Comment 3 for bug 1732681

> Looking at the Xenial numbers I believe -3 would result reasonably fast
> builds with a tolerable increase in size compared to xz -6's results.

As you've seen, the rpi2 server image is the only place we're using this code today. We are not overly concerned with the build speed as these are highly-parallelizable builds that are not part of a critical pipeline. I believe we should go with xz -6; certainly for bionic and forward where we have actual parallelization, the build time penalty appears to be negligible.