On Tue, Apr 30, 2019 at 10:58 AM Peter Geoghegan <pg@bowt.ie> wrote:j
> I have found what the problem is. I simply neglected to make a
> conservative assumption about suffix truncation needing to add a heap
> TID to a leaf page's new high key in nbtsort.c (following commit
> dd299df8189), even though I didn't make the same mistake in
> nbtsplitloc.c. Not sure how I managed to make such a basic error.
Attached is a much more polished version of the same patch. I tried to
make clear how the "page full" test (the test that has been fixed to
take heap TID space for high key into account) is related to other
close-by code, such as the tuple space limit budget within
_bt_check_third_page(), and the code that sets up an actual call to
_bt_truncate().
I'll wait a few days before pushing this. This version doesn't feel
too far off being committable. I tested it with some of the CREATE
INDEX tests that I developed during development of the nbtree unique
keys project, including a test with tuples that are precisely at the
1/3 of a page threshold. The new definition of 1/3 of a page takes
high key heap TID overhead into account -- see _bt_check_third_page().
--
Peter Geoghegan