Op 29-03-2022 om 12:50 schreef Matthias van de Meent:
> On Tue, 29 Mar 2022 at 11:10, Kyotaro Horiguchi <horikyota.ntt@gmail.com> wrote:
>>
>> At Tue, 29 Mar 2022 17:06:21 +0900 (JST), Kyotaro Horiguchi <horikyota.ntt@gmail.com> wrote in
>>> At Mon, 28 Mar 2022 18:36:46 +0900 (JST), Kyotaro Horiguchi <horikyota.ntt@gmail.com> wrote in
>>> Then, I found that portal->holdSnapshot is that. I came up with the
>>> attached. It does the follows:
>>>
>>> 1. Teach PlannedStmtRequiresSnapshot() to return true for FetchStmt.
>>>
>>> 2. Use holdSnapshot in RunFromStore if any.
>>>
>>>
>>> The rerpducer is reduced to as small as the following.
>>>
>>> CREATE TABLE t (a text);
>>> INSERT INTO t VALUES('some random text');
>>> BEGIN;
>>> DECLARE c CURSOR FOR SELECT * FROM t;
>>> FETCH ALL IN c;
>>>
>>> But I haven't come up with a reasonable way to generate the 'some
>>> random text' yet.
>>
>> I gave up and took a straightforward way to generate one.
>>
>> I don't like that it uses a fixed length for the random text, but
>> anyway it works for now...
>
> An shorter (?) reproducer might be the following, which forces any
> value for 'a' to be toasted and thus triggering the check in
> init_toast_snapshot regardless of value length:
>
> CREATE TABLE t (a text);
> ALTER TABLE t ALTER COLUMN a SET STORAGE EXTERNAL;
> INSERT INTO t VALUES ('toast');
> BEGIN;
> DECLARE c CURSOR FOR SELECT * FROM t;
> FETCH ALL IN c;
Excellent. That indeed immediately forces the error.
(and the patch prevents it)
Thanks!
>
> Enjoy,
>
> -Matthias
