Re: BUG #18947: TRAP: failed Assert("len_to_wrt >= 0") in pg_stat_statements

On 2025/06/08 11:39, Michael Paquier wrote:
> On Sat, Jun 07, 2025 at 10:19:53PM +0530, Dilip Kumar wrote:
>> I didn't get time to debug, I might look into it tomorrow if someone
>> doesn't do it before that, but I am able to reproduce the issue and
>> seems like problem is with respect to computing the quer_loc in below
>> loop, and due to that 'quer_loc' is becoming bigger than query_len and
>> hitting the assert.
>
> The loop in charge of assigning the constants in the queries needs to
> be smarter in terms of the new locations assigned to sub-queries, here
> the VALUES one.
>
> 499edb09741b is the origin of the regression, so open item assigned to
> me (will look into it later), adding Anthonin in CC.

The issue seems to be that the commit missed handling cases where
a clause follows a SELECT query wrapped in parentheses.

If the following clause includes a constant (e.g., LIMIT 1 or
FETCH FIRST 1 ROW ONLY), it can trigger the reported assertion failure.
If it doesn't include a constant (e.g., FETCH FIRST ROW ONLY),
the assertion failure doesn't occur, but only the inner SELECT query (inside
the parentheses) appears in pg_stat_statements, i.e.,
the "FETCH FIRST ROW ONLY" part is missing from pg_stat_statements.query.

+ SelectStmt *n = (SelectStmt *) $2;
+
+ /*
+ * As SelectStmt's location starts at the SELECT keyword,
+ * we need to track the length of the SelectStmt within
+ * parentheses to be able to extract the relevant part
+ * of the query. Without this, the RawStmt's length would
+ * be used and would include the closing parenthesis.
+ */
+ n->stmt_len = @3 - @2;

Shouldn't this part be skipped in cases where the SELECT with parens is
followed by a clause? At least in those cases, this logic doesn't seem
appropriate.

Regards,

--
Fujii Masao
NTT DATA Japan Corporation