At Sun, 17 Jan 2021 23:02:18 -0800, Noah Misch <noah@leadboat.com> wrote in
> On Sun, Jan 17, 2021 at 10:36:31PM -0800, Noah Misch wrote:
> > I wrote the above based on the "PageGetLSN(page) > (snapshot)->lsn" check in
> > TestForOldSnapshot(). If the LSN isn't important, what else explains
> > RelationAllowsEarlyPruning() checking RelationNeedsWAL()?
>
> Thinking about it more, some RelationAllowsEarlyPruning() callers need
> different treatment. Above, I was writing about the case of deciding whether
> to do early pruning. The other RelationAllowsEarlyPruning() call sites are
> deciding whether the relation might be lacking old data. For that case, we
> should check RELPERSISTENCE_PERMANENT, not RelationNeedsWAL(). Otherwise, we
> could fail to report an old-snapshot error in a case like this:
>
> setup: CREATE TABLE t AS SELECT ...;
> xact1: BEGIN ISOLATION LEVEL REPEATABLE READ; SELECT 1; -- take snapshot
> xact2: DELETE FROM t;
> (plenty of time passes)
> xact3: SELECT count(*) FROM t; -- early pruning
> xact1: SAVEPOINT q; SELECT count(*) FROM t; ROLLBACK TO q; -- "snapshot too old"
> xact1: ALTER TABLE t SET TABLESPACE something; -- start skipping WAL
> xact1: SELECT count(*) FROM t; -- no error, wanted "snapshot too old"
>
> Is that plausible?
Thank you for the consideration and yes. But I get "snapshot too old"
from the last query with the patched version. maybe I'm missing
something. I'm going to investigate the case.
regards.
--
Kyotaro Horiguchi
NTT Open Source Software Center
