On Fri, 2007-06-01 at 19:11 +0400, Teodor Sigaev wrote:
> > <2007-06-01 16:28:51.708 CEST:%> LOG: GIN incomplete split root:11118
> > l:45303 r:111740 at redo CA/C8243C28
> ...
> > <2007-06-01 16:38:23.133 CEST:%> LOG: GIN incomplete split root:11118
> > l:45303 r:111740 at redo CA/C8243C28
>
> Looks like a bug in GIN. I'll play with it. Can you provide more details about
> your test?
Agreed, sorry about that.
I note that forgetIncompleteSplit doesn't report anything if it fails to
find an incomplete split when it looks for it.
ISTM that either we have
1) sometimes a starting split, but no ending split
2) a starting split, but we try to remove the wrong split later, for
some reason. i.e. we make the change correctly, just misjudge the xlog
somehow.
(1) would show up on data retrieval tests, but (2) would show up only on
recovery. So I would think it is (2).
I'd suggest we throw an error, as shown in the enclosed patch. Frank,
can you give that a whirl to provide Teodor with something more to work
with? Thanks.
Neither GIST nor B-tree seems to throw an error in corresponding
locations also, so the potential for not being able to track this is
high. I'd want to throw errors in those locations also.
--
Simon Riggs
EnterpriseDB http://www.enterprisedb.com