ModelChoiceField generates unused temporary cache

[Thomas C]

There is a huge memory consumption when using some admin pages with big production datasets (see also ​https://groups.google.com/forum/#!topic/django-users/YYlLWyBH_go).

I looked into the ModelChoiceIterator used to generate the choices for a ModelChoiceField, and noticed that an unused temporary queryset cache was generated due to the use of .all() (there is no variable pointing to the cloned queryset after the iteration).

With the proposed PR, this cache is no longer generated. After applying the patch, I noticed that the maximum memory usage dropped from ~200MB to ~150MB for a model with about 30 000 instances.

[Tim Graham]

Have you seen #22841? I think that is related to this.

[Thomas C]

I missed that ticket, thanks. My understanding of #22841 is that it proposes a way to "tell" the ModelChoiceField to reuse a queryset instead of reevaluating it at each time.

I think that my PR is related but compatible with that proposed change as it handles the case where we don't want to reuse the queryset. That case is currently handled by iterating on queryset.all() — which creates a cloned unevaluated queryset — to construct the choices. My proposition is to use iterator() instead as this is designed precisely for those cases where you don't want to use cache nor to populate it.

[Tim Graham <timograham@…>]

In fa534b92dda0771661a98a1ca302ced264d0a6da:

Fixed #23623 -- Reduced memory consumption when generating ModelChoiceField choices

[Simon Charette]

It looks like this introduced a regression when passing a queryset using the prefetch_related() which is not compatible with iterator() by design #25496.

I would suggest we conditionally use iterator() if no related data should be fetched instead of simply reverting this optimization. Unfortunately there's no public API to determine whether or not a QuerySet is prefetching data and we would have to rely on the qs._prefetch_related_lookups property.