On 22.10.2013 19:55, Heikki Linnakangas wrote:
> I fixed the the same problem in GiST a few years back, by making it
> tolerate missing downlinks, and inserting them lazily. The B-tree code
> tolerates them already on scans, but gets confused on insertion, as seen
> above. I propose that we use the same approach I used with GiST, and add
> a flag to the page header to indicate "the downlink hasn't been inserted
> yet". When insertion (or vacuum) bumps into a flagged page, it can
> finish the incomplete action by inserting the downlink.
This is what I came up with.
One thing I'm not totally happy about is the way page deletions of
incompletely split pages are handled. Basically, it just bails out and
refuses to delete a page that is part of an incomplete split. That's
probably OK in practice, as incomplete splits should be very rare
anyway, but it's a bit dissatisfying to not handle the case because at
first glance it seems like it should be even simpler than usual to
delete a page that has no downlink. Nevertheless, I decided to just skip
that for now.
After this patch, deleting the only child of a parent and the parent
itself is still a multi-WAL-record operation that needs to be tracked
during recovery, and completed at the end of recovery. I'd like to
eliminate that too, but that's another patch.
- Heikki
