Matt Robenolt

Stalk me:

The Django ORM & Subqueries

Note: At the time of writing, Django is at v1.4.1

So what actually happens when you have a Django query like the following?

cities = City.objects.all()
venues = Venue.objects.filter(city__in=cities)

Django is smart enough to know that you want to use the results from the first unevaluated QuerySet as the input for the __in clause, and in turn, generates a subquery. The raw query generated is something like this:

SELECT ...
FROM   "venue"
WHERE  "venue"."city_id"
IN (
    SELECT U0."id"
    FROM "city" U0
) 

Awesome. Django saved us an extra round trip by inlining the query instead of evaluating the first query separately.

What's wrong?

In this scenario, nothing at all. We've saved a round trip to the database, all is good. Let's throw caching into the mix.

cities = cache.get('cities')
if cities is None:
  cities = City.objects.all()
  cache.set('cities', cities)
venues = Venue.objects.filter(city__in=cities)

In this scenario, we want to cache the entire list of cities. The City QuerySet gets cached for us like we expect, but when Django gets to the filter(), it sees that you're passing it a QuerySet and generates a subquery again. We've effective cached nothing because the exact same query is being run.

This also applies when you want to use the same __in across multiple filters.

cities = City.objects.all()
venues = Venue.objects.filter(cityin=cities)
users = UserAccount.objects.filter(cityin=cities)

Again, the exact same subquery is repeated across both queries.

Well, how can we fix this?

The reason why Django is doing this is solely because the argument passed into filter() is a QuerySet. If we just pass a list, Django will use an array of ids inside the IN SQL.

Let's look at an improved version with caching.

cities = cache.get('cities')
if cities is None:
  cities = list(City.objects.all())  # Force query evaluation
  cache.set('cities', cities)
venues = Venue.objects.filter(city__in=cities)

The only thing that has changed is forcing the QuerySet to be evaluated by converting it to a list. Now, cities is a list, so Django won't try to perform a subquery anymore. This could be simplified to just caching the primary keys as a list as well, depending on what you're trying to achieve.

The resulting SQL will be something like:

SELECT ...
FROM   "venue"
WHERE  "venue"."city_id"
IN (
    1, 2, 3
) 

One more gotcha!

We all know that Django's QuerySets are lazy, right? No query is actually performed until the QuerySet is iterated over. Once a QuerySet has been iterated over, it won't query again for the results. The results get cached into an internal _results_cache object.

At the moment, Django is not smart enough to detect that a QuerySet has already been evaluated before generating a subquery, effectively causing the query to be run again when we already know the results.

Picture this scenario:

cities = City.objects.all()
names = [city.name for city in cities]  # QuerySet cache is filled
# ... do something awesome ...
venues = Venue.objects.filter(city__in=cities)

Since we've already evaluated the QuerySet before passing it to the __in clause, it'd be smart to just use the ids that we've already calculated from before, but it doesn't. So pay attention and be careful.

I've submitted a patch to Django core to try and get that behavior fixed: