Re: [Bug] Add the missing RTE_GRAPH_TABLE case to transformLockingClause()

On Wed, Jun 10, 2026 at 1:03 AM Ayush Tiwari
<ayushtiwari(dot)slg01(at)gmail(dot)com> wrote:
>
> Hi,
>
> On Wed, 13 May 2026 at 10:57, Ashutosh Bapat <ashutosh(dot)bapat(dot)oss(at)gmail(dot)com> wrote:
>>
>> > Repro:
>> >
>> > CREATE TABLE v(id int PRIMARY KEY, vname text);
>> > CREATE PROPERTY GRAPH g VERTEX TABLES (v);
>> > SELECT * FROM GRAPH_TABLE(g MATCH (a) COLUMNS (a.vname)) gt
>> > FOR UPDATE OF gt;
>> > -- ERROR: unrecognized RTE type: 8
>> >
>> > Attached a patch that returns ERRCODE_FEATURE_NOT_SUPPORTED "FOR ... cannot be
>> > applied to a GRAPH_TABLE" with a position pointer, matching the convention used by
>> > the function/tablefunc etc.
>>
>> In a fully matured SQL/PGQ support, a row mark on GRAPH_TABLE should
>> be pushed down to the resulting subqueries in rewriteGraphTable(). I
>> see this similar to how the row mark is pushed down into views during
>> the rewrite phase. I don't think the code changes will be massive, it
>> will be a matter of invoking markQueryForLocking() on the query that
>> replaces the graph pattern in rewriteGraphTable(). I am not sure
>> whether we want to do that now when stabilizing the feature. Peter,
>> what do you think?
>
>
> I wanted to implement this by trying to push FOR UPDATE/SHARE
> down into the relational tables underlying the GRAPH_TABLE (the
> "fully matured SQL/PGQ" direction mentioned). But had
> some doubts around the implementation.
>
> I'd like input on a few design questions from you/Peter/hackers:
>
> Q1. Is the right place to do the pushdown rewriteGraphTable(),
> mirroring what ApplyRetrieveRule() does for views (i.e. call
> markQueryForLocking() on the freshly built subquery)? That
> would require exposing markQueryForLocking() outside
> rewriteHandler.c.

That's right. I think making markQueryForLocking non-static is fine
since we are using it in rewriter code itself.

>
> Q2. How should FOR UPDATE behave when the MATCH expands to more
> than one path and rewriteGraphTable() produces a UNION?
> markQueryForLocking() refuses to descend into set-operations,
> same as it does for views. Options:
> (a) reject FOR UPDATE on multi-path MATCH with a clear
> "cannot lock rows from a set-operation result" error,
> matching the view behaviour;
> (b) silently fall back to the current
> "raise relation lock only" behaviour and document it.
> I think it should be (a).
>

Building a UNION query is an implementation artifact of GRAPH_TABLE,
locking clause should not depend upon it. I think we should build the
subqueries within the setop with a locking clause of their own. That's
supported.

> Q3. SQL/PGQ does not really define UPDATE/DELETE semantics for
> graph paths. Is the intent here that FOR UPDATE/SHARE on
> GRAPH_TABLE is purely a concurrency hint rather than a
> precursor to UPDATE WHERE CURRENT OF? If so, that should
> probably be documented next to the GRAPH_TABLE syntax in
> the SQL reference.
>

I don't understand the WHERE CURRENT OF reference here. FOR
UPDATE/SHARE is generally used with an intent of locking the read rows
so that they don't get modified while the transaction is in progress.
One can use just UPDATE to update the FOR UPDATE locked rows as well.

>>
>> If we decide to not to support it, the code changes look on the right
>> path. We need to update the following comment which appears earlier in
>> the function to mention GRAPH_TABLE, /* ignore JOIN, SPECIAL,
>> FUNCTION, VALUES, CTE RTEs */. I think it will be better to specify
>> that this restriction is temporary in the comments at both the places;
>> but that's arguable.
>>
>> > Patch includes tests for all four locking strengths.
>> > Since the code path looks simple we can just keep one of them as well and trim other
>> > tests. Thoughts?
>>
>> I don't think we need all the four tests, but we need a test for
>> locking clause without relation since that takes a different code
>> path.
>
>
> Attached is v2-0001, rebased on master and addressing
> the comments you added in case we plan on parking the feature
> work for once v20 opens.

Thanks. I revised it a bit. Reworded the commit message to describe
user facing behaviour instead of the internals, which are quite
obvious from the code changes. Other cases in transformLockingClause()
don't mention XXX, they just have ereport(). Removed that comment.
Also massaged comments in the test - we usually don't describe the
expected behaviour or error when it is obvious from the output.
Attached revised patch.

--
Best Wishes,
Ashutosh Bapat