blog posts to appear below this line
We run a lot of tests against Firefox, and running them takes a long time. Ideally, a best effort test selection is done on try, and if nothing obvious shakes out of that a change is landed on inbound. Inbound runs a selection of test jobs pared down based on coarse historical data (SETA). Modulo a handful of periodic build types, the remaining integration branches run a full set of tests and builds.
With the exception of the (entirely optional) selection made by try syntax, none of these steps codify the idea that some tests might fail given a particular change, and others are unlikely to fail or never will.
This post outlines one approach to implementing smarter (semi-automatic) test selection for Gecko.
This is a short introduction to mach try, a mach command that simplifies the process of selecting tests and pushing them to the try server.
If you’ve pushed to Mozilla’s try server recently, you’ve probably noticed some of your failures being re-triggered automatically. This happens courtesy of the trigger bot tool implemented as a pulse listener that invokes self-serve in response to certain events.
Two trigerring rules are currently implemented:
- When a test fails it’s triggered one more time, up to a certain portion of all jobs for the push.
--rebuild is found in try syntax, all test jobs for that push are triggered the specified number of times.
--rebuild is specified, nothing gets triggered when a test fails, and if
--no-retry is found in try syntax, no triggering happens for the push.
This was enabled at the end of June, and I’ve gotten a bit of feedback (almost mostly positive). My own curiosity is about the overall volume of these jobs, so I pulled some stats from self-serve to see how much we’ve ended up triggering. For the first 2 weeks of July, there were 373,772 jobs on try, 13,099 of which were initiated by the trigger-bot (3.5%).
This is a larger proportion than I think it should be, but we’ve recently landed changes that mean we will trigger fewer hidden jobs. Hidden jobs are usually hidden due to failing persistently, so this should reduce overall triggering.
The rudimentary script I used to gather these stats is available on the trigger-bot repo
This is a short post to get things up and running and understand markdown.