While hacking on the GiST insertion algorithm, I noticed a bug with
temporary GiST indexes. The scan algorithm uses LSNs to detect a
concurrent page split, but for a temporary index, we just use a constant
LSN on the assumption that there can't be anyone else modifying the
index concurrently. But that assumption is not true. While a temporary
index can't be modified by other backends, it's entirely possible to
insert new tuples to the index within the same backend while a scan is
in progress.
Here's a test case, using btree_gist. If you modify it to use a
non-temporary table, it works fine, but with a temporary table the 2nd
SELECT misses some rows, when new rows are inserted in the middle of the
scan.
Operations on temporary tables are not WAL-logged, so we don't have LSNs
to use, but we just need a monotonically increasing sequence of numbers
for this purpose. Attached patch provides a function to generate such
fake LSNs.
--
Heikki Linnakangas
EnterpriseDB http://www.enterprisedb.com