Regression Ratios

Regression is a nasty issue. Ongoing regression from bug fixes can be a pretty clear indicator that there are some serious problems with your code base, your process, your team or all of the above. As an example case consider the effect of moving from 1 in 5 of all bug fixes causing an unrelated issue to crop up to 1 in 10. Given an imaginary scenario where 1000 bugs are found (a medium size project) and a team is closing 20 of those bugs a day then in an extremely simple model we have just added two weeks to our timeline simply from regression issues.  (handy chart from google spreadsheets below)



Of course its much more insidious then that though in the real world. For one, regression bugs only pop out at you the minute they are introduced if you’re lucky (they are obvious). More likely is that about half of them will show themselves and the other half will start cropping up near the release date during regression passes. If you alter the model above to show that you’ll see little bulges near the end of the line that start to make time lines feel very elusive and untrustworthy. And given that no QA process is perfect you have to believe that whatever is causing you to introduce those bugs that you are seeing is also causing bugs that are yet to be found.

The importance of quality can never be over-stated, and the effect of poor quality can have a really dramatic impact on a team as well as your customers. Here you are, slaving over your own code, crafting a well thought out and tested solution to a problem, only to have someone else come along and break your code with their fix. Here you are, proudly about to successfully bring a feature out of the QA phase and into the hands of customers and all these elements that you had already seen work are failing. How quickly your trust in your fellow developer begins to wane.

Here though is a problem that I have come to see the light on as being truly systemic. Yes there are horrid developers, and yes there are some difficult technologies, and yes some level of regression is inevitable…. but levels of regression like those that I’ve seen can usually be traced back to process, to your design and architecture, to your requirements and most importantly to your culture.

In terms of process there are some stock answers to bringing this number down like TDD and constant refactoring. I believe in both of these, but in my experience these can be difficult to enfuze into a culture. Without full buy-in and a cultural shift in developers these are just more TLA’s to throw on the steaming pile of buzzwords that will help you magically improve.

Having just come to the end of a release of our product in the past few weeks I am looking forward to whittling away at our own regression ratios.

Changes we’re making :

  • No more sharing of the bug load. Every developer owns their module/feature/code from design through to maintenance. This include assigning ownership to old modules whose owners may no longer be around. If you write bad code, you will fix more bugs. If you fix more bugs, you will have less time to spend on new bug-ridden features. This will mean our actual bug trends won’t be the pretty descent we’ve worked so hard on achieving. But the overall effect of a self correcting system should help. 
  • More atomic checkins. Checkins that span branch points, or handling merges of bugs can be extremely error prone with subversion and our current branching strategy. We hope to address this somewhat by eliminating bug fixes that span multiple revisions… Ideally this will involve :
    • Patch files attached to the bugtracker for code reviews to avoid multiple checkins when reviews spot errors
    • When changes still need to be made to a bug when things are caught in QA then we also revert the original checkin, and checkin again with all changes at once in the new revision. 
    • Bugs will be closed diligently and new bugs opened rather than morphing the bug over time as new side effects or slightly related bugs are spotted. 
  • More automated testing. This is last on the list not because it’s least important, but because we don’t expect as quick of an impact as with the other two steps. A major number of the issues we run into remain UI issues. And while WatiN does do wonders for us, it’s painstakingly slow to get the number of tests to where they should be (and keep them up to date). 
Lastly just for interest here’s the graph of how we’ve actually done. Black shade are high severity bugs. The first mountain was the release previous where we amalgamated all previous releases into this (migration of existing clients) we had some pretty serious regression in that release after Christmas. The  second range is the most recent release which is looking amazing in comparison, though at this resolution you are missing some definite climbs amongst that fall. Definite progress anyway….