This is a follow-up to my earlier findings about add-ons and startup times. In this post I am going to dig deeper between the relations of add-ons and shutdown times. A slow shutdown doesn’t seem to be a big deal. If one considers though that a new instance of Firefox can’t be launched if the old one is still shutting down, the issue becomes more serious.
It turns out that for shutdown times a simple linear model is not good enough while a log-linear instead has a reasonably good performance. Log transforming the shutdown times slighly complicates the meaning of the coefficients as they have to be interpreted as the percentage change of the average shutdown time. E.g. if an add-on has a coefficient of 100%, it means that it might (correlation is not causation!) slow down shutdown by 2 times. The idea of using a log-linear model comes from our contributors Jeremy Atia and Martin Gubri , which discovered the relationship during a preliminary analysis.
Unlike in the startup case, there are fewer stronger relationships here. Some patterns start to emerge though, the Yandex add-on for instance seems to be associated with both slower startup and shutdown timings.
We started to keep track of those results on a weekly basis through a couple iacomus dashboard: one for startup and the other for shutdown times correlations. The dashboards are surely not going to win any design award but they get the job done and didn’t require any effort to setup. I am confident that by spotting consistently ill-behaved add-ons through the time-series we should be able to spot real tangible offenders.
 If you love probability, statistics and machine learning and are looking for an open-source project to contribute to, Firefox is a cool place to start! Get in touch with me if that sounds interesting.