Could we transition to having taskcluster-based builds and only the test
suites that have no equivalent (so dump e.g. all of mochitest(-*), but
not js shell as Nicolas brought up) ? Or would that be an inordinate
amount of work? It seems like that might lead to faster + cheaper
builds, plus a good amount of gain from being able to switch off the
majority of the testing immediately.
I have no particular insight in expenses here, but I do know that the
linux build results tend to be the fastest to show up as a result of a
(try) push, and esp. for 32-bit build failures having to wait for the
windows builds would delay when we notice that there are problems.
~ Gijs