It is no secret to my Twitter followers, and perhaps beyond the Twitter-sphere, that I am on a crusade of sorts to get people considering other ways besides estimating when it comes to costing software development projects and tasks. Such a view remains controversial, even among Agile practitioners. People argue that there is no alternative; customers want estimates, so we must provide. Stakeholders need to know when things will get done. Estimation is seemingly one of the few remaining immutable practices hanging over from the Waterfall era.
One of the common criticisms of my view is that it is unduly dismissive. When asked by our boss or a customer for an estimate, we can’t simply palm them off and say “I don’t estimate! Talk to the hand, sir!”
Of course this is true. But I should point out that I actually see nothing wrong with being asked for an estimate of how long something will take. What I object to is being asked to carry out (or ask my team to carry out) estimation rituals whose results will then be used for making important business decisions.
We cannot palm people off, but what we can do is offer alternative, empirical approaches to traditional and “Agile” forms of estimating, explain exactly how we will provide the required information and why such approaches offer advantages over guessing “how long” or “how big”.
First off, I would suggest that there are many problems with the “how long/big” approach, the biggest of which is that such an estimate does not take into account the:
- Inherent unpredictability of building software
- Current work in progress (i.e. the team/dev may not be able to start the work “now”, or even for a few days, weeks or longer)
- Capacity to do the work (i.e. the team/dev may make the estimate based on certain assumptions of team size which turn out to be false, or a colleague being there who ends up not being), nor
- Any upcoming changes in priorities (i.e. something may jump above the piece of work in priority).
From a task point of view, what is estimated as a “10 minute job” may end up taking a day or longer due to one or more of the above. I’m sure you have seen this situation many times over. From a project point of view, this situation is magnified and can be hugely costly, even catastrophically so. 3 month projects become 6 months. 1 year projects become 3 years.
In a situation where there are small tasks flowing through from the customer to the development team that are unpredictable in their timing (e.g. BAU work queues, feature development, etc.), a far better, probabilistic approach to get some semblance of predictability is to do the following:
- Measure actual lead times of every piece of work and plot them in a Lead Time Distribution graph
- Measure throughput (you can start by simply counting the number of cards in the “done” column at the end of every week)
- Use a fixed WIP limit on cards in progress (start, if you like, with the natural limit of team size)
- You can now use Little’s Law to calculate average lead time for a card at position n in the queue, i.e. (WIP + n) / throughput:
- e.g. Number of cards done in 1 week = 20, therefore throughput = 4 cards/day
- Team size = 2, therefore WIP = 2
- Lead time = (2+1)/4 = 0.75 days (i.e. on average it will take three quarters of a day for a card at the top of the queue to be delivered)
With the same formula you can predict where a card 2nd, 3rd or xth in the queue will get done, which is very helpful for guiding your prioritisation:
e.g. Using the same example above, a card 2nd in the queue will likely be done in 4/4 = 1 day, while a card 6th in the queue will likely be done in 8/4 = 2 days
Bear in mind the only way this formula can provide useful numbers is by having a WIP limit that is fixed (as far as possible). There will of course be variability in how long each card takes, but the law of large numbers will even this out to an acceptable average and it’s certainly far more scientific than asking people to estimate each card.
Note that if you use Scrum, and thus the team breaks down features into small tasks just-in-time at the beginning of every Sprint, you can use the same principles as above to determine when a new feature might be delivered (Scrum has a WIP limit over the Sprint length of the number of tasks in the Sprint Backlog, throughput is the number of “done” stories/tasks divided by the Sprint length, etc.).
Over time you can achieve a higher level of confidence with the predictions as you start to identify and split out different work types, determine probability of delivery times using your Lead Time Distribution graph, etc.
What about “how long will this project take?” !! Warning !! You can scale this approach up to the portfolio level. But… do bear in mind that building an entire software product rarely has a finite end point or a repeatable result because it is not possible (nor desirable) to define all of the scope required to deliver a delightful, valuable outcome. Use such predictions with extreme caution. There is no substitute in software product development for creating certainty around costs and delivery times via fixed agile teams delivering working software early and often, short feedback loops with the customer, etc.
So, next time you’re asked “how long” or “how big” about a software project or task, don’t palm off your boss or your customer with simply “I don’t estimate!”. Perhaps you might consider answering: “I don’t estimate! But… here is how we can save ourselves the cost of estimation meetings and make empirical predictions going forward to answer these questions with more confidence.”