SQL (Hive) group-by using nulls as wildcard

Question

I have a table like this:

group   val1   val2   val3

group1   5      .      .
group1   .      2      1
group1   .      .      3
group2   1      4      .
group2   .      .      8
group2   2      .      7

I need to count the occurrences of all possible combinations for each group in Hive, using null values (.) as a wildcard. This would give me results like this:

group   val1   val2   val3  cnt

group1   5      2      1     2
group1   5      2      3     2
group2   1      4      8     2
group2   2      4      8     1
group2   2      4      7     1

I know I can do this by selecting all distinct group-val1 pairs, full joining this with all distinct group-val2 pairs, and full joining this with all distinct group-val3 pairs. This gives me all possible combinations for each group, which I can then inner join with my table, counting cases where a row of my original data is a subset of a combination.

Something like this:

create table my_results as 

with combos as (
select *
from (select distinct group, val1 from data) A
full join (select distinct group, val2 from data) B
    on A.group = B.group
full join (select distinct group, val3 from data) C
    on A.group = C.group 
)

select A.group, A.val1, A.val2, A.val3, count(*)
from combos A
inner join data B
    on A.group = B.group
    and (A.val1 = B.val1 OR B.val1 is null)
    and (A.val2 = B.val2 OR B.val2 is null)
    and (A.val3 = B.val3 OR B.val3 is null)
group by A.group, A.val1, A.val2, A.val3

But, My dataset is very large (100s of millions of rows). and the number of all possible combinations I can expect is also very large (10s of thousands). Such a join is just too big.

Is there another way? I wondered if I could use regular expressions, but I don't know where to start.

Answer 1

In your sample data, only the third column has multiple values. So, you could just fill in the one value for the two other columns:

select group,
       max(max(col1)) over (partition by group) as col1,
       max(max(col2)) over (partition by group) as col2,
       col3,
       count(*)
from data
group by group;