Monday, February 27, 2017

Continuous Planning

"In preparing for battle I have always found that plans are useless, but planning is indispensable." -- Dwight D. Eisenhower

There is a tension between engineering on the one hand, and on the other hand those who would like to know when the task will be done. A product must be marketed, documented, sold, and supported. "When it's done," is useless when you're trying to sell to a customer against a market full of competitors. However, the software we write gets more complex each day, and the process for bringing it to life is complex. Complexity means unknowns, and unknowns mean uncertainty. A software project is like a hurricane with a cone of uncertainty preceding it. This tension between the desire to know and the reality of uncertainty is a fundamental part of working a software project (and probably other kinds of projects).

Before going too much farther I will state my assumption: a completion date is an output not an input, and the most effective tool for managing a completion date is changing the amount of work you want to do (i.e. "scope").

You cannot take a date and work backwards. This is no different than taking a date and working forwards. Well actually, there is a big difference. In working forward, you can always push the completion date out. In working backwards you cannot start any earlier than now. The completion date inevitably follows from when you start, how quickly you can work, and how much you are trying to do.

You can spend money on tools, training, consultants, but these each have a time cost.

You can add more people, but in order to establish a context on a project a new person must learn a code base, tools, technologies, personalities of the team, and to do so he or she must take time from an otherwise productive member of the team.

You can have the current team work overtime, but too much of that will cause quality issues and burnout.

You can relax expectations about quality, but that is just trading your future time to get something done more quickly and temporarily.

The best thing you can do to manage a completion date, is to cut the amount of "stuff" you are trying to do, or to rearrange the order of when you will do it, so you get the things you want earlier than you otherwise would have.

Given that a date is an output, as engineers and managers we try to navigate this tension between the desire to know and the reality of uncertainty with planning, but there's a problem with plans: they're useless. Imagine planning a single task. When will it be done? Well, if you ask one engineer she will give you an estimate based on her skill and experience. If the task is ever given to another engineer, then that estimate is invalidated. On top of that, an engineer (or human really) is notorious for estimating only the amount of work she must do. She doesn't think about QA testing, deployment, and data migrations, among other things. Nor does she think to factor in overhead like meetings, filling out time cards, learning new skills, bonding with coworkers, etc.

That is just at the most atomic level of estimation. Once you start to think about collaboration things get more complex. Does our engineer need to get a review from a coworker? That coworker is now being taken off of his task to do the review, which can lead to delays. What if our engineer needs assistance from someone more familiar with a particular technology or part of the code base? What if our engineer wants to brainstorm with another engineer? If our engineer gets delayed then any tasks that were dependent on her task also get delayed.

Now imagine making a plan for a product that spans several teams and tens or hundreds (or thousands??) of people. If you don't know everything that everyone is working on and how they are all related, then you can't with certainty plan out anything, and that is only taking into account everything that can be known, there are still unknowns (like if someone may get sick). This is the uselessness of a plan.

Well, a plan is not entirely useless. It is probably very accurate for the tasks that will be started in a few days, but entirely inaccurate for the tasks that will be started in three or six months.

So, there are two problems with a plan: 1) it must be updated to reflect new information, and 2) it fails to take into account the "cone of uncertainty."

Updating a plan seems easy enough, however the larger the plan the more work it will take to keep updated. One could certainly employ an army of project managers who verify that the task breakdown, estimates, and dependencies have not changed; that you've taken into account every meeting, vacation plan, all the testing, deployment, and overhead. Ideally the plan would be updated continuously (so its more of a "dashboard" than a "plan"). More valuable than knowing that three months ago we thought a task should be complete on such and such a date would be knowing when we think it will be completed as of now with all the latest information we have access to, but that would create quite a drag on the entire team.

Even if you could keep the plan up-to-date, it gives the false impression that one can know precisely when a task will be complete. You may be able to predict the completion date of a task that starts tomorrow, but not for a task that will start in three months. Three months provides plenty of time for both knowns and unknowns to change when the task could even start, let alone when it would complete.

Usually, this uncertainty is handled by "padding" the date, but this is not enough. A single point-in-time completion date conveys certainty, and this is certainly wrong. The completion of a task should always be a range, one that is narrow for the near future and wide for the distant future.

Incidentally, I think even agile burndown charts get this wrong. In my opinion, there is (and should be) variability to a team's velocity. Simply taking some velocity value and running it out a few months to predict a single point in time when a task will be complete is at odds with reality.

What does Continuous Planning look like? Well I don't really know, because I just made it up! At a high level I would summarize it as: plan using real data, with task completion ranges, over as long a term as you want, in aggregate, on average, continuously. The task completion ranges are the key. You can plan over as long a term as you want, however, the ranges will get wider. If you can reduce the variability in your process---and prove it with your data---then you can narrow the ranges. Planning is done in aggregate and on average, because it is impossible to know and manage every possible factor, so we must abstract away much of the minutiae. Finally, to plan continuously implies some kind of tool to facilitate.

To the extent that I have thought about how this would work out practically, this is what I would do:

Each team would estimate their tasks by each member recording the number of hours he or she thinks it would take him or her to complete the task, given that there are no other distractions. This is a kind of pure estimation that engineers usually make. It would be helpful to discuss the task as a team, and try to elicit different opinions on the complexity of the task, so the estimates will be as complete as possible.

Why not use story points? I have been a fan of story points precisely because they abstract away hours. Hours can vary depending on a persons skill and experience. Hours can get lengthened by interruptions and discovering additional complexity. Hours give a false impression that they would map directly to calendar time, and you can accurately predict when a task will complete.

However, the first thing a person asks is how long does X points take. Usually you have to pick as a standard comparison some "golden story" for a certain number of points. People will usually consciously or unconsciously come up with some rule of thumb like, "an eight point story should take about a sprint to complete." So in the end you are estimating in hours, but they're a convoluted form of hours.

Hours are a natural unit for estimations. The danger in using hours is actually trusting the estimate for a precise completion date. We've already rejected precise completion dates with Continuous Planning, and the rest of the process is designed around (automatically) finding an accurate scale with which to judge these estimates. I would actually advocate that the estimates and velocities be hidden variables, and (other than your own estimate) you only see the completion range for a task. This would hopefully reduce some confused expectations around what it means for an estimate to be denominated in hours.

The estimates from each team member would be combined together into a single estimate for the task. The method for combination could be taking an average. It could involve throwing out extreme values first, or doing some sophisticated statistical analysis.

Having done these estimates, a task tracking system would keep track of when tasks started and when they completed, or how long they've been started even if not complete. This actual data can be used to calculate a velocity. The velocity could be calculated at several levels. You could calculate the velocity for an individual task, for a particular team member, for the team as a whole. You could even calculate the velocity for a feature epic cutting across several teams.

To calculate a date range for completion, you can take the average plus or minus a standard deviation for a single velocity calculation over time and get an optimistic and pessimistic velocity. You could get an optimistic and pessimistic velocity by taking the minimum and maximum of the most recent velocity calculations at two different level (task and team, for example). I'm not sure which would work best; it warrants some research.

Tasks would be a hierarchical tree. An epic is really just a task with subtasks. The velocities and estimates can flow up the tree for the purposes of calculating estimated completion ranges for epics.

If you wanted to get fancy, you could draw dependencies between tasks, and the system could then attempt some kind of topological sort of the tasks, and using a prioritized backlog and team assignments to each task, construct a plan for what could be done in parallel, and---based on velocities calculated from real data--calculate a completion range for each task, epic, and the project as a whole.

As you can see there are still many questions to be answered. I think this is an idea worth exploring. In my experience, the usual tools fall flat at resolving the tension between the desire to know and the reality of uncertainty.

To effectively attack this tension requires abstracting away much of the minutiae of detailed planning by embracing the variability of the process. The plan is useless. Planning is indispensable. Therefore, plan continuously

In preparing for battle I have always found that plans are useless, but planning is indispensable.
Read more at:
In preparing for battle I have always found that plans are useless, but planning is indispensable.
Read more at:
In preparing for battle I have always found that plans are useless, but planning is indispensable.
Read more at:
In preparing for battle I have always found that plans are useless, but planning is indispensable.
Read more at:
In preparing for battle I have always found that plans are useless, but planning is indispensable.
Read more at:
n preparing for battle I have always found that plans are useless, but planning is indispensable.
Read more at:
n preparing for battle I have always found that plans are useless, but planning is indispensable.
Read more at:
n preparing for battle I have always found that plans are useless, but planning is indispensable.
Read more at:


Ian said...

Have you ever looked at Joel Spolsky's "Evidence Based Scheduling"?

The key feature is using monte carlo methods to statistically combine past estimates and actuals to create a range of possible future outcomes, from which you can say things like we have a 90% chance of finishing by a given date.

The #NoEstimates movement does something similar, but also says that the variability in task size is lost in the noise of the variability of estimates anyway, so we might as well assume all tasks are the same size - hence we get a range of possible future outcomes almost as good without even estimating anything.

Paul Stadig said...

Very interesting! Thanks for sharing those resources.

I figured that I can't be the first to think along these lines.