On 06/07/2016 06:11 AM, Marc Tamlyn wrote:
> I may be "too close" to knowing the implementation of this feature to be
> able to comment on whether the behaviour is surprising to most people,
> but it doesn't surprise me. It's certainly analogous to that when you do
> `MyModel.objects.create()` it doesn't change an already executed
> queryset. There's a question of where you draw the line as well - what
> about `related_set.update()`?
>
> I think it's worth documenting the behaviour, also noting that you can
> "force" the execution of a new queryset by chaining another .all().
Hmm, I have the opposite instinct. I don't find it analogous to the case
of some unrelated queryset object failing to update its internal cache
when the database changes. In this case we have a related-manager with
an internal cache, and we make changes _via that same manager_. I find
it quite surprising that the manager doesn't automatically clear its
cache in that case.
A much stronger precedent, I think, is the fact that Queryset.update()
does clear the queryset's internal result cache. In light of that, I
think the current behavior with prefetched-data on a related manager is
a bug that should be fixed (though it certainly should be mentioned in
the release notes).
Carl