractive to more developers if it focuses on these widely recognized pain points. Which is why the mystics drive me crazy--if they get their way they will find that their playground is gradually becoming a ghost town, but only after it's too late.

No comments: