RE: btree split logic is fragile in the presence of large index items

> -----Original Message-----
> From: pgsql-hackers-owner(at)hub(dot)org [mailto:pgsql-hackers-owner(at)hub(dot)org]On
> Behalf Of Tom Lane
>
> I've been chewing some more on the duplicate-key btree issue, and
> digging through CVS logs and Postgres v4.2 to try to understand the
> history of the code (and boy, this code does have a lot of history
> doesn't it?)
>
> One thing I'm still not too clear on is how we handle backwards
> indexscans. After looking at Lehman and Yao's paper it seems like
> only forward scans are guaranteed to work when other processes are
> busy splitting pages. Anybody know how that's handled?
>

There's the following comment for backwards scan in _bt_step()
in nbtsearch.c

/*
* If the adjacent page just split,
then
we may have
* the wrong block. Handle this
case.
Because pages
* only split right, we don't have
to wo
rry about this
* failing to terminate.
*/

Seems backwards index scans sometimes move right(scan forward).

Regards.

Hiroshi Inoue