On 29.04.26 16:09, Ashutosh Bapat wrote:
On Mon, Apr 27, 2026 at 11:34 AM SATYANARAYANA NARLAPURAM
<[email protected]> wrote:
Hi,
On Sun, Apr 26, 2026 at 8:10 AM Junwang Zhao <[email protected]> wrote:
Hi SATYANARAYANA,
On Sun, Apr 26, 2026 at 3:53 AM SATYANARAYANA NARLAPURAM
<[email protected]> wrote:
Hi hackers,
transformRangeGraphTable() calls transformExpr() and
assign_list_collations() for COLUMNS expressions but missed calling
resolveTargetListUnknowns(). As a result, literals such as 'val1'
in a COLUMNS clause retained type "unknown", causing failures with
ORDER BY, UNION, and output conversions.
Fix by calling resolveTargetListUnknowns() on the columns target
list right after assign_list_collations(), similar to SELECT target lists in
transformSelectStmt().
Attached a patch to fix this, which also includes test cases to reproduce.
I can reproduce this and the patch fixes it.
One question: why is resolveTargetListUnknowns called after
assign_list_collations?
I'm asking because in transformSelectStmt, resolveTargetListUnknowns
is invoked before assign_query_collations. It might not matter, but keeping
the order consistent would be good for readers.
Updated the patch. It should be before.
Do we really need a test for ORDER BY on a literal column? I replaced
all the test queries with a single one which covers all the scenarios
covered by those queries.
The patch needed to consider pstate->p_resolve_unknowns. Unknown
literals are not resolved as text always. See the test query.
I couldn't find a commit to apply this patch cleanly (the subject says
patch 5/5, so maybe you had some unpublished local changes?). After
applying the test case manually, it looks like the test output is
already correct without the code change. So if this patch is still
required, we need a better test case.
