Christopher Kings-Lynne <chriskl@familyhealth.com.au> writes:
>> The way you had it coded, it was generating
>> new index entries pointing at the old version of the tuple ...
> Strange. I guess I must have been testing with a database that had
> short enough system catalogs that the indexes were never used?
Could be, especially if you were using client-side queries to check the
result. I think most of the system's internal catalog fetches are
hard-wired to use indexes except under the most dire circumstances
(mainly, a standalone backend with -P). But a client-issued query
would do whatever the planner thought best.
regards, tom lane