Um,the most important rationale isn't mentioned, which is worrying. Part of the original reasoning behind the "best practice" is that it creates bloated images

Um,the most important rationale isn't mentioned, which is worrying. Part of the original reasoning behind the "best practice" is that it creates bloated images. Since the overlay filesystem captures the deltas from the base image, when you upgrade you create a bigger container than required since it contains both the original base image with un-upgraded files PLUS the deltas with the upgrade differences. If the base isn't being reused with other images on a machine this is wasteful (even if only conceptually). As always the decision to upgrade anyway is an engineering trade-off, but clearly - if possible - starting from an un-upgraded base image would be ideal.