Bernd Helmle <mailings@oopsware.de> writes:
> So i've saved the index file (normal BTree index with a single bigint
> column), did a REINDEX and the problem was gone. Looking at the index file
> with pg_filedump and pgbtreecheck from Alvaro gave me the following output:
> ...
Don't know how careful pgbtreecheck is. The pg_filedump output isn't
very helpful because you filtered away the flags, so we can't tell if
any of these pages are deleted. (If they are, the duplicate-looking
links might not be errors, since we intentionally don't reset a deleted
page's left/right links when deleting it.)
Could we see the whole special-space dump for each of the pages you're
worried about?
One thought that occurs to me is that POWER is a weak-memory-ordering
architecture, so that it's a tenable idea that this has something to do
with changing page links while not holding sufficient lock on the page.
I don't see btree doing that anywhere, but ...
BTW, how long has this installation been around, and when did you start
seeing funny behavior? Can you say with reasonable confidence that the
bug was *not* present in any older PG versions?
regards, tom lane