Thursday, March 15, 2012

Unifying the Agile and Spiral models

If, like me, you have never heard of the Spiral model of development, here is a summary: it is a bit like Agile, except that all the code is thrown away at the end of the iteration. Also, the scope increases each time, so the iterations grow longer and longer.

Yes, I know this sounds like a terrible idea, but that's only when compared to Agile as a reference point. From that angle, Waterfall is the model where you only get one, very big, very long iteration, and hope to get it right the first time. When put in the context in which it was invented, doesn't the Spiral idea of starting with a smaller version of the project seem like a great improvement?

Not throwing away the code all the time sounds like an even better improvement, but when I first heard of the Spiral model (just a few hours ago), I thought the opposite. Throwing code away! What a novel and intriguing idea! Should I try? Maybe? Being in an Agile team doesn't guarantee good design, after all. And changing a bad design is sometimes a large endeavor. With Substrate, for example, I haven't been able to implement any story at all this iteration, because I was busy refactoring the entire application to use MVC.

Instead of being forced to either drop all the existing code or keep all of it, I suggest we should be allowed to do either. And those two possibilities are just two ends of a continuum: using version control, we can easily revert to a number of intermediate steps between the two extremes, whichever is most convenient.

An Example

Let's start with an empty project, in the middle of Design Space, the space of all possible designs.

We discuss the product idea, draw a bit of UML, and end up with a particular design (a particular point in Design Space).

A few iteration go by, moving the project toward the goal established in the design.

While implementing new features, suppose we come up with a new design for the project. Could happen! Otherwise, we would still be doing big design up front and it would work great. The goal is moved a bit to the left.

Since we are using iterations, we adapt to this unexpected change without a sweat. We just start writing the new code with the new code in mind!

But what if the new design was very, very different from the first?

In that case, starting from scratch is less work.

But this is not an all or nothing decision. Yes, you can start from scratch, but you can also start from any other past point. Just pick the one which represents the least amount of work!

Future Work

Now, as I said, I only heard about the Spiral model today, so there are still many rough corners in my generalization of it and Agile. For one thing, going back in time means removing some features, and if those features have shipped already, maybe you don't want to remove them. Feature Space is not necessarily aligned with Design Space.

The other big problem is that history is linear, while Design Space is multidimensional. When your goal changes its position, it probably won't move across all the dimensions, only some of them. So while it is true that some of your code can be thrown away, most of it probably shouldn't; in order to reach the commit you want to restart from, you will have to throw away some commits which are still good, just because they were submitted after the restarting commit. I recommend using a history-rewriting tool such as git, but when I tried, solving all the merge conflicts was still a lot of work.

Ideally, each new feature could be implemented on a bare-bones version of the application, containing only the elements on which the new feature builds. In that setting, implementing the new feature is much easier, because there are no unexpected interactions with the other features, the codebase is small enough to be understood all at once, and the debug-edit-compile cycle is fast. Theoretically, using feature branches, maybe this ideal could even be reached!

The problem, of course, is that the unexpected interactions bite back once we try to merge all of the feature branches into a complete product.

I think I am going to try anyway. If I can afford wasting an entire iteration redesigning my code, maybe I can also waste one redesigning my process?

No comments: