Optimizing DB query involving annotate and aggregate

[ST]

Hi,

I'm trying to optimize the run-time of getting total credit and debt values out of our database. Ideally I'd like to formulate it as a Django query. This is the raw SQL query I have, which produces the right answer and is very fast (milliseconds):

SELECT sum(tg.total) FROM

  (

  SELECT sum(t.amount) AS total, t.member_id AS member_id

  FROM club_transaction AS t

  WHERE t.member_id IS NOT NULL

  GROUP BY t.member_id

  ) AS tg

WHERE tg.total < 0

(plus a second query for > 0)

My Django implementation was:

        m = Member.objects.annotate(balance=Sum('transaction__amount'))

        m_debt = m.filter(balance__lt=0).aggregate(total=Sum('balance'))

        m_credit = m.filter(balance__gt=0).aggregate(total=Sum('balance'))

which looks a lot nicer, is easier to understand and maintain.

However, it results in the following SQL query (slightly redacted):

SELECT SUM(balance) FROM

  (

  SELECT "club_member"."id" AS "id", {all the other fields}, SUM("club_transaction"."amount") AS "balance"

  FROM "club_member"

  LEFT OUTER JOIN "auth_user" ON ("club_member"."user_id" = "auth_user"."id")

  LEFT OUTER JOIN "club_transaction" ON ("club_member"."id" = "club_transaction"."member_id")

  GROUP BY "club_member"."id", {all the other fields}, "auth_user"."last_name", "auth_user"."first_name"

  HAVING SUM("club_transaction"."amount") < 0

  ORDER BY "auth_user"."last_name" ASC, "auth_user"."first_name" ASC

  ) subquery

(again, plus another one for > 0)

which is very slow (almost 1.5 seconds).

How can I construct a Django query which doesn't request (and group by) all the unnecessary other fields ?

I already tried playing around with only() and values() but never got it to work.

Looking forward to your responses!

best regards,

ST

[Anssi Kääriäinen]

Something like this might work:
    Transaction.objects.values_list('member_id').annotate(total=Sum('amount')).filter(total__gt=0).aggregate(Sum('total'))

That is, don't start from Member, Django isn't smart enough to get rid of the non-necessary joins. Instead go directly for the query you wrote in SQL. In the ORM .values() is the way to control GROUP BY clause. Getting the order of .values() and .annotate() right is important, and it is sometimes hard to figure out how to do it correctly. Also, if you have some extra ordering going on in the query just issue an empty .order_by() call - that will get rid of all ordering.

 - Anssi

[ST]

On Wednesday, February 5, 2014 8:01:53 PM UTC, Anssi Kääriäinen wrote:

Something like this might work:
    Transaction.objects.values_list('member_id').annotate(total=Sum('amount')).filter(total__gt=0).aggregate(Sum('total'))

This didn't work - it produced a "SELECT  FROM" query, which obviously didn't work - tried adding 'amount' to the values_list, but that didn't help either. Eventually got it to work by using .only('member', 'amount') instead, and it *was* fast, but it didn't quite do what we need: we only want the *outstanding* credit/debt, i.e. some member might have many transactions, but they all add up to zero - so instead of increasing both total_debt and total_credit by that value, it shouldn't affect the totals. Which is why I wanted to total up by members first... any idea how to get that to work ?

[Erik Cederstrand]

Den 06/02/2014 kl. 12.34 skrev ST <sej...@googlemail.com>:

> This didn't work - it produced a "SELECT FROM" query, which obviously didn't work - tried adding 'amount' to the values_list, but that didn't help either. Eventually got it to work by using .only('member', 'amount') instead, and it *was* fast, but it didn't quite do what we need: we only want the *outstanding* credit/debt, i.e. some member might have many transactions, but they all add up to zero - so instead of increasing both total_debt and total_credit by that value, it shouldn't affect the totals. Which is why I wanted to total up by members first... any idea how to get that to work ?

Instead of spending a lot of time trying to bend the Django ORM to do what you want, just keep the raw SQL that you know works, and add tests to your test suite. The ORM is really nice, but sometimes it’s just not worth the hassle.

Erik

[Arnold Krille]

On Wed, 5 Feb 2014 10:11:29 -0800 (PST) ST <sej...@googlemail.com>
wrote:

I did something similar a few years back. Its next to impossible with
django-orm, just do it in raw sql.
The django-orm can't really do any advanced group_by clauses. And you
can't give them with extra(). Well, you can give them but they won't be
used;-)

Doing two (or three) orm-queries and then joining the data in python
will actually be slower then doing it all in hand-crafted sql.

So just do the sql by hand. And then optimize by having three columns,
one with the SUM(amount) if amount>0, one with the SUM(amount) if
amount<0 and one with the SUM(amount). Total credits, total depts and
balance all in one query (if possible)...

- Arnold

[ST]

Managed it in the end, so for reference here is what I ended up with:

def total_credit_debt(self):

        transaction_totals = Transaction.objects.filter(member__isnull=False).values('member').annotate(total=Sum('amount')).order_by()

        creditors = transaction_totals.filter(total__gt=0)

        debtors = transaction_totals.filter(total__lt=0)

        total_credit = sum(row['total'] for row in creditors)

        total_debt = sum(row['total'] for row in debtors)

        return total_credit, total_debt

(@classmethod on TransactionManager)

Unfortunately I didn't manage to get it to do the second sum inside the database as well, but this is fast enough - there's only a few hundred members with non-zero balances. If you think there's any way this could be improved / cleaned up, I'd be happy to hear:)

Thanks!