Adjusted known related objects handling of target fields which relies on
from and to_fields and has the side effect of fixing a bug bug causing
N+1 queries when using reverse foreign objects.
Thanks Carsten Fuchs for the report.
Data loaded in migrations were restored at the beginning of each
TransactionTestCase and all the tables are truncated at the end of
these test cases. If there was a TransactionTestCase at the end of
the test suite, the migrated data weren't restored in the database
(especially unexpected when using --keepdb). Now data is restored
at the end of each TransactionTestCase.
Previously, foreign relationships were followed only one level deep which
prevents foreign keys to foreign keys from being resolved appropriately.
This was causing issues such as improper database value conversion for
UUIDField on SQLite because the resolved expression's output field's
internal type wasn't correct. Added tests to make sure unlikely foreign
reference cycles don't cause recursion errors.
Refs #24343.
Thanks oyooyo for the report and Wayne Merry for the investigation.
Even though good databases tend to keep the result sorted by the/one
window expression and the planners are smart enough to not resort if not
required, it is not valid to rely on this.
MariaDB specifically did return whatever order it wanted, which is
completely okay. Now we sort towards the expected data for all databases.
Checked the following locations:
* Model.save(): If there are parents involved, take the safe way and use
transactions since this should be an all or nothing operation.
If the model has no parents:
* Signals are executed before and after the previous existing
transaction -- they were never been part of the transaction.
* if `force_insert` is set then only one query is executed -> atomic
by definition and no transaction needed.
* same applies to `force_update`.
* If a primary key is set and no `force_*` is set Django will try an
UPDATE and if that returns zero rows it tries an INSERT. The first
case is completly save (single query). In the second case a
transaction should not produce different results since the update
query is basically a no-op then (might miss something though).
* QuerySet.update(): no signals issued, single query -> no transaction
needed.
* Model/Collector.delete(): This one is fun due to the fact that is
does many things at once.
Most importantly though: It does send signals as part of the
transaction, so for maximum backwards compatibility we need to be
conservative.
To ensure maximum compatibility the transaction here is removed only
if the following holds true:
* A single instance is being deleted.
* There are no signal handlers attached to that instance.
* There are no deletions/updates to cascade.
* There are no parents which also need deletion.
