queryset.count() does not work correctly with sliced querysets

[eth_anton@…]

Makes for counter-intuitive behavior when slicing...

From an interactive shell session:

>>> all = NewsItem.objects.all()
>>> all.count()
11L
>>> all[:5].count()
11L
>>> set = NewsItem.objects.all()[:5]
>>> set.count()
11L
>>>

[James Bennett]

Hm.

We're definitely explicitly dropping the limit and offset if they're present, but there's no comment to explain why. When I get a little time, I'll try the test suite with the offending lines commented out (lines 183 and 184 in django/db/models/query.py).

[Malcolm Tredinnick]

We should either be raising an error in these cases or behaving as we are now.

Count() returns a single value and it doesn't make sense to apply limit and offset to a singleton list like that (although it works. A non-zero offset is going to lead to disappointment in the number of results you get back, though). It's almost certainly not what the caller intended and it's not going to work as a way to count how many things were returned after the limit and offset are applied.

LIMIT/OFFSET are applied after the result set is determined (and in this case it consists of one item, so it's pointless).

[cheap_hydrocodone]

<a href=​http://user.aol.com/yourbestmeds/cheap_hydrocodone.html > cheap_hydrocodone

</a>. <a href=​http://user.aol.com/yourbestmeds/buy_hydrocodone_online.html > buy_hydrocodone_online
</a>. <a href=​http://user.aol.com/yourbestmeds/ambien_online.html > ambien_online
</a>. <a href=​http://user.aol.com/yourbestmeds/ambien.html > ambien
</a>. <a href=​http://user.aol.com/yourbestmeds/carisoprodol.html > carisoprodol
</a>.

[Collin Grady <cgrady@…>]

fixing spambot edits

[Collin Grady <cgrady@…>]

woops, fixing invalid fix

[Chris Beaven]

Malcom: .count() is used to determine proper page length for the paginator so it may be what the caller intended.

My opinion is that .count() should only return up to the maximum length that the sliced queryset could return.

[Chris Beaven]

oops, refreshed after Collin's changes and reverted his fixes :P

[Chris Beaven]

Even if we don't change the existing behaviour, existing code should be commented as to why we are doing it the way we are.

[Collin Grady <cgrady@…>]

On a whim I checked this using raw SQL and a LIMIT didn't affect the output of COUNT, so maybe that's why the code currently drops it - not much point adding it if the db won't respect it...

Though if there was a way to apply the limit first, and tell sql to count those without killing performance, it would definitely be more intuitive.

[Chris Beaven]

Thanks for checking that Collin.

If we do still want count to take into account the LIMIT/OFFSET, then it should be easy enough to do some basic maths to return the correct value, don't you think?

[Chris Beaven]

Documentation if it's needed

[Chris Beaven]

Malcom's comments aren't making much sense to me - no-one was talking about a singleton list.

It seems pretty clear to me that the limit/offset should be respected for a count, and my patch is sound so I'm marking as ready for checkin.

[Chris Beaven]

with tests

[Chris Beaven]

I left the doc patch separate, because I'm not sure we need to explicitly document this change in behaviour in the DB api - the old behaviour was never implied.

[Malcolm Tredinnick]

My earlier comment was that count() and limit/offset do not mix at the SQL level. You have worked around that by manually applying those parameters inside the Python code. I think this is counter-intuitive, since it's now no longer SQL-intuitive or Python-intuitive (slicing the single count() result changes the result?! We implemented limit/offset as slices because it was analogous -- you were slicing a sequence of results). But I can see the flip side, too, so I'm not going to veto this.

One problem with the patch is

>>> articles[8:100].count()
-1

(using the same test data setup). You need to make line 224 be something like count = max(0, count - offset). Fix that and add a test and we're good to go, I think.

[Chris Beaven]

Maybe I'm not following, but you can't slice a count() result as it does not return a QuerySet.
It seems pretty Python-intuitive that len(articles[:5]) == articles[:5].count().

Thanks for picking up that bug, I'll fix now.

[Malcolm Tredinnick]

(In [4488]) #fixed #2256 -- Made count() interact with slicing on QuerySets. Patch from
SmileyChris.

[redduck666]

In [1]: from django.contrib.auth.models import User

In [2]: User.objects.all()[:0].count()
Out[2]: 4296

In [3]: len(User.objects.all()[:0])
Out[3]: 0

is this considered normal?

[Karen Tracey]

Alex has opened #10202 for this.