Allow QuerySet.bulk_create() to set pk of created objects when ignore_conflicts=True

[saber solooki]

As requested in ticket:28668#comment:24, here's a ticket for setting the primary keys of created objects. I tried it in PostgreSQL 10.6 and this could be supported. When ignore_confilicts=True, Django doesn't put ' RETURNING "mymodel"."id" ' in the query.

[Дилян Палаузов]

At https://code.djangoproject.com/ticket/28668 are attached patches, that accomplish this.

[Tim Graham]

​PR

[Hasan Ramezani]

Based on the ticket description we should add ' RETURNING "mymodel"."id" ' to the query when ignore_confilicts=True. But based on the ​Simon comment on the PR the database will not return any conflicting row ids.
I am a little confused because of the Database just returns newly created object ids. Check Simon's example in the comment.
What should we do?

[Simon Charette]

Hasan, this was mentioned in https://code.djangoproject.com/ticket/28668#comment:21 but was dismissed as something the proposed implementation would take of.

It does so by doing a LEFT JOIN over the existing rows and would require a set of columns against which to join to be provided to the call.

Implementing this in a proper manager would likely require a lot of work. It would require ignore_conflicts to accept an iterable of constraint names or tuple of fields corresponding to Field.unique=True or Meta.unique_together definitions. These entries could then be matched to constraint names and used as conflict_target in the ON CONFLICT clause and in the join USING clause. I think we'd also like to allow exclusion constraints to be treated differently because joining on fields won't help here as an insert could be conflicting with more than one result.

I assume callers would also like to be able to differentiate between newly inserted and conflicting entries so the return value of bulk_create would need to be formalized/changed. I assume we'd like to return a list of tuple of the form (bool: inserted, Model: instance, List[Model]: conflicting_instances) and only assign the pk on non-conflicting rows.

In summary this ticket involves a lot of research work to get right and not paint ourselves in a corner with regards to the desired API.