Query resolution can be at least 3x slower in 3.2

[Phillip Cutter]

I recently upgraded to Django 3.2 from 3.1.8 and found running my queries were taking nearly 15 seconds to run compared to a few milliseconds in 3.1.8. To help bring this issue to light I've created a basic Django project on GitHub with a test case that shows this issue. Running "manage.py test" on Django 3.2 and Django 3.1.8 shows that 3.2 is taking roughly 3x as long to build the query, while not executing the query.

Repository with test case: ​https://github.com/mrfleap/django-db-test

The underlying query is admittedly repetitive, but I still feel it this degradation in performance should not go overlooked. Through some basic debugging I believe the issue may lie django/utils/tree.py on line 93 (​https://github.com/django/django/blob/stable/3.2.x/django/utils/tree.py#L93) where it checks to see if a node is within a list of children, one of my basic tests found the operation to take 9 seconds alone, rather than the fractions of a millisecond it usually takes. I've attached a screenshot of my breakpoint to showcase that it's taking ~9 seconds to determine if one node exists in an array of two.

I'd also like to point out that I am using Django Guardian quite heavily in this scenario, which is leading to the complicated query in the first place. However, as far as I am aware, there is no difference in the generated query between both versions, and the difference in speed is based on how Django is resolving the internal objects into a SQL statement.

Please let me know if there's any other information I can provide to help fix this issue.

[Phillip Cutter]

Screenshot showing breakpoint set during query building in a separate repository

[Mariusz Felisiak]

Thanks for the report. Tentatively accepted, I'm not sure if it's feasible to improve a performance and keep the fix for #32143.

Performance regression in bbf141bcdc31f1324048af9233583a523ac54c94.
Reproduced at 1351f2ee163145df2cf5471eb3e57289f8853512.

[Simon Charette]

Alright so I did a bit of digging in the past hour and here are my findings

Before bbf141bcdc31f1324048af9233583a523ac54c94 Query.__eq__ was wrongly returning True due to making it a subclass BaseExpression in 35431298226165986ad07e91f9d3aca721ff38ec. Before these changes Query.__eq__ was not implemented which meant that Model.objects.filter(foo=bar).query != Model.objects.filter(foo=bar).query as object.__eq__ is based of id(self) == id(other). This is problematic for expressions such as Exists and Subquery which are based off sql.Query and need to be able to compare to each other to implement the expression API.

Since sql.Query is an immutable internal object we could possibly cached_property(identity) and move along as that cuts the slowdown in about half, that does feel like a bandaid solution though. It's interesting to me that Node.add would compare all its children in the first place as that seems like a costly operation with little benefit in the first place.

After playing around with a few things here's a list of two proposed solutions

1. Remove the tree.Node.add branch for if data in self.children. While ​it's covered by the suite only ​a single test in the whole suite reaches it the branch under arguably convoluted conditions while the check is performed 99% of the time when filter and exclude are called to perform expensive comparisons of contained expressions. I wouldn't be surprised if it was at the origin of others recent slowdowns in the ORM query construction due to the introduction of BaseExpression.identity to needs for deconstruc-ability imposed by Index(condition) and Constraint(condition) for little benefit.
2. Make Subquery (and Exists) explicitly non-deconstructible; in practice they already are because sql.Query doesn't properly implement deconstruct so there's little value in implementing identity for them in the first place.
3. Remove the sql.Query(BaseExpression) subclassing as it's no longer necessary in practice. That'll have the side effect of removing support for Model.objects.filter(foo=bar).query == Model.objects.filter(foo=bar).query which is a nice property but since we don't internally have a need for it yet there's little point in the complexity it incurs.

I'd suggest we backport 2. and 3. to address this ticket and move the discussion about 1. in a new Cleanup/optimization one. Let me know if that makes sense to you and I'll commit to a PR before the end of April.

[Mariusz Felisiak]

Thanks for the investigation!

2 & 3 make sense for me, however I would like to see an implementation first. This can be too massive for backporting. I'm a bit afraid that we can open the Pandora's box.

Will we be able to keep support for combining Q() objects ​with boolean expressions?

[Simon Charette]

2 & 3 make sense for me, however I would like to see an implementation first. This can be too massive for backporting. I'm a bit afraid that we can open the Pandora's box.

Hopefully the proposed patch is not too invasive, I could try to figure out another approach. From an invasivity reduction perspective removing the Node.add branch seems like a clear choice for this particular issue but I wouldn't be surprised if other code paths were affected by the slowness introduced by sql.Query.__eq__.

Will we be able to keep support for combining Q() objects ​with boolean expressions?

Yep, it shouldn't be an issue. The main change is really that Subquery doesn't partially implement the deconstruct API which is only really useful for migrations anyway.

​PR

[Mariusz Felisiak <felisiak.mariusz@…>]

In 4f600673:

Refs #32632 -- Added tests for returning a copy when combining Q() objects.

[Mariusz Felisiak <felisiak.mariusz@…>]

In c8b65943:

Fixed #32632, Fixed #32657 -- Removed flawed support for Subquery deconstruction.

Subquery deconstruction support required implementing complex and
expensive equality rules for sql.Query objects for little benefit as
the latter cannot themselves be made deconstructible to their reference
to model classes.

Making Expression @deconstructible and not BaseExpression allows
interested parties to conform to the "expression" API even if they are
not deconstructible as it's only a requirement for expressions allowed
in Model fields and meta options (e.g. constraints, indexes).

Thanks Phillip Cutter for the report.

This also fixes a performance regression in bbf141bcdc31f1324048af9233583a523ac54c94.

[Mariusz Felisiak <felisiak.mariusz@…>]

In d5add5d3:

[3.2.x] Fixed #32632, Fixed #32657 -- Removed flawed support for Subquery deconstruction.

Subquery deconstruction support required implementing complex and
expensive equality rules for sql.Query objects for little benefit as
the latter cannot themselves be made deconstructible to their reference
to model classes.

Making Expression @deconstructible and not BaseExpression allows
interested parties to conform to the "expression" API even if they are
not deconstructible as it's only a requirement for expressions allowed
in Model fields and meta options (e.g. constraints, indexes).

Thanks Phillip Cutter for the report.

This also fixes a performance regression in bbf141bcdc31f1324048af9233583a523ac54c94.

Backport of c8b659430556dca0b2fe27cf2ea0f8290dbafecd from main