> isn't really specific to ANALYZE. I'm inclined to think that the most
> reasonable fix is to make get_sort_group_operators() and related
Hm, patch is in attach but it doesn't solve all problems. Initial bug is still
here for array of row type, but when I tried to change that with recursive call
get_sort_group_operators() as it done for row type then 'gmake check' fails
because lookup_rowtype_tupdesc fails to find anonymous composite type. As I can
see anonymous composite type are identified by (RECORD_OID, typmod) pair and
typmod aren't available here. So, my plan was to add typmod to
get_sort_group_operators() but I have no idea where is typmod value for element
type.
In runtime problems are solved by using HeapTupleHeaderGetTypMod() for record /
element of array.
With modified get_sort_group_operators() for arrays check actually fails for
query 'select * from search_graph order by path;' at file
src/test/regress/sql/with.sql. get_sort_group_operators() is called from
addTargetToSortList() and fails.
It seems to me that anonymous composite type could force us to teach
vacuum/analyze code to fallback to simpler analyze algorithm.
--
Teodor Sigaev E-mail: teodor@sigaev.ru
WWW: http://www.sigaev.ru/