a thinko in b676ac443b6

On 7/27/21 4:28 AM, Amit Langote wrote:

Hi,

I noticed $subject while rebasing my patch at [1] to enable batching
for the inserts used in cross-partition UPDATEs.

b676ac443b6 did this:

-           resultRelInfo->ri_PlanSlots[resultRelInfo->ri_NumSlots] =
-               MakeSingleTupleTableSlot(planSlot->tts_tupleDescriptor,
-                                        planSlot->tts_ops);
...
+           {
+               TupleDesc tdesc =
CreateTupleDescCopy(slot->tts_tupleDescriptor);
+
+               resultRelInfo->ri_Slots[resultRelInfo->ri_NumSlots] =
+                   MakeSingleTupleTableSlot(tdesc, slot->tts_ops);
...
+               resultRelInfo->ri_PlanSlots[resultRelInfo->ri_NumSlots] =
+                   MakeSingleTupleTableSlot(tdesc, planSlot->tts_ops);

I think it can be incorrect to use the same TupleDesc for both the
slots in ri_Slots (for ready-to-be-inserted tuples) and ri_PlanSlots
(for subplan output tuples). Especially if you consider what we did
in 86dc90056df that was committed into v14. In that commit, we
changed the way a subplan under ModifyTable produces its output for an
UPDATE statement. Previously, it would produce a tuple matching the
target table's TupleDesc exactly (plus any junk columns), but now it
produces only a partial tuple containing the values for the changed
columns.

So it's better to revert to using planSlot->tts_tupleDescriptor for
the slots in ri_PlanSlots. Attached a patch to do so.

Yeah, this seems like a clear mistake - thanks for noticing it! Clearly
no regression test triggered the issue, so I wonder what's the best way
to test it - any idea what would the test need to do?

I did some quick experiments with batched INSERTs with RETURNING clauses
and/or subplans, but I haven't succeeded in triggering the issue :-(

regards

--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

On Wed, Jul 28, 2021 at 1:07 AM Tomas Vondra
<tomas.vondra@enterprisedb.com> wrote:

On 7/27/21 4:28 AM, Amit Langote wrote:

I think it can be incorrect to use the same TupleDesc for both the
slots in ri_Slots (for ready-to-be-inserted tuples) and ri_PlanSlots
(for subplan output tuples). Especially if you consider what we did
in 86dc90056df that was committed into v14. In that commit, we
changed the way a subplan under ModifyTable produces its output for an
UPDATE statement. Previously, it would produce a tuple matching the
target table's TupleDesc exactly (plus any junk columns), but now it
produces only a partial tuple containing the values for the changed
columns.

So it's better to revert to using planSlot->tts_tupleDescriptor for
the slots in ri_PlanSlots. Attached a patch to do so.

Yeah, this seems like a clear mistake - thanks for noticing it! Clearly
no regression test triggered the issue, so I wonder what's the best way
to test it - any idea what would the test need to do?

Ah, I should've mentioned that this is only a problem if the original
query is an UPDATE. With v14, only INSERTs can use batching and the
subplan does output a tuple matching the target table's TupleDesc in
their case, so the code seems to work fine.

As I said, I noticed a problem when rebasing my patch to allow
cross-partition UPDATEs to use batching for the inserts that are
performed internally to implement such UPDATEs. The exact problem I
noticed is that the following Assert tts_virtual_copyslot() (via
ExecCopySlot called with an ri_PlanSlots[] entry) failed:

Assert(srcdesc->natts <= dstslot->tts_tupleDescriptor->natts);

srcdesc in this case is a slot in ri_PlanSlots[] initialized with the
target table's TupleDesc (the "thinko") and dstslot is the slot that
holds subplan's output tuple ('planSlot' passed to ExecInsert). As I
described in my previous email, dstslot's TupleDesc can be narrower
than the target table's TupleDesc in the case of an UPDATE, so the
Assert can fail in theory.

I did some quick experiments with batched INSERTs with RETURNING clauses
and/or subplans, but I haven't succeeded in triggering the issue :-(

Yeah, no way to trigger this except UPDATEs. It still seems like a
good idea to fix this in v14.

--
Amit Langote
EDB: http://www.enterprisedb.com

On 7/28/21 3:15 AM, Amit Langote wrote:

On Wed, Jul 28, 2021 at 1:07 AM Tomas Vondra
<tomas.vondra@enterprisedb.com> wrote:

On 7/27/21 4:28 AM, Amit Langote wrote:

I think it can be incorrect to use the same TupleDesc for both the
slots in ri_Slots (for ready-to-be-inserted tuples) and ri_PlanSlots
(for subplan output tuples). Especially if you consider what we did
in 86dc90056df that was committed into v14. In that commit, we
changed the way a subplan under ModifyTable produces its output for an
UPDATE statement. Previously, it would produce a tuple matching the
target table's TupleDesc exactly (plus any junk columns), but now it
produces only a partial tuple containing the values for the changed
columns.

So it's better to revert to using planSlot->tts_tupleDescriptor for
the slots in ri_PlanSlots. Attached a patch to do so.

Yeah, this seems like a clear mistake - thanks for noticing it! Clearly
no regression test triggered the issue, so I wonder what's the best way
to test it - any idea what would the test need to do?

Ah, I should've mentioned that this is only a problem if the original
query is an UPDATE. With v14, only INSERTs can use batching and the
subplan does output a tuple matching the target table's TupleDesc in
their case, so the code seems to work fine.

As I said, I noticed a problem when rebasing my patch to allow
cross-partition UPDATEs to use batching for the inserts that are
performed internally to implement such UPDATEs. The exact problem I
noticed is that the following Assert tts_virtual_copyslot() (via
ExecCopySlot called with an ri_PlanSlots[] entry) failed:

Assert(srcdesc->natts <= dstslot->tts_tupleDescriptor->natts);

srcdesc in this case is a slot in ri_PlanSlots[] initialized with the
target table's TupleDesc (the "thinko") and dstslot is the slot that
holds subplan's output tuple ('planSlot' passed to ExecInsert). As I
described in my previous email, dstslot's TupleDesc can be narrower
than the target table's TupleDesc in the case of an UPDATE, so the
Assert can fail in theory.

I did some quick experiments with batched INSERTs with RETURNING clauses
and/or subplans, but I haven't succeeded in triggering the issue :-(

Yeah, no way to trigger this except UPDATEs. It still seems like a
good idea to fix this in v14.

OK, thanks for the explanation. So it's benign in v14, but I agree it's
better to fix it there too. I'll get this sorted/pushed.

regards

--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company