ST_CollectionExtract: inconsistent EMPTY results

[strk]

strk=# select ST_AsEWKT(ST_CollectionExtract('LINESTRING EMPTY', 2));
    st_asewkt     
------------------
 LINESTRING EMPTY
(1 row)

strk=# select ST_AsEWKT(ST_CollectionExtract('GEOMETRYCOLLECTION EMPTY', 2));
       st_asewkt       
-----------------------
 MULTILINESTRING EMPTY
(1 row)

Why is it a _single_ empty in the first case but a _multi_ empty in the second ? How to transform a multi to a simple ? (ST_Homogenize missing)

[pramsey]

Actually the behaviour is consistent: when passed a non-collection, the function just echoes it back. This is true with both empties and non-empties.

postgis20=# select ST_AsText(ST_CollectionExtract('GEOMETRYCOLLECTION(LINESTRING(0 0, 1 1))', 2));
         st_astext          
----------------------------
 MULTILINESTRING((0 0,1 1))
(1 row)

postgis20=# select ST_AsText(ST_CollectionExtract('LINESTRING(0 0, 1 1)', 2));
      st_astext      
---------------------
 LINESTRING(0 0,1 1)
(1 row)

The return for non-empty collections is always a homogeneous multi-something, even if the input collection has only one member.

[strk]

Replying to pramsey:

Actually the behaviour is consistent: when passed a non-collection, the function just echoes it back.

Not true. Was changed some time ago to always return the requested type:

=# select ST_AsText(ST_CollectionExtract('POINT(0 0)', 2));
    st_astext
------------------
 LINESTRING EMPTY
(1 row)

See #457

It would be consistent to always return the MULTI* version of the requested type...

[pramsey]

Returns all results as multi*

[pramsey]

Not sure I like this improvement entirely, but what the heck.

[pramsey]

Ah, and how I know why I don't like this improvement, it breaks regression (of course, I knew that, but it's more obvious when the regression tests fail, which they do). So this change would change the behaviour of the 2.0 branch. I'm pushing it forward to 2.1 on principle.

[pramsey]

Changed to always return MULTI* at r9764

[pramsey]

...aaand, backed out again at r9766

strk, if you want this you'll have to apply it yourself and fix/change the topology code to match.

[robe]

punt