Re: Improving REINDEX for system indexes (long)

Bruce Momjian <pgman(at)candle(dot)pha(dot)pa(dot)us> writes:
> Tom, would you summarize what REINDEX currently _doesn't_ do?

As of CVS tip I think the only deficiency is that indexes on the shared
catalogs (pg_database, pg_shadow, pg_group) have to be reindexed in
place, rather than being rebuilt with a new relfilenode as is done for
CLUSTER or TRUNCATE. In-place reindexing isn't crash-safe, since if
you fail you're left with a half-built (effectively corrupt) index.

I don't see any way to avoid that, though, since we cannot change the
relfilenode value for a shared index.

I was toying with the notion of changing btree index build to not write
the metapage until the index is fully built; in this way, at least the
corrupted state of the index would be obvious. (You'd get "not a btree"
failures.)

regards, tom lane