The rule of thumb that I try to live by is to not attempt to abstract something until I'm doing it for the third time. Why do I do that?

Well the first time I guarantee that I don't know how similar examples will vary and won't be able to pick a good abstraction. So attempting to abstract at that point will result in more code that is less to the point (and therefore less clear), but I'm very unlikely to find an abstraction that is useful for my next time. Which means that my abstraction has lots of real costs and no real benefits.

The second time I now have a point of comparison. I could abstract and it might work. But in my experience I still don't know enough for the cost/benefit to work out. Also at this point I am consciously fighting the second system effect.

The third time I know enough about the problem and have enough experience that I am confident that I can come up with the right abstractions. Abstractions that pay off enough in future maintenance to justify the cognitive overhead of having to learn them, and the initial effort of writing them.

Moving back to this specific example, you are perfectly right that if I had several variations of HTML format that I had to deal with, an OO layer would be a great way to go. However if I don't, adding an OO layer would add code and make it more difficult to see what is going on. It would also make other kinds of changes more difficult to make.

Let's make this concrete. In the several years since I wrote this I have never once wanted to have several similar markup languages that just varied in the syntax of the tag. (Which is what your proposed OO layer would give me.) However I have encountered other needs, such as emoticons and the need to have tags that are only allowed to open when another tag (eg a table tag) is open. Those changes would be complicated by your proposed OO interface since they cut across that interface and require changes to both sides.

I could do that (no OOP), but that would be premature under-abstraction :-)

No seriously, you make a great point about your having used this module for years and not needing that particular abstraction. And I know you have loads and loads of experience. But let me give a counter story.

One of my first jobs out of school was supervising a team trying to put together a consolidated data model and data dictionary for about 300 coorporate information systems. I learned a lot more than I ever wanted to know about the vagaries of abstraction and language. And most of all I came to the conclusion that there will never be one way to do things and that the best one could do to control confusion was to create good frameworks for doing the same thing many ways. (Maybe that is why I like Perl so much?).

The main reason for OOPing up an implementation is that it provides a framework for TIMTOWTDI. When software is released publicly or across a corporation, it is virtually guarenteed that needs will change across user groups and time. Even during development, customers rarely can articulate or correctly prioritize the corner cases until they've seen a few demos, and by that time, an awful lot of code may be written.

However, just to be clear, I was not suggesting a full scale OOP evangelist style of OOPing up. Most OOP implementations go way overboard creating a mess of interconnected objects that would make a circuit board engineer's eyes cross. They assume that all objects must be mutable. That danger lurks wherever data is accessible to outsiders, etc, etc. Yada, yada, yada..

Rather I had in mind something much simpler. Something that would allow for the possibility that an OOPish need will arise without destroying the overall clarity of your code. How would I do that?

As a first step, I would make changes that do nothing more than make it possible for consumers of the parser to call your parser using an OOP syntax. Internally, aside from the ever-present $self, the object wouldn't really be OOP at all. This would involve little more than

• adding a dummy new(...) method that did nothing more than bless and return $self.
• adding $self as a parameter to all of your methods
• making the wrapper around the dispatcher a method and adding $self as a parameter to the dispatch call

I'm not even sure I would bother giving new(...) parameters (other than $sClass) or even moving the configuration data into a hash. And though I would likely add getter methods for configuration data, I certainly wouldn't bother adding mutators or fancy methods to hide data.

My reasons would be the same as yours - we don't know enough about the way it should be abstracted (yet). Maybe the users don't need to configure the object, they only want to override token processing. Maybe they want to configure, but are ok with doing it only when the object is constructed. Maybe they need a fancy way to add new syntax elements or may be they don't. Maybe they are happy to rely on training and social controls to keep data private and maybe they aren't. We just don't know. All of these can be added later without disrupting existing code. So I would keep it simple.

Ah...but if the changes are so little - why bother? If you really need to, you can do those kind of changes later on. For now they are just fluff...

Why? Because for very little cost, it reduces (not eliminates) the risk that we will have to go out and review 100K or more lines to find all the places we need to change from a function interface to an OOP interface. It reduces the risk that we will need to release a backwards incompatible product to clients. It reduces the risk that we will have to redesign the class to support both a procedural and OOP interface (which will definitely muck up the clarity of your code). It reduces the risk that we have to choose an awkward ordering of parameters just so we can guess which interface is being used. For very little cost and virtually no premature abstraction we create an interface that is more likely to grow in a managable way.

So yes, I agree with minimalism, but the minimalism should be in the design of the object, not in the avoidance of objects.

Best, beth

Sorry about taking a while to respond, I have a lot going on right now.

First let me comment that there is a bit of a bait and switch here. Not long ago you were suggesting using objects to allow polymorphism on tag recognition. I pointed out why that wasn't a good place to put an abstraction layer. Now you're suggesting using objects to hide the overall engine. That's a change in design.

Secondly it seems to me that you're having trouble recognizing abstraction that doesn't come with an OO bow tie on top. The style of programming used in that example offers equivalent if not greater opportunities for abstraction than OO does. Admittedly you abstract in different ways in different places. You have more flexibility in how to abstract at a cost of not providing a standardized framework for others to recognize. But under-abstraction is not a problem with that style.

What flexibility would you hope to gain in this specific situation by wrapping the interface with an object? Make that clear and I'll show you how to do the same thing without objects. Please trust me when I say that I am not taking this position out of any ignorance of how useful OO can be. Quite the reverse in fact. I believe I can quite honestly say that I understand OO much better than most programmers and I understand a number of other programming techniques as well. Therefore I am in a better position than most to decide when OO is the right abstraction technique to look for.

For more on the general limitations of OO as an abstraction pattern, The world is not object oriented may be of interest. But, lest I sound entirely negative about OO, I should hasten to add that there are cases when I think it is appropriate and useful. One important example being when you have a series of different decisions that need to dispatch in a coordinated fashion. In which case you can do the dispatches by making method calls on objects.

For the record I am not alone in my views. The ever quotable Paul Graham has even stronger views. They were quite quotably stated in his essay The Hundred Year Language:

I don't predict the demise of object-oriented programming, by the way. Though I don't think it has much to offer good programmers, except in certain specialized domains, it is irresistible to large organizations. Object-oriented programming offers a sustainable way to write spaghetti code. It lets you accrete programs as a series of patches. Large organizations always tend to develop software this way, and I expect this to be as true in a hundred years as it is today.

Now it is easy to dismiss Paul as a gadfly, or to assume that he doesn't know much about programming. But before doing so I really, really, really recommend reading his On Lisp cover to cover with comprehension. Once you see how he tackles a variety of problems using different techniques, you may take more seriously his contention that OO's toolkit is not the only, and is often not the best, way to approach problems. (Incidentally in one example he implements an admittedly basic OO system in 8 lines of code. He then shows how to add features to it.) While you are at it you may wish to read Structure and Interpretation of Computer Programs for a related but somewhat different set of approaches to abstraction. (That book is intentionally more introductory.) If reading Lisp is a block for you, Higher-Order Perl covers related ideas in Perl.

Those are, of course, all oriented towards a functional style. But there are perfectly good abstraction techniques for straight imperative code. The classic Code Complete 2 has as good an explanation of the basic ones as anything else does. Be warned that it is a weighty tome, but in my opinion virtually everything in that book is something that every competent programmer should know inside out and backwards. Even if you know 80% of what is in that book, the remaining 20% makes the purchase and reading worthwhile IMO.