count(col) returns incorrect values if col is NULL

[avikivity]

count(col) (unlike count(*) or count(1)) is supposed not to count rows where col is NULL. However, it does, if the column has a collection type:

cqlsh> create table ks1.tab (id int PRIMARY KEY , i int, l frozen<list<int>>);                                                                         
cqlsh> insert into ks1.tab (id, i) values (1, 1);                                                                                                  
cqlsh> insert into ks1.tab (id, l) values (2, [2]);
cqlsh> select * from ks1.tab;

 id | i    | l
----+------+------
  1 |    1 | null
  2 | null |  [2]

(2 rows)

cqlsh> select count(i) from ks1.tab;

 system.count(i)
-----------------
               1

(1 rows)

cqlsh> select count(l) from ks1.tab;

 count
-------
     2

(1 rows)

Checked with 5.2.1. 5.0.5 known to be good.

[avikivity]

Likely due to

    } else if (name.has_keyspace()
                ? name == COUNT_NAME
                : name.name == COUNT_NAME.name) {
        auto arg_types = get_arguments(COUNT_NAME.name);
        if (arg_types.size() != 1) {
            throw std::runtime_error("count() function requires only 1 argument");
        }

        auto& arg = arg_types[0];
        return aggregate_fcts::make_
            return aggregate_fcts::make_count_rows_function();
        }

which converts COUNT_NAME to COUNT_ROWS_NAME.

[avikivity]

In 5.0 it works by accident. count(frozen<list<int>>) happens to match the signature of count(blob) because of the weakly_assignable thing.

[avikivity]

5.1 is also broken.

This is difficult to backport due to all the refactoring, but perhaps not difficult to rework.