The Disparate Musings of Pixie

One of my philosophies about exceptions is that they should be very specific. It's not enough to throw a generic Exception, or even a LibraryException, you should throw a specific exception for any event you want to handle separately. For instance, at a previous employer, our database layer would catch the generic exceptions (PDOException, in our case), inspect them, and re-throw as very specific exceptions (DuplicateEntry, CannotConnect, etc.).

Why

I've formed this opinion over the years, seeing things like the following code from a real application, which decides whether a nonced token has been seen before, and rejects the token if so:

try {
  // insert ...
  return false;
} catch (PDOException $e) {
  // 23000 is SQL error code for duplicate entry
  if (23000 == $e->errorInfo[0]) {
    return true;
  }
  throw $e;
}

That's ugly, and an annoying amount of boilerplate. It's much clearer if you can do something like:

try {
  // insert ...
  return false;
} catch (DuplicateEntry $e) {
  return true;
}

Particularly when you start getting into multiple scenarios, or start needing to support multiple database types. It can also be handy in the case of testing, since your mock can literally just throw the scenario you're trying to test.

If you're looking at that DuplicateEntry exception and wondering why it doesn't just check if the value exists first: race conditions, man.

Fun fact: that particular codebase has over half a dozen instances of catching DuplicateEntry for varying tables. Some of which are caught all the way in controllers to expose information to the user (hey, we already have an account with that e-mail address). The pain saved is very real.

Don't Wrap That Exception

This of course leads to another opinion about exception handling: libraries shouldn't wrap exceptions thrown by lower levels. At least, not in a generic, everything-is-now-a-LibraryException sense; clearly I'm okay with wrapping an exception to make it more specific.

There's a conflict of interest here: what exceptions get thrown is absolutely a part of your public API, and you want to be able to swap out the libraries you happen to be using underneath. Wrapping exceptions enables you to do that, whereas passing them through does not.

But here's the thing: in the age of Dependency Injection, your caller knows more about the exceptions thrown by the implementing object than you do. So if you're a Frobnicator, and your caller gives you a Fetcher implementation, you're actually hiding information the caller expects by catching exceptions thrown by the Fetcher and wrapping them in your FrobnicatorException. That's actively hostile to your calling code, because now to catch the exceptions they were expecting they must first catch your exception, unwrap it, then rethrow.

I guess another way of saying that is any layer in the middle should refine exceptions, not conglomerate them. Please, be mindful of which you're doing.

• The Basic Problem
• The Method
• Inheritance Is… Uh… Missing?
• Clojure
• Go
• Subjective Comparisons
• Recommendation
• Available For Work

Before my now-previous employer folded, we were being courted by a buyer (boy, was that time period an emotional roller-coaster!). This would have changed our development direction somewhat, and so at the behest of my boss I spent a couple of weeks exploring language options for our upcoming greenfield projects.

I was asked to explore Clojure and Go, and compare them to our existing language, PHP. I decided to evaluate the languages by porting a fairly stand-alone portion of our existing app into both Go and Clojure, and comparing it to the PHP version.

The Basic Problem

That relatively stand-alone portion of our app was our recipe importer. It would allow users to import recipes from external sites, for use within our system.

There are a number of standards for recipe formats: the hRecipe microformat, schema.org's Recipe format, RDF, data-vocabulary.org's Recipe format, and probably a few others. In addition, most of the sites we explicitly supported either fail to mark up the entire recipe, interpreted the spec they purported to follow in strange ways, or merely referenced a recipe from another site, requiring us to follow the link chain to a parsable source.

In other words: it's a real (non-toy) problem, it's sufficiently complicated to get a reasonable feel for a language, and it's simple enough it can be ported within a relatively short timeframe.

The Method

I spent a week in each of Clojure and Go. This was not enough time for a full feature-parity port with the PHP version, but sufficiently long to present most of the complexity of the problem (inheritance, parser registration, and so forth).

Inheritance Is… Uh… Missing?

Neither Go nor Clojure support "classical" OO inheritance. It may be my brain is broken through years of shoehorning things into inheritance hierarchies, but it seems to me this particular problem happens to be a pretty good fit for classic OO. "Site X uses format Y, with these overrides."

So the biggest stumbling block in both languages was simply coming up with a way to model the problem. To be perfectly honest, the solution I ended up in both languages looks pretty much like classic OO shoehorned into whatever the language does support (embedded structs in Go, multimethods in Clojure), and would almost certainly not be considered idiomatic by people with more experience in each language. (But then, neither would my PHP, and I've been doing that for a very long time.)

Clojure

As a lisper bent on proving that a lisp was a viable option, I started with Clojure.

I spent approximately three days figuring out how to even model the problem without classic OO inheritance, and the following two building three generic parsers and two site-specific ones.

It was one of the most enjoyable weeks of my tenure. Clojure is as flexible as you'd expect a Lisp to be, the code was compact, reasonably fast, and I was reminded of how much I'd been missing interactive development.

When I presented it alongside the PHP it was a port of, reactions were mixed. Some developers remarked that they had a very difficult time understanding it, and that it would likely prevent the developers who worked on other codebases (e.g., our mobile developers) from swooping in to make a quick fix. My boss, who had apparently spent a significant amount of time attempting to do something useful in Clojure without much success, remarked that the comparison made Clojure much easier to understand for him than it had been previously.

Go

Next I tried Go. Having already faced a lack of inheritance, and presented with Go's much smaller field of options to choose from, I was able to jump right in, and hit parity with the Clojure version in short order.

I did not particularly enjoy programming in Go. While Go is strongly opinionated in surprisingly good ways, the edit, compile, debug cycle is obnoxious and slow (goauto may have helped reduce this pain), the code ends up verbose and repetitive, and there are little to no facilities for abstracting out patterns.

The Go code received less objection when presented. The lines were significantly longer than either the PHP or Clojure versions, and vertically it was very slightly less dense than the PHP version due to all the error checks.

Subjective Comparisons

Code Feel

PHP	Clojure	Go

Is it any wonder I prefer lisps? So much meatier.

Learning Curve / Programmer Effectiveness

Go was the easiest to pick up, but there isn't really anywhere left to go once you do. Clojure is definitely harder to become proficient with, but you can leverage that into increased programmer efficiency later. PHP has so many pitfalls that anyone starting out is likely to end up negatively productive at first, writing a horribly insecure buggy mess.

Developer Emotions

Clojure inspired a mixture of enthusiasm and curiosity, with strong concerns for the practicality of finding and training developers in a language which is so different from mainstream ones. Go had an enthusiastic champion in my boss, but was largely met with ennuitic "meh"s by other developers (golang is basically boring-by-design). PHP was loved by none and hated by some, with the only argument for it being "we've already got a bunch of code written in PHP".

Recommendation

I liked Clojure. It's fun to write in, it feels very productive, and the power lisps provide to build your own abstractions is unparalleled. But as varied as the quality of code was in our PHP, providing room for even more variability is a hard case to argue. While I would have loved to recommend Clojure, asking a team of devs who think in terms of C++ to make the transition was more than I could have sold.

Go, on the other hand, solves a lot of problems that we had with PHP. Stuff the wrong type? Go would have caught that. Inconsistently-formatted code? gofmt would have fixed that. Bugs because PHP did something unexpected unless you thoroughly read and retained every aspect of the documentation? Go's house is not built from bundles of thorns.

So while Go may not be the language for me, it would have been the right call for our team.

Available For Work

As you may have guessed by the mention of my previous employer folding, I am presently available for work. Feel free to contact me if you have something you think I may be interested in.