Sunday, February 28, 2010

Estimation Poker

Estimation is a necessary part of software development. Product owners want to know how much work can get done by a deadline, project managers need to make commitments, and developers want to know if they committed to a reasonable amount of work.  While estimates are often inaccurate, estimates  provide landmarks along the way of a project to gauge progress. So, estimates are an inevitable, and useful part of the software development process. Many complain that the process of getting to those estimates, estimation, takes too long, so planning sessions are cut short and teams don't have enough time to discuss issues that have uncertainty. By appropriate use of planning poker, you can balance the needs for good estimates while minimizing time spent estimating.

Sometimes when a team is asked to estimate a backlog item, one or more people with expertise in an area are asked to estimate the item, but this is not the best way for an agile team  to get a good estimate.

There are benefits to involving a larger part of the team in the estimation process. The challenge is that people feel that involving the whole team is wasteful if the estimation process takes too much time. On the other hand, inaccurate estimates have their own costs for the team and the other stakeholders.

Planning Poker, an estimating method popular with Agile Teams can address some of these issues. Briefly, planning poker involves getting the developers on a team together to estimate stories using a deck of cards that have numbers that represent units of work. The numbers are often spaced in a Fibonacci sequence, the theory being that the larger the estimate, the lower the precision. Planning planning poker can be a really useful tool to both improve estimation and discover uncertainty in requirements.

People resist planning poker for reasons like:
  • It seems inaccurate if the person doing the estimating does not having the "appropriate" expertise. A UI developer may not feel qualified to estimate a story that seems to be mostly backend processing, for example.
  • It seems like a waste of time because people believe that one person can estimate for everyone.
  • It seems inaccurate since the person who's been assigned the work should estimate it based on their skills.
Even if you find yourself throwing a wild guess at a planning poker session, the fact that you don't understand the scope of the issue is useful information. The benefit of having the entire cross-functional team understanding and estimating stores is that you can identify challenges across the application. What might be easiest to do in the back-end can add work to the application tier or UI, and also make testing harder. Having one person estimate  can make it hard to identify misunderstandings and issues because we tend to want to agree with "the expert," and there is no forum for identifying misunderstandings.  It's not always clear at the start of a project who the best person for task will be, both for the reasons I just mentioned and because assigning the work up front can lead to inefficiencies if work takes more or less time than estimated.

If you find that your estimates are inaccurate, or your estimation process takes too long, consider the following approach:
  • Gather team members who are working on all aspects of the application. You need not have the whole team, but be sure to represent each "architectural layer". If your team is less than 7 people or so, include everyone.
  • Look at the description of each story or problem report in priority order. Ask the team to pick cards based on what they read.
  • See how close the estimates are. 
    • If they are close, ask someone to explain what they envisioned doing to implement the issue. If someone has a vastly different idea, they should speak up. 
    • If they are different, as someone with one of the extreme estimates to explain their reasoning. This will start a conversation about what the requirement means, and what implementation strategy makes sense.
This process helps you to focus discussion time on the hardest, highest priority issues. You will want to be sure that to allocate an appropriate amount of time to planning and estimating relative to your sprint length. You may still run out of time, but even if you do, you'll  have discussed and estimated the highest priority items as accurately as you could have, knowing what you knew. 

The biggest challenges to having accurate estimates are not having consensus on the "what" and not understanding the details of the "how." The process above is one way to focus discussion on the high-risk items in your backlog, while keeping the time spent on estimating reasonably low.  

Saturday, February 20, 2010

The Indivisible Task

One of the things that makes agile work well is a daily sense of progress that can be reflected in, for example,  a burn-down chart.  For burn-down charts to be meaningful, the estimate of amount of work remaining in a sprint need to be accurate. Re-estimating work remaining in a task is helpful,  but the best measure of progress is the binary "done/not done" state of the items in your backlog.

Assuming that you have a clear definition of "done" for a task,  it's easiest to measure progress when you have tasks that are small enough that you can mark them complete on a daily (or more frequent) basis. Breaking work down into a reasonable number of reasonably sized tasks is something many find challenging. (Note: I'm talking here about development tasks as part of a sprint backlog, rather than splitting User Stories in a product backlog, though there are some parallels.)

 I've worked on teams when people refused to break down large task into 1-day or smaller parts. The common excuse for not breaking down work  is that the person who signed up for the work understood what the work was and the estimate was accurate. Of course, we had no way of knowing that the estimate was at wrong until the work was not done at the end of the week or so.

What was interesting to me is that those most resistant to decomposition weren't less experienced programmers, but rather the people the team acknowledged as "experts" and "good designers" who were good at decomposition as it applied to designs. So the theory of attacking complexity through looking at smaller pieces was something they were comfortable with. Not only that, they actually worked in a way that led to discrete units of work being completed throughout the project, whether in terms of frequent commits or even simply being able to finish a work day with a sense of accomplishment, even if the great task was still incomplete.

Breaking down work isn't as hard as some make it sound.   From a development centric perspective some of the things you already do which can guide you in task breakdown:
  • Thinking about when you might commit code. It's good practice to commit code frequently; consider the Task-Level-Commit pattern from Software Configuration Management Patterns
  • Considering what the tests you write as (or before) you code.  
  • Deciding what you want to accomplish before leaving work each day so that you end the day with a sense of accomplishment.  
What these items have in common is that the define natural boundaries in the development process.

The main difference between doing this kind of planning and good programming practice is making your plan visible to others. This takes discipline, and a certain amount of risk, since if your plan goes awry it's visible.  Part of being a successful agile team is understanding that plans can be wrong, and using that experience to figure out how to do better in the future.

You may discover part way through your planning that the task breakdown you did no longer makes sense in light of something you discovered. That's OK, but at least you have a good sense of what work was done, and you can figure out what tasks are left (and estimate them!)

By breaking down work into smaller parts you have the ability to:
  • Evaluate your progress in a definitive way. as it is often easier to define "done" for a smaller task.
  • Get feedback from your colleagues before you dive in to a problem. 
  • Share effort if any of the work can be done in parallel.
  • Simplify updates and merges, as the changes to the codeline will all be small at any point in time.
It is possible to come up with too many sub-tasks, such that the overhead of tracking them on a backlog negates their value as a tracking tool. In that case,  there is nothing to prevent you from taking note of the very small things you do each day and combing some of them into day or half day items that do appear on the backlog. And if you really only want to have large tasks on the sprint backlog, consider doing your own fine grained breakdown that you can use to help you give a better estimate for time remaining. I tend to favor backlogs with tasks of a half to one day, and then making personal notes about smaller steps  to complete  those tasks on my own.

The value of working in small steps isn't a new idea. in 1999 Johanna Rothman wrote about the value and mechanics of managing your work in terms of  which she calls inch-pebbles (as opposed to "milestones") and Fred Brooks advised "allow no small slips" in  The Mythical Man-Month, and being able to identify these slips is key to having effective sprints.

Agile teams work because they have mechanisms to give frequent feedback on status. Accurate estimates of work remaining are an essential tool for evaluating progress, and small tasks help you estimate accurately. Decomposing work is not easy, and takes discipline, but the benefits are great.

Tuesday, February 16, 2010

97 Things Every Programmer Should Know is Done

The book 97 Things Every Programmer Should Know: Collective Wisdom from the Experts is finally available, and the title is on the mark.

Kevlin Henney, who I first met at the 1998 PLoP conference, asked me to participate in this project  in September of 2008.  I am honored to be a part of the list of contributors, which includes Kevlin,  Bob Martin, Michael Feathers, Giovanni Asproni, and many others who have important things to say about how to build great software. Kevlin did an amazing job coordinating and editing, and the the book represents an excellent cross-section of the many contributions that formed the basis for the final version.

Reading this book gives you a chance to learn from the experiences of people who've worked hard not just at writing good code, but at creating good software systems. Some of the advice may be things you already know. Some items may be surprising. Read this book to learn, be challenged, and to understand why programming isn't just about languages and syntax.

For more info, you can look at the associated wiki site. And feel free to share with me any thoughts you have about my contributions: Deploy Early and Often and Own (and Refactor) the Build.

Sunday, February 7, 2010

Tracking what Matters

I'm a big fan of burn-down charts for tracking sprint and release progress.  The basic idea of a burn-down chart is that the team starts with estimates for all of the tasks in the sprint, and then on daily (or more frequent) basis re-estimates the amount of work remaining.

With a burn-down chart, you are tracking the new estimate based on the work that you have done. As you work on the sprint backlog you get a better understanding of the tasks, and thus you can revise estimates for tasks that span more than one day. This is reasonable since the original estimate is, well, an estimate.

Sometimes if you spend 4 hours on an 8 hour task, you'll have 4 hours of work left. Most of the time the time left  will not be the original estimate less the time spent, but more or less. At the end of 4 hours, the remaining work estimate for the same 8 hour task could be 2 hours, or it could be 10 hours if you discovered a snag.  This is important information for everyone involved in the project and allows the team to identify a problem at the daily scrum. Re-estimating is harder than just doing subtraction, but it's valuable.

One thing that happens when teams use an issue tracking tool (like Jira and Greenhopper) to manage their backlog is that re-estimating and effort tracking are combined. The only way to re-estimate is to "log work." You're required to enter the amount of time spent, and the tool will kindly offer to change your estimate to the difference between the original estimate and the time spent. There are two problems with this:

  • It's important to think about the time left for the task based on the assumption that your original estimate had a margin of error. For all but trivial cases, the "calculated new estimate" is always wrong.
  • The "time spent" value isn't really useful to stakeholders. In the best case you are only doing one thing during the time in question, so your time spent entry is accurate, but doesn't answer the question: "when will it be done!" In the worst case, you're not tracking your time accurately, and the time spent number is inaccurate, and provide no real  information.

Like all things agile, when looking at your project tracking approach you need to be clear about what you want to track and why. The main concern for stakeholders on an agile project is whether they will get the functionality they want at the end of the sprint. So the time-remaining number is important.

There are some good reasons for tracking time spent including evaluating estimation accuracy and billing
But in both of these cases you need to evaluate the overhead of the tracking time relative to the value. Tracking total effort for the sprint relative to estimated work done may be more useful than per-task effort to estimate tracking, and analyzing the results in a retrospective may yield more useful information than per-task tracking.

When doing sprint tracking:

  • Make sure that everyone understands the goals of the tracking process so that you get uniformly valuable results. You definitely want to track how close you are to "done," but explain how important tracking effort is. 
  • Make sure that, whatever the goals, that the data are updated daily. If the burn-down chart doesn't change for 2 days is it because people didn't update their estimates, or that the project is at a stand-still?
  • Remind everyone that the estimates are just that: "estimates," and an informed guess that turns out to be wrong is better than no estimate at all. (And the inaccuracy of  the estimate helps to identify unknown complexity.)

Burn-down charts can be a simple, valuable, tool to identify problems during a sprint as long are your teams breaks out of the habit of tracking "effort" as opposed to "effort remaining."

Site Reliability Engineering; The Book and The Practices

Site Reliability Engineering It’s difficult to walk into a software development organization without hearing about the discipline of Site ...