So, in reading these patches, it came to me that we might want to have
pg_upgrade mark indexes invalid if we in the future change the
implementation of some opclass. For instance, the inclusion opclass
submitted here uses three columns: the indexed value itself, plus two
booleans; each of these booleans is a workaround for some nasty design
decision in the underlying datatypes.
One boolean is "unmergeable": if a block range contains both IPv4 and
IPv6 addresses, we mark it as 'unmergeable' and then every query needs
to visit that block range always. The other boolean is "contains empty"
and is used for range types: it is set if the empty value is present
somewhere in the block range.
If in the future, for instance, we come up with a way to store the ipv4
plus ipv6 info, we will want to change the page format. If we add a
page version to the metapage, we can detect the change at pg_upgrade
time and force a reindex of the index.
Thoughts?
--
Álvaro Herrera http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services