The Art of Writing Software

Measure your improvements

Tags [ kaizen, lean engineering, metrics, process improvement ]

Metrics are an important part of any development group’s toolset. If we want to continually improve our ability to develop software (through a lean engineering kaizen approach, or simply as a learning organization), then we need to have a way to figure out:

This is where process metrics come into play. I’ll start with my definition of a metric, which is a numerical measurement of something. If you can count it, it can be a metric. So “number of outstanding bugs” is a metric, but “software quality” is not. The term “quantitative metric” is redundant, and “qualitative metric” is an oxymoron.

There are generally two types of metrics we can capture:

  1. causal metrics: these are metrics that have a direct business impact: for example, ROI for a feature, unique monthly visitors, click-through ad rate, etc.

  2. symptomatic metrics: these are metrics that do not directly affect ROI (although we might believe they do) but are downstream indicators for one or more causal metrics. The number of outstanding bugs in a product, the number of bugs caught in a certain phase of development, percentage of code covered by unit tests, etc. are all symptomatic metrics.

My general observation, based on reading articles around the use of metrics for improving your processes, is that a lot of metrics-based improvement projects fail to distinguish between these two types of metrics. Partly, I think this is because while the causal metrics properly align your improvement efforts with your business’s interests, they are also harder to define and measure. By contrast, a lot of symptomatic metrics are easy to find and measure, but their relationship to the business may be less clear.

For example, consider the balance between software quality and time to market. You can take a longer time when developing a feature or product to reduce the number of bugs that show up at deployment, or you can ship a feature more quickly, knowing that there may be both known and unknown bugs present. In this case, you can measure both number of known bugs at deployment time, and you can measure overall time-to-production for a feature (time from feature conception to deployment).

Now, if you can decrease time-to-production without increasing the bugginess of your code, that’s a win. Similarly, if you can reduce bugginess without lengthening your production time, that’s a win. However, both of these things will probably require some effort to implement. Another interesting possibility would be to simply make an adjustment of where you sit on this balance. For example, simply spend more time looking for and fixing bugs in your QA phase, to tradeoff fewer bugs for a slower time-to-market. Or vice versa to get to market more quickly, possibly with more bugs. Both of these adjustments are probably relatively painless to implement, in that no one has to change what they do, just how long they do it for.

So the question is, which one of these things ought we to do? My argument is that bugginess and time-to-production, both being symptomatic metrics, don’t give us the answer directly. It all depends on our product environment. For example, when producing software to run medical equipment, a company reputation for quality might be more important than shipping new features quickly; or, in a highly innovative internet space, time to market might be king in terms of how much market share you can capture.

It’s management’s job to both help implement win-win changes as well as to set the “slider” of the quality/time tradeoff at the right spot. The trick, of course, is that it might be hard to measure this directly; there are several symptomatic things we could measure, including:

Now really, the profit over some window of time (e.g. for a website, revenue vs. development spend for a given month) is the thing we want to optimize. The interesting idea here is for management to be able to run a series of experiments: if I increase/decrease QA or development time, how does it affect ROI for my product? How does the relative bugginess of a release affect its profitability? For certain “sliders” in the business, it is relatively simple to take a series of measurements to find a current “sweet spot”.

An interesting idea here is that sometimes we work through things backwards. For example, we try to estimate “how long will it take to fully regression test a release”, or “how long will it take to code up a feature”, rather than “how buggy will the release be if we test it for X amount of time”, or “how much of this functionality can you develop in X amount of time.” In other words, rather than deriving the time-to-market from a set of estimates for all the steps, instead set the time to market by timeboxing those steps, and see what the outcome is. This is a powerful notion of metrics-based management that is hinted at (in the notion of Scrum timeboxed iterations) but which I have not seen explicitly suggested anywhere [ 1 ]. (Please post all the references to things that I’ve missed in the comments section–I’m sure there are plenty).

At the end of the day, however, it is hard to optimize things we can’t measure. I think important metrics to gather are:

We need to be aware which metrics are causal, and which are merely symptomatic, so that we are measuring things that directly affect the business somehow. This approach permits empirical management–adjust something you can control, see how it affects your causal metrics, rinse, repeat.

1. Scrum timeboxes an entire iteration, but does not timebox an individual feature, so a team may be able to spend all their time on one feature, or spread their effort across many features. The closest thing I’ve seen here is the notion of the “Small” in INVEST user stories, where stories are limited to a certain amount of complexity. However, the story points in this case are still estimates of the work involved, rather than timeboxes around how much time to spend implementing a feature; the “small” requirement is really to permit more accurate estimation rather than to timebox the amount of effort (although it does secondarily have this effect, I’ve not seen this stressed in articles about this).