Wednesday, April 19, 2006

Agile and Estimation

Over on Tyner Blain, Scott states:
"There is nothing that prevents a waterfall project from reviewing estimates throughout the course of the project."
Recall that agile processes use iteration and frequent releases to adjust to changing requirements and other discoveries during product development. Waterfall processes, on the other hand, schedule along a critical path that culminates in a single release.

While Scott endorses agile practices, he does not believe they lead to improved estimation. I don't agree with Scott on this last point. Here's why.

Reliable estimates hinge upon a holistic understanding of requirements, design, implementation, and testing.

With waterfall, you certainly can review and adjust estimates halfway through the process, but you will not be able to incorporate the full feedback effects. For example, you could be halfway through requirements and design and re-estimate the project, but you wouldn't have the benefit of the discoveries that inevitably occur during implementation and testing.

With incremental delivery (agile), you iterate on all of these phases early and often. Consequently, you quickly learn the impact of, for instance, testing on requirements. You are thus better informed when you review and adjust your estimates for the project.

6 comments:

Scott Sehlhorst said...
This comment has been removed by a blog administrator.
Scott Sehlhorst said...

Hey Roger, thanks for joining in on the conversation.

I think this is a tough one to grasp, but improved estimation doesn't come from incremental releases. It does come from getting insight into the development process along the way. This can happen for both waterfall and incremental delivery projects.

In my post I added the statement "The benefit of waterfall process estimation is that we can confidently predict how long it will take us to implement the wrong requirements."

The real benefit of learning that (differentially) comes from incremental releases is in fixing bad (or less than great) requirements with feedback from users and stakeholders.

The feedback that comes from those end users doesn't change how long we estimate that it would take to implement X, it changes X to Y.

While this absolutely gives us a better product, it doesn't give us better estimates.

Scott

Roger L. Cauvin said...

Scott, I think you're neglecting two important factors.

First, the learning that comes from incremental releases does not merely stem from feedback from stakeholders. We make requirements discoveries during design, implementation, and testing. We make design discoveries during implementation and testing. And so forth.

Second, you can generally make better informed estimate adjustments in an iterative process. You simply have more information to make the estimate due to having partially confronted the requirements, design, implementation, and testing risks on an iterative basis.

Better estimates are not merely a function of experience. They depend on information. Iterative processes bring to light the information with the most impact early in the cycle.

Roger L. Cauvin said...

Just to be clear. You wrote:

"The feedback that comes from those end users doesn't change how long we estimate that it would take to implement X, it changes X to Y."

That's true, but the information we get from iterating isn't just what we will implement, but how long it will take to implement it.

For example, without changing the requirements in any way, we may find by iterating that the architecture is much more complex than originally anticipated, and that therefore the level of effort required to implement the same requirements will be higher than our original estimate.

Scott Sehlhorst said...

I completely agree that design, implementation and testing are the sources of information that make our estimations better. My point is that we can do all of that with a waterfall process too.

I dev-managed a project that (for reasons outside my control) had to be delivered "at the end" (a waterfall project). I chose to run that project in a continuous release-ready state. Every night we had an automated build and test cycle. We continuously applied the insights from our build and test findings to refine our estimates.

These are the benefits that you are associating with incremental delivery. I associate these benefits with incremental construction. Since we didn't deliver to the customer and get feedback that allowed us to improve our requirements, it wasn't an incremental delivery process.

Roger L. Cauvin said...

Scott, I see distinctions you're trying to make, but the terminology is not quite right. You write:

"I completely agree that design, implementation and testing are the sources of information that make our estimations better. My point is that we can do all of that with a waterfall process too."

and

"I chose to run that project in a continuous release-ready state. Every night we had an automated build and test cycle. We continuously applied the insights from our build and test findings to refine our estimates."

What you describe is not a pure waterfall process. A continuous or iterative release-ready state is characteristic of an agile approach, not a waterfall one.

You also write:

"These are the benefits that you are associating with incremental delivery. I associate these benefits with incremental construction. Since we didn't deliver to the customer and get feedback that allowed us to improve our requirements, it wasn't an incremental delivery process."

You make an important distinction between incremental delivery and incremental construction. Even so, the notion of "delivery" is actually itself ambiguous. Some advocates of incremental delivery don't believe it's essential in every "release" to deliver product to a customer. You can realize a lot of benefits beyond mere incremental development if you deliver the product to your own team as if you were delivering to a customer.

Either way, incremental delivery entails incremental development, which does improve estimation in a manner that pure waterfall processes cannot.