My thought experiments about pseudo-pages and avoiding the split stuff were not intended to get the patch kicked out. I thought for a while that hash indexes were a special case and could benefit from dispensing with those trickier problems. Upon further reflection, for interesting size hash indexes pure hash value predicate tags wouldn't be much better. Furthermore, if we do decide we want to use using x % max_predicate_locks_per_relation to avoid having to escalate to relation predicate locks at the cost of slightly higher collision rate then we should consider that for the whole system (including heap page predicate locking), not just hash indexes. Please consider those ideas parked for now.
OK. While our potential pseudo-pages are identified as
"hash_value % some_constant_modulus", real bucket pages are very roughly
identified as "hash_value % number_of_index_pages". So, page number is
adoptive to index size, despite it costs us handling page split. In the same way,
locking in other index access methods is adoptive to an index size, so
that should be considered as useful feature which should be present in hash index
