Question

I’m re-asking this question in a simplified and expanded manner.

Consider these sql statements:

create table foo (id INT, score INT);

insert into foo values (106, 4);
insert into foo values (107, 3);
insert into foo values (106, 5);
insert into foo values (107, 5);

select T1.id, avg(T1.score) avg1
from foo T1
group by T1.id
having not exists (
    select T2.id, avg(T2.score) avg2
    from foo T2
    group by T2.id
    having avg2 > avg1);

Using sqlite, the select statement returns:

id          avg1      
----------  ----------
106         4.5       
107         4.0       

and mysql returns:

+------+--------+
| id   | avg1   |
+------+--------+
|  106 | 4.5000 |
+------+--------+

As far as I can tell, mysql’s results are correct, and sqlite’s are incorrect. I tried to cast to real with sqlite as in the following but it returns two records still:

select T1.id, cast(avg(cast(T1.score as real)) as real) avg1
from foo T1
group by T1.id
having not exists (
    select T2.id, cast(avg(cast(T2.score as real)) as real) avg2
    from foo T2
    group by T2.id
    having avg2 > avg1);

Why does sqlite return two records?

Quick update:

I ran the statement against the latest sqlite version (3.7.11) and still get two records.

Another update:

I sent an email to [e-mail] about the issue.

Myself, I’ve been playing with VDBE and found something interesting. I split the execution trace of each loop of not exists (one for each avg group).

To have three avg groups, I used the following statements:

create table foo (id VARCHAR(1), score INT);

insert into foo values ('c', 1.5);
insert into foo values ('b', 5.0);
insert into foo values ('a', 4.0);
insert into foo values ('a', 5.0);

PRAGMA vdbe_listing = 1;
PRAGMA vdbe_trace=ON;

select avg(score) avg1
from foo
group by id
having not exists (
    select avg(T2.score) avg2
    from foo T2
    group by T2.id
    having avg2 > avg1);

We clearly see that somehow what should be r:4.5 has become i:5:

I’m now trying to see why that is.

Final edit:

So I’ve been playing enough with the sqlite source code. I understand the beast much better now, although I’ll let the original developer sort it out as he seems to already be doing it:

http://www.sqlite.org/src/info/430bb59d79

Interestingly, to me at least, it seems that the newer versions (some times after the version I’m using) supports inserting multiple records as used in a test case added in the aforementioned commit:

CREATE TABLE t34(x,y);
INSERT INTO t34 VALUES(106,4), (107,3), (106,5), (107,5);  

Answer 1

I tried to mess with some variants of query.

It seems, like sqlite has errors in using of previous declared fields in a nested HAVING expressions.

In your example avg1 under second having is always equal to 5.0

Look:

select T1.id, avg(T1.score) avg1
from foo T1
group by T1.id
having not exists (
    SELECT 1 AS col1 GROUP BY col1 HAVING avg1 = 5.0);

This one returns nothing, but execution of the following query returns both records:

...
having not exists (
    SELECT 1 AS col1 GROUP BY col1 HAVING avg1 <> 5.0);

I can not find any similar bug at sqlite tickets list.