In Django, what is the most efficient way to check for an empty query set?

Question

I've heard suggestions to use the following:

if qs.exists():
    ...

if qs.count():
    ...

try:
    qs[0]
except IndexError:
    ...

Copied from comment below: "I'm looking for a statement like "In MySQL and PostgreSQL count() is faster for short queries, exists() is faster for long queries, and use QuerySet[0] when it's likely that you're going to need the first element and you want to check that it exists. However, when count() is faster it's only marginally faster so it's advisable to always use exists() when choosing between the two."

Answer 1

query.exists() is the most efficient way.

Especially on postgres count() can be very expensive, sometimes more expensive then a normal select query.

exists() runs a query with no select_related, field selections or sorting and only fetches a single record. This is much faster then counting the entire query with table joins and sorting.

qs[0] would still includes select_related, field selections and sorting; so it would be more expensive.

The Django source code is here (django/db/models/sql/query.py RawQuery.has_results):

https://github.com/django/django/blob/60e52a047e55bc4cd5a93a8bd4d07baed27e9a22/django/db/models/sql/query.py#L499

def has_results(self, using):
    q = self.clone()
    if not q.distinct:
        q.clear_select_clause()
    q.clear_ordering(True)
    q.set_limits(high=1)
    compiler = q.get_compiler(using=using)
    return compiler.has_results()

Another gotcha that got me the other day is invoking a QuerySet in an if statement. That executes and returns the whole query !

If the variable query_set may be None (unset argument to your function) then use:

if query_set is None:
    # 

not:

if query_set:
   # you just hit the database

Answer 2

exists() is generally faster than count(), though not always (see test below). count() can be used to check for both existence and length.

Only use qs[0]if you actually need the object. It's significantly slower if you're just testing for existence.

On Amazon SimpleDB, 400,000 rows:

• bare qs: 325.00 usec/pass
• qs.exists(): 144.46 usec/pass
• qs.count() 144.33 usec/pass
• qs[0]: 324.98 usec/pass

On MySQL, 57 rows:

• bare qs: 1.07 usec/pass
• qs.exists(): 1.21 usec/pass
• qs.count(): 1.16 usec/pass
• qs[0]: 1.27 usec/pass

I used a random query for each pass to reduce the risk of db-level caching. Test code:

import timeit

base = """
import random
from plum.bacon.models import Session
ip_addr = str(random.randint(0,256))+'.'+str(random.randint(0,256))+'.'+str(random.randint(0,256))+'.'+str(random.randint(0,256))
try:
    session = Session.objects.filter(ip=ip_addr)%s
    if session:
        pass
except:
    pass
"""

query_variatons = [
    base % "",
    base  % ".exists()",
    base  % ".count()",
    base  % "[0]"
    ]

for s in query_variatons:
    t = timeit.Timer(stmt=s)
    print "%.2f usec/pass" % (1000000 * t.timeit(number=100)/100000)

Answer 3

It depends on use context.

According to documentation:

Use QuerySet.count()

...if you only want the count, rather than doing len(queryset).

Use QuerySet.exists()

...if you only want to find out if at least one result exists, rather than if queryset.

But:

Don't overuse count() and exists()

If you are going to need other data from the QuerySet, just evaluate it.

So, I think that QuerySet.exists() is the most recommended way if you just want to check for an empty QuerySet. On the other hand, if you want to use results later, it's better to evaluate it.

I also think that your third option is the most expensive, because you need to retrieve all records just to check if any exists.

Answer 4

@Sam Odio's solution was a decent starting point but there's a few flaws in the methodology, namely:

1. The random IP address could end up matching 0 or very few results
2. An exception would skew the results, so we should aim to avoid handling exceptions

So instead of filtering something that might match, I decided to exclude something that definitely won't match, hopefully still avoiding the DB cache, but also ensuring the same number of rows.

I only tested against a local MySQL database, with the dataset:

>>> Session.objects.all().count()
40219

Timing code:

import timeit
base = """
import random
import string
from django.contrib.sessions.models import Session
never_match = ''.join(random.choice(string.ascii_uppercase) for _ in range(10))
sessions = Session.objects.exclude(session_key=never_match){}
if sessions:
    pass
"""
s = base.format('count')

query_variations = [
    "",
    ".exists()",
    ".count()",
    "[0]",
]

for variation in query_variations:
    t = timeit.Timer(stmt=base.format(variation))
    print "{} => {:02f} usec/pass".format(variation.ljust(10), 1000000 * t.timeit(number=100)/100000)

outputs:

           => 1390.177710 usec/pass
.exists()  => 2.479579 usec/pass
.count()   => 22.426991 usec/pass
[0]        => 2.437079 usec/pass

So you can see that count() is roughly 9 times slower than exists() for this dataset.

[0] is also fast, but it needs exception handling.

Answer 5

I was also in this trouble. Yes exists() is faster for most cases but it depends a lot on the type of queryset you are trying to do. For example for a simple query like: my_objects = MyObject.objets.all() you would use my_objects.exists(). But if you were to do a query like: MyObject.objects.filter(some_attr='anything').exclude(something='what').distinct('key').values() probably you need to test which one fits better (exists(), count(), len(my_objects)). Remember the DB engine is the one who will perform the query, and to get a good result in performance, it depends a lot on the data structure and how the query is formed. One thing you can do is, audit the queries and test them on your own against the DB engine and compare your results you will be surprised by how naive sometimes django is, try QueryCountMiddleware to see all the queries executed, and you will see what i am talking about.

Answer 6

I would imagine that the first method is the most efficient way (you could easily implement it in terms of the second method, so perhaps they are almost identical). The last one requires actually getting a whole object from the database, so it is almost certainly the most expensive.

But, like all of these questions, the only way to know for your particular database, schema and dataset is to test it yourself.