I'm really hoping someone can provide a more straightforward method of doing this, but here's my shot:
with my_array as
(select ARRAY[[['abc', 'def'], ['ghi', 'jkl']], [['mno', 'pqr'], ['stu', 'vwx']]] as arr),
deconstructed_array as (
select a.elem, a.nr, x, y, z
FROM my_array, unnest(arr) WITH ordinality as a(elem, nr)
JOIN (
SELECT x, y, z, row_number() over (ORDER BY x, y, z) as row
FROM my_array,
generate_series(array_lower(arr, 1), array_upper(arr, 1)) g1(x),
generate_series(array_lower(arr, 2), array_upper(arr, 2)) g2(y),
generate_series(array_lower(arr, 3), array_upper(arr, 3)) g3(z)
) array_dims ON nr = row
)
select array_agg(elems) FROM (
select array_agg(elem) as elems, x, y
FROM deconstructed_array GROUP BY x, y
) first_level
WHERE x = 1
group by x
;
array_agg
-----------------------
{{abc,def},{ghi,jkl}}
(1 row)
Explanation:
We use generate_series to associate the various dimensions of the array with unnested rows. Unfortunately, we need to know that this is a three dimensional array in this case, but the length of each array shouldn't matter. The output of that second CTE looks something like this:
elem | nr | x | y | z
------+----+---+---+---
abc | 1 | 1 | 1 | 1
def | 2 | 1 | 1 | 2
ghi | 3 | 1 | 2 | 1
jkl | 4 | 1 | 2 | 2
mno | 5 | 2 | 1 | 1
pqr | 6 | 2 | 1 | 2
stu | 7 | 2 | 2 | 1
vwx | 8 | 2 | 2 | 2
From there, we just use array_agg to group the arrays back together.