Subquery expressions unnecessarily added to group by

[Jonny Fuller]

Hi friends,

My Django/SQL skills are not good enough to properly isolate the problem independently of my use case detailed below. I believe the problem is Subqueries being forced in to the group_by clause because they are select expressions. Per the below if I remove the subqueries manually from sql group_by, my query runs perfectly. I tried to manually edit the qs.query.group_by, but because it is created by compiler.get_group_by() I cannot fix the ORM query group by clause.

Are Subquery expressions always supposed to be included in group_by? If this is desired behavior is it possible to toggle it off so the ORM can produce the accurate query?

"""
Problem Statement: The individual annotations work fine when run independently, but when chained the query takes 5 minutes. This is due to the final group by clause unexpectedly receiving the Subquery as extra fields.
"""
​
# relevant models and querysets
class ClaimQuerySet(models.QuerySet):
    def annotate_all(self):
        """Adds ``results``, ``latest_note_text``, and ``latest_assessment_text`` to the queryset."""
        return self.annotate_latest_results().annotate_most_recent_note().annotate_most_recent_assessment()
​
    def prefetch_all(self, annotate_sum=True):
        return self.prefetch_notes().prefetch_latest_results(annotate_sum)
​
    def prefetch_latest_results(self, annotate_sum: bool=True):
        """Prefetches the most result :class:`RulesEngineResult` object and optionally
        preload its :attr:`RulesEngineResult.score`.
​
        Args:
            annotate_sum:
        """
        latest_runs = self.latest_runs
        if annotate_sum:
            latest_runs = latest_runs.annotate(_score=Sum('results__value'))
        return self.prefetch_related(Prefetch(
            'rules_engine_results', queryset=latest_runs, to_attr='_last_run')
        )
​
    def prefetch_notes(self):
        """Prefetches all related notes and assessments."""
        return self.prefetch_related('notes', 'assessments')
​
    @property
    def latest_runs(self):
        """Shortcut for :attr:`RulesEngineResultQuerySet.get_latest_runs`"""
        return RulesEngineResult.objects.get_latest_runs()
​
    def annotate_latest_results(self) -> 'ClaimQuerySet':
        """Annotates the queryset with a new field ``results`` whose value is the Sum
        of the last :attr:`RulesEngineResult.results` for the claim.
        """
        # Only Sum on runs in the above set.
        filter_q = Q(rules_engine_results__in=self.latest_runs)
        # noinspection PyTypeChecker
        return self.annotate(results=Sum('rules_engine_results__results__value', filter=filter_q))
​
    def annotate_most_recent_note(self) -> 'ClaimQuerySet':
        """Annotates the queryset with a field ``latest_note_text`` whose value is the last
        entered :attr:`Note.text` for the claim or ``None`` if there are no associated notes.
        """
        return self._annotate_most_recent_basenote(Note, 'latest_note_text')
​
    def annotate_most_recent_assessment(self) -> 'ClaimQuerySet':
        """Annotates the queryset with a field ``latest_assessment_text`` whose value is the last
        entered :attr:`Assessment.text` for the claim or ``None`` if there are no associated assessments.
        """
        return self._annotate_most_recent_basenote(Assessment, 'latest_assessment_text')
​
    def _annotate_most_recent_basenote(self, model: Type['BaseNote'], field_name: str) -> 'ClaimQuerySet':
        newest = model.objects.filter(claim=OuterRef('id')).order_by('-created')
        annotate_kwargs = {
            field_name: Subquery(newest.values('text')[:1])
        }
        # noinspection PyTypeChecker
        return self.annotate(**annotate_kwargs)
​
​
class Claim(BaseClaim):
    """Concrete :class:`~mkl.fraud_django.models.BaseClaim` for
    :mod:`~mkl.fraud_django.workers_comp` claims.
    """
    objects = ClaimQuerySet.as_manager()
    first_rules_engine_run = models.DateField()
​
    @property
    def latest_note(self) -> 'Note':
        """Returns the latest :class:`Note`."""
        return self.notes.latest()
​
    @property
    def latest_assessment(self) -> 'Assessment':
        """Retrieves the latest :class:`Assessment`."""
        return self.assessments.latest()
​
    @property
    def latest_rulesengine_run(self) -> 'RulesEngineResult':
        """Returns the most most recent run.
​
        .. note::
​
            Use :meth:`ClaimQuerySet.prefetch_latest_results` to
            prefetch the last_run, falls back on querying for the latest value.
​
            Note, if used in a prefetched queryset the value could be stale.
        """
        return self._get_latest(RulesEngineResult, '_last_run')
​
    def _get_latest(self, model: Type[models.Model], cache_attr: str):
        """Handler to return None if a latest related object does not exist,
        checks the cache first."""
        if hasattr(self, cache_attr):
            try:
                return getattr(self, cache_attr)[0]
            except IndexError:
                return None
        try:
            return model.objects.filter(claim=self).latest()
        except model.DoesNotExist:
            return None
​
    def __unicode__(self):
        return self.claim_number
​
​
class BaseNote(models.Model):
    """Abstract Base Model for both Notes and Assessments.
​
    Use this base for any claim related editable field whose
    historical data is important.
​
    On the claim we can write functions to retrieve the latest.
​
    .. note:: The related name will be the class name lower case with an 's'.
​
    Attributes:
        text (str): The user provided content
        created (datetime.datetime): Created time stamp
        claim (:class:`Claim`): The claim related to the note.
    """
    id = models.AutoField(primary_key=True)
    text = models.TextField(max_length=1000)
    created = models.DateTimeField(auto_now_add=True)
    claim = models.ForeignKey('Claim', on_delete=models.PROTECT, related_name='%(class)ss')
​
    class Meta:
        abstract = True
        get_latest_by = 'created'
        ordering = ('-created',)
​
​
class Note(BaseNote):
    """Concrete class for Notes, related_name will become ``notes``."""
​
​
class Assessment(BaseNote):
    """Concrete class for Assessment, related_name will become ``assessments``."""
    CHOICES = (
        ('01', 'Will open a case'),
        ('02', 'Will not open a case'),
        ('03', 'Previously opened'),
        ('04', "Appears suspicious but won't open"),
        ('05', 'Not enough info to determine'),
        ('06', 'Existing vendor request'),
    )
    text = models.CharField(max_length=1000, choices=CHOICES)
​
    def get_choice_value(self) -> str:
        """Returns the value as the choice human readable text."""
        db_text = self.text
        return dict(self.CHOICES)[db_text]
​
​
class RuleResult(models.Model):
    """The result of running the engine for a particular claim against a :class:`Rule`.
​
    Attributes:
        rule: The rule to be checked
        value: The numeric weight of the result
        result: The rules engine result of all rules run against the claim
    """
    id = models.AutoField(primary_key=True)
    rule = models.ForeignKey('Rule', on_delete=models.PROTECT)
    value = models.IntegerField()
    result = models.ForeignKey('RulesEngineResult', on_delete=models.PROTECT, related_name='results')
​
​
class RulesEngineResultQuerySet(models.QuerySet):
    def get_latest_runs(self):
        """Filters to only the most recent :class:`RulesEngineResult`\s."""
        annotated = self.annotate(
            latest=Max('claim__rules_engine_results__created')
        )
        return annotated.filter(created=F('latest'))
​
​
class RulesEngineResult(models.Model):
    """
    RulesEngine run result.
​
    Attributes:
        claim (:class:`Claim`): The claim run through the RulesEngine.
        results (List[:class:`RuleResult`]): Collection of results for each rule.
    """
    id = models.AutoField(primary_key=True)
    created = models.DateTimeField(auto_now_add=True)
    claim = models.ForeignKey('Claim', on_delete=models.PROTECT, related_name='rules_engine_results')
    objects = RulesEngineResultQuerySet.as_manager()
​
    class Meta:
        get_latest_by = 'created'
​
    @property
    def score(self) -> int:
        """Returns the aggregate score of all related results. Checks prefetched cache first."""
        if hasattr(self, '_score'):
            return self._score
        d = self.results.aggregate(score=models.Sum('value'))
        return d['score']
​
​
"""
Individual Query rendering
"""
# Recent Note
qs = models.Claim.objects.annotate_most_recent_note()
​
SELECT "workers_comp_claim"."id",
       "workers_comp_claim"."claim_number",
       "workers_comp_claim"."first_rules_engine_run",
       (SELECT U0."text"
        FROM "workers_comp_note" U0
        WHERE U0."claim_id" = ("workers_comp_claim"."id")
        ORDER BY U0."created" DESC
        LIMIT 1) AS "latest_note_text"
FROM "workers_comp_claim"
​
# Recent Assessment
qs = models.Claim.objects.annotate_most_recent_assessment()
​
SELECT "workers_comp_claim"."id",
       "workers_comp_claim"."claim_number",
       "workers_comp_claim"."first_rules_engine_run",
       (SELECT U0."text"
        FROM "workers_comp_assessment" U0
        WHERE U0."claim_id" = ("workers_comp_claim"."id")
        ORDER BY U0."created" DESC
        LIMIT 1) AS "latest_assessment_text"
FROM "workers_comp_claim"
​
# Latest Results (Run)
qs = models.Claim.objects.annotate_latest_results()
​
SELECT "workers_comp_claim"."id",
       "workers_comp_claim"."claim_number",
       "workers_comp_claim"."first_rules_engine_run",
       SUM("workers_comp_ruleresult"."value") FILTER (WHERE "workers_comp_rulesengineresult"."id" IN (SELECT U0."id"
                                                                                                      FROM "workers_comp_rulesengineresult" U0
                                                                                                             INNER JOIN "workers_comp_claim" U1 ON (U0."claim_id" = U1."id")
                                                                                                             LEFT OUTER JOIN "workers_comp_rulesengineresult" U2 ON (U1."id" = U2."claim_id")
                                                                                                      GROUP BY U0."id"
                                                                                                      HAVING U0."created" = (MAX(U2."created")))) AS "results"
FROM "workers_comp_claim"
       LEFT OUTER JOIN "workers_comp_rulesengineresult"
                       ON ("workers_comp_claim"."id" = "workers_comp_rulesengineresult"."claim_id")
       LEFT OUTER JOIN "workers_comp_ruleresult"
                       ON ("workers_comp_rulesengineresult"."id" = "workers_comp_ruleresult"."result_id")
GROUP BY "workers_comp_claim"."id"
​
​
"""
When chained the query renders incorrectly like this
"""
qs = models.Claim.objects.annotate_latest_results().annotate_most_recent_note().annotate_most_recent_assessment()
​
SELECT "workers_comp_claim"."id",
       "workers_comp_claim"."claim_number",
       "workers_comp_claim"."first_rules_engine_run",
       SUM("workers_comp_ruleresult"."value") FILTER (WHERE "workers_comp_rulesengineresult"."id" IN (SELECT U0."id"
                                                                                                      FROM "workers_comp_rulesengineresult" U0
                                                                                                             INNER JOIN "workers_comp_claim" U1 ON (U0."claim_id" = U1."id")
                                                                                                             LEFT OUTER JOIN "workers_comp_rulesengineresult" U2 ON (U1."id" = U2."claim_id")
                                                                                                      GROUP BY U0."id"
                                                                                                      HAVING U0."created" = (MAX(U2."created")))) AS "results",
       (SELECT U0."text"
        FROM "workers_comp_note" U0
        WHERE U0."claim_id" = ("workers_comp_claim"."id")
        ORDER BY U0."created" DESC
        LIMIT 1)                                                                                                                                  AS "latest_note_text",
       (SELECT U0."text"
        FROM "workers_comp_assessment" U0
        WHERE U0."claim_id" = ("workers_comp_claim"."id")
        ORDER BY U0."created" DESC
        LIMIT 1)                                                                                                                                  AS "latest_assessment_text"
FROM "workers_comp_claim"
       LEFT OUTER JOIN "workers_comp_rulesengineresult"
                       ON ("workers_comp_claim"."id" = "workers_comp_rulesengineresult"."claim_id")
       LEFT OUTER JOIN "workers_comp_ruleresult"
                       ON ("workers_comp_rulesengineresult"."id" = "workers_comp_ruleresult"."result_id")
GROUP BY "workers_comp_claim"."id", (SELECT U0."text"
                                     FROM "workers_comp_note" U0
                                     WHERE U0."claim_id" = ("workers_comp_claim"."id")
                                     ORDER BY U0."created" DESC
                                     LIMIT 1), (SELECT U0."text"
                                                FROM "workers_comp_assessment" U0
                                                WHERE U0."claim_id" = ("workers_comp_claim"."id")
                                                ORDER BY U0."created" DESC
                                                LIMIT 1)
​
​
"""
Why is Django performing the group by with the Subqueries? How do I make it render correctly like this:
"""
SELECT "workers_comp_claim"."id",
       "workers_comp_claim"."claim_number",
       "workers_comp_claim"."first_rules_engine_run",
       SUM("workers_comp_ruleresult"."value") FILTER (WHERE "workers_comp_rulesengineresult"."id" IN (SELECT U0."id"
                                                                                                      FROM "workers_comp_rulesengineresult" U0
                                                                                                             INNER JOIN "workers_comp_claim" U1 ON (U0."claim_id" = U1."id")
                                                                                                             LEFT OUTER JOIN "workers_comp_rulesengineresult" U2 ON (U1."id" = U2."claim_id")
                                                                                                      GROUP BY U0."id"
                                                                                                      HAVING U0."created" = (MAX(U2."created")))) AS "results",
       (SELECT U0."text"
        FROM "workers_comp_note" U0
        WHERE U0."claim_id" = ("workers_comp_claim"."id")
        ORDER BY U0."created" DESC
        LIMIT 1)                                                                                                                                  AS "latest_note_text",
       (SELECT U0."text"
        FROM "workers_comp_assessment" U0
        WHERE U0."claim_id" = ("workers_comp_claim"."id")
        ORDER BY U0."created" DESC
        LIMIT 1)                                                                                                                                  AS "latest_assessment_text"
FROM "workers_comp_claim"
       LEFT OUTER JOIN "workers_comp_rulesengineresult"
                       ON ("workers_comp_claim"."id" = "workers_comp_rulesengineresult"."claim_id")
       LEFT OUTER JOIN "workers_comp_ruleresult"
                       ON ("workers_comp_rulesengineresult"."id" = "workers_comp_ruleresult"."result_id")
GROUP BY "workers_comp_claim"."id";

[Tim Graham]

See TicketClosingReasons/UseSupportChannels for places to ask "is it a bug/how do I?" questions. Most likely you'll need to simplify your problem. If you find Django at fault, please create a more concise, minimal ticket.

[Simon Charette]

Additional details about which database engine and Django version you are using would be also appreciated.

[Simon Charette]

I suspect Subquery.get_group_by_cols needs to be adjusted to return an empty list. Since #30099 (2.2a1+ only) ​it returns [self] but before this change ​it returned all the filter conditions with a left hand side which is even worst.

Could you try subclassing Subquery and overriding get_group_by_cols to return [], use the subclass with your query, and see if it helps?

[Jonny Fuller]

Replying to Simon Charette:

Additional details about which database engine and Django version you are using would be also appreciated.

Hi Simon. I am using Django 2.1.5 and Postgre.

[Jonny Fuller]

Replying to Simon Charette:

I suspect Subquery.get_group_by_cols needs to be adjusted to return an empty list. Since #30099 (2.2a1+ only) ​it returns [self] but before this change ​it returned all the filter conditions with a left hand side which is even worst.

Could you try subclassing Subquery and overriding get_group_by_cols to return [], use the subclass with your query, and see if it helps?

Wow I don't know what black magic that did but it works!

[Simon Charette]

I just ran the full suite with the get_group_by_cols override detailed and it worked fine, this case is simply untested.

Jonny, would you be interested in submitting a Github PR with the changes suggested above with the addition of a regression test to make sure the GROUP BY doesn't occur? I'd be happy to offer assistance like I did for #30099 in ​https://github.com/django/django/pull/10846.

[Jonny Fuller]

Replying to Simon Charette:

I just ran the full suite with the get_group_by_cols override detailed and it worked fine, this case is simply untested.

Jonny, would you be interested in submitting a Github PR with the changes suggested above with the addition of a regression test to make sure the GROUP BY doesn't occur? I'd be happy to offer assistance like I did for #30099 in ​https://github.com/django/django/pull/10846.

Sure thing Simon. I just need to figure out how to write a generic test case...

[Simon Charette]

This happens to be a bit related to #29542.

[Simon Charette]

Jonny, by playing a bit with the code I noticed that it's a bit more complex than simply returning an empty list in Subquery.get_group_by_cols.

Can you confirm that both queries were returning the appropriate results but that the one where subqueries were added to the GROUP BY was performing significantly slower?

If that's the case then this is more of an optimization problem where subqueries can sometimes be removed from the GROUP BY and sometimes not but not in all cases.

e.g.

Publisher.objects.annotate(
    has_long_books=Exists(
        Book.objects.filter(
            publisher=OuterRef('pk'),
            pages__gt=800,
        ),
    ),
).values_list('has_long_books').annotate(
    total=Count('*'),
)

When there's an explicit grouping by a subquery (or exists) then it must be honoured but I believe in other cases it's doesn't have to.

[Simon Charette]

I renamed the ticket because I'm pretty confident that the generated query doesn't cause the wrong results to be returned.

The fact that the ORM doesn't have the introspection abilities to determine if some annotated expressions can be trimmed from the GROUP BY clause on aggregation ​is a known limitation and there's probably other existing issues for other type of expressions.

Since the initial approach of making get_group_by_cols return an empty list is not viable for the aforementioned reason I think we should begin by merging ​regression tests to prevent this edge case from being overlooked in the future.

[Simon Charette]

Finally came up with an elegant solution. By changing Expression.get_group_by_cols's signature to accept an optional alias parameter that is only provided when the annotated expression is being explicitly grouped against. That allows Subquery.get_group_by_cols to either return a reference to it's selected alias when grouped against or no columns at all when it's not the case which is what this ticket is tracking.

[Tim Graham <timograham@…>]

In 9dc367d:

Refs #30158 -- Added alias argument to Expression.get_group_by_cols().

[Tim Graham <timograham@…>]

In fb3f034f:

Fixed #30158 -- Avoided unnecessary subquery group by on aggregation.

Subquery annotations can be omitted from the GROUP BY clause on aggregation
as long as they are not explicitly grouped against.

Thanks Jonny Fuller for the report.

[Tim Graham <timograham@…>]

In e595a713:

Refs #29542, #30158 -- Enabled a HAVING subquery filter test on Oracle.

Now that subquery annotations aren't included in the GROUP BY unless
explicitly grouped against, the test works on Oracle.

[GitHub <noreply@…>]

In 350123f:

Moved release note for refs #30158 from deprecated to backwards incompatible changes.

[Mariusz Felisiak <felisiak.mariusz@…>]

In 5e33ec80:

Refs #30158 -- Made alias argument required in signature of Expression.get_group_by_cols() subclasses.

Per deprecation timeline.

[Mariusz Felisiak <felisiak.mariusz@…>]

In c6350d5:

Refs #30158 -- Removed alias argument for Expression.get_group_by_cols().

Recent refactors allowed GROUP BY aliasing allowed for aliasing to be
entirely handled by the sql.Query.set_group_by and compiler layers.