The original problem was that queryset cloning was really expensive
when filtering with F() clauses. The __deepcopy__ went too deep copying
_meta attributes of the models used. To fix this the use of
__deepcopy__ in qs cloning was removed.
This commit results in some speed improvements across the djangobench
benchmark suite. Most query_* tests are 20-30% faster, save() is 50%
faster and finally complex filtering situations can see 2x to order
of magnitude improvments.
Thanks to Suor, Alex and lrekucki for valuable feedback.
The guarantee that no queries will be made when accessing results is
done by new EmptyWhere class which is used for query.where and having.
Thanks to Simon Charette for reviewing and valuable suggestions.
When &'ing or |'ing querysets, wrong values could be cached, and crashes
could happen.
Thanks Marc Tamlyn for figuring out the problem and writing the patch.
The added promotion logic is based on promoting any joins used in only
some of the childs of an OR clause unless the join existed before the
OR clause addition.
The ORM didn't reuse joins for direct foreign key traversals when using
chained filters. For example:
qs.filter(fk__somefield=1).filter(fk__somefield=2))
produced two joins.
As a bonus, reverse onetoone filters can now reuse joins correctly
The regression was caused by the join() method refactor in commit
68847135bc
Thanks for Simon Charette for spotting some issues with the first draft
of the patch.
This is necessary because get_model() checks are case insensitive, and if the swapable check isn't, the
swappable logic gets tied up in knots with models that are partially swapped out.
Thanks to chris@cogdon.org for the report and extensive analysis, and Preston for his work on the draft patch.
Cleaned up the implementation of base convert_values() a little, and
made sure it accepts None as a value for numeric fields.
There are no tests attached. The reason is that not all of the
convert_values() accept None as a value for numeric fields (for example
sqlite3.convert_values()).
The reason the base convert_values() needs to accept None is that this
situation might arise in custom compilers for 3rd party backends. It
is easy to keep the convert_values() working, so lets do that.
This is a rather large refactoring. The "lookup traversal" code was
splitted out from the setup_joins. There is now names_to_path() method
which does the lookup traveling, the actual work of setup_joins() is
calling names_to_path() and then adding the joins found into the query.
As a side effect it was possible to remove the "process_extra"
functionality used by genric relations. This never worked for left
joins. Now the extra restriction is appended directly to the join
condition instead of the where clause.
To generate the extra condition we need to have the join field
available in the compiler. This has the side-effect that we need more
ugly code in Query.__getstate__ and __setstate__ as Field objects
aren't pickleable.
The join trimming code got a big change - now we trim all direct joins
and never trim reverse joins. This also fixes the problem in #10790
which was join trimming in null filter cases.
The trim argument was used by split_exclude() only to trim the last
join from the given lookup. It is cleaner to just trim the last part
from the lookup in split_exclude() directly so that there is no need
to burden add_filter() with the logic needed for only split_exclude().
Connection creation was done in db backend ._cursor() call. This
included taking a new connection if needed, initializing the session
state for the new connection and finally creating the connection.
To allow easier modifying of these steps in subclasses (for example to
support connection pools) the _cursor() now calls get_new_connection()
and init_connection_state() if there isn't an existing connection. This
was done for all non-gis core backends. In addition the parameters used
for taking a connection are now created by get_connection_params().
We should also do the same for gis backends and encourage 3rd party
backends to use the same pattern. The pattern is not enforced in code,
and as the backends are private API this will not be required by
documentation either.
F() expressions reuse joins like any lookup in a .filter() call -
reuse multijoins generated in the same .filter() call else generate
new joins. Also, lookups can now reuse joins generated by F().
This change is backwards incompatible, but it is required to prevent
dict randomization from generating different queries depending on
.filter() kwarg ordering. The new way is also more consistent in how
joins are reused.
The select_related code got confused when it needed to travel a
reverse relation to a model which had different parent than the
originally travelled relation.
Thanks to Trac aliases shauncutts for report and ungenio for original
patch (committed patch is somewhat modified version of that).
The problem is the same as in #10888 which was reintroduced when
bulk_insert was added. Thanks to Jani Tiainen for report, patch and
also testing the final patch on Oracle GIS.
Simon Charette
Anssi Kääriäinen
Claude Paroz
mpaolini
Aymeric Augustin
Russell Keith-Magee
orblivion
Tai Lee
Mike Johnson