Firstly is that caching is only a small benefit itself: it saves us
doing precisely identical work for the life of the cache. However
caching in the model layer is a prerequisite to *prepopulating the
caches*, which is where we get the major wins: that lets us eliminate
repeated similar work by doing one query for a set of data rather than
one query per related thing *per row*. So the gain of caching
everything vs caching what we've figured out is safe is relatively
small, because the incremental gain from both of these things is small
without other work, and that other work will apply to read only and
readwrite transactions.
