Inconsistent results of QuerySet count() method using PostgreSQL backend prior and post the QuerySet evaluation

[kamandol]

Using a PostgreSQL backend database, in some cases with QuerySets involving annotations, values or values_list and ordering(order_by) clauses the resulting QuerySet count() method fails to predict the real row result unless that QuerySet is evaluated.

For instance:

q = BundleVersion.objects.order_by('-id').values_list("port__id", "asset_bundle__id").annotate(max=models.Max("id"))

This QuerySet using PostgreSQL would, in fact, group the results by BundleVersion.id instead of the tuple ("port__id", "asset_bundle__id") due to the order_by clause using a column not in the values_list. This is a documented behavior.

The problem is that if q is not evaluated yet, calling q.count() will return an amount of rows as if the grouping was done on the tuple ("port__id", "asset_bundle__id"). After evaluating q, or calling len(q), the result of q.count() will be the expected, as if the grouping was done using 'BundleVersion.id'.

[Tim Graham]

Could you provide a test for Django's test suite in tests/queries/tests.py that demonstrates the issue?

[kamandol]

Replying to timgraham:

Could you provide a test for Django's test suite in tests/queries/tests.py that demonstrates the issue?

Here is the commit with the demonstration:

​https://github.com/sp-ricard-valverde/django/commit/c50afafdb02880e4c941c2a72a215bee80de3aed

[Iacopo Spalletti]

Tests looks legitimate, but I don't get this one ​https://github.com/sp-ricard-valverde/django/commit/c50afafdb02880e4c941c2a72a215bee80de3aed#diff-ce9b52a66d03e851a9828377263dc04bR3864 ; or rather I don't get why is labeled as 'OK' as it the concatenation of the previous ones

[kamandol]

Replying to yakky:

Tests looks legitimate, but I don't get this one ​https://github.com/sp-ricard-valverde/django/commit/c50afafdb02880e4c941c2a72a215bee80de3aed#diff-ce9b52a66d03e851a9828377263dc04bR3864 ; or rather I don't get why is labeled as 'OK' as it the concatenation of the previous ones

Sorry, my bad, it's a bad choice for a test title. What it demonstrates is that the second assert in that test is successful even though a new QuerySet is built and, according to the other failed tests it should fail too.

[kamandol]

Updated test comments and clearer test names

​https://github.com/sp-ricard-valverde/django/commit/42c1002a00b434331fd2c9970b05baf02f0a262f

[Tim Graham]

If you could put the tests on a branch of you fork other than "master" that seems a bit safer. There is probably no need for skipUnlessDBEngine - is there a reason the queries shouldn't also work on other databases? Anyway, the usual way to skip tests is based on database features.

[kamandol]

I will change the branch as requested.

You're right, the faulty behavior was detected on PostgreSQL, but on MySQL worked correctly so the test should pass. Could not test other db engines though. I will remove that extra piece of code.

[Can Sarıgöl]

​PR

[Can Sarıgöl]

I want to discuss when collecting group by columns why are order by columns added into extensions.

For example:
Model.objects.values("col_a").annotate(max=Max("col_b")).order_by('col_c')
current query:

select col_a, MAX(col_b) as max 
from Model 
group by col_a, col_c
order by col_c

I think the expected behavior has should be like that:
django.db.utils.ProgrammingError: column "Model.col_c" must appear in the GROUP BY clause or be used in an aggregate function

Because this behavior changes my group by and result. if I'm lucky, my result doesn't change.

If we check test_annotation_with_value, 'name' column is being added into group_by because Book model has Meta.ordering.