Avoid unnecessary clear of cached reference

[Barry Johnson]

Consider this case of ORM models "Parent" and "Child", where Child has a foreign key reference to Parent (and the database can return generated IDs following insert operations):

parent = Parent(name='parent_object')
child = Child(parent=parent)
parent.save()
child.save()
print(child.parent.name)

The print statement will cause an unnecessary lazy read of the parent object.

In the application where this behavior was first observed, the application was creating thousands of parent and child objects using bulk_create(). The subsequent lazy reads occurred when creating log entries to record the action, and added thousands of unwanted SELECT queries.

Closed ticket 29497 solved a problem with potential data loss in this situation by essentially executing child.parent_id = child.parent.pk while preparing the child object to be saved. However, when the child's ForeignKeyDeferredAttrbute "parent_id" changes value from None to the parent's ID, the child's internal cache containing the reference to "parent" is cleared. The subsequent reference to child.parent then must do a lazy read and reload parent from the database.

A workaround to avoid this lazy read is to explicitly update both the "parent_id" and "parent" cache entry by adding this non-intuitive statement:

child.parent = child.parent

after executing parent.save()

But it appears that a simple change could avoid clearing the cache in this narrow case.
Within Model._prepare_related_fields_for_save(), replace

setattr(self, field.attname, obj.pk)

with

self.__dict__[field.attname] = obj.pk

This suggested code has -not- been tested.

This change would set the associated "parent_id" attribute to the desired value without affecting the cache. In this spot of the code, "obj" is currently set to the cached parent object that we want to preserve, and we're just reconciling the associated copy of the parent's primary key.