On Wed, Apr 19, 2017 at 1:03 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
david.g.johnston@gmail.com writes: > This doesn't, and should since the number of elements in the non-empty array > shouldn't change the dimensionality logic.
> SELECT array_agg(CASE WHEN a = ARRAY[]::text[] THEN ARRAY['N/A']::text[] > ELSE a END) > FROM ( VALUES (1, ARRAY[]::text[]), (1, ARRAY['1','2']::text[]) ) vals (v, > a)
Why do you think that should work? You're asking array_agg to accumulate a 1-D length-1 array and then a 1-D length-2 array. There's no way to make a rectangular 2-D array out of that, except perhaps by inventing entries which isn't in array_agg's charter.
All this and all I really want is a friggin' "array_append / array_concat" aggregate function that accepts either scalars or matching "primary dimension" arrays - and treats empty arrays as no-ops.
In the end I realized that serializing the arrays to text would work just fine. The extra I/O for converting from and to an actual array type is immaterial in my situation. It still doesn't remove my actual desire for an aggregate array_concat type function, and the for a rectangular array is a bit annoying (i.e., allow non-rectangular and just report an out-of-bounds error on attempts to explicitly access non-existent elements) but that's easy enough to toss in with the other peculiarities in this area.
SELECT unnest(array_agg)::text[]
FROM (
SELECT array_agg(a)
FROM ( VALUES (1, ARRAY[]::text[]::text), (1, ARRAY['1','2']::text[]::text) ) vals (v, a)