Tom,
> I'm toying with the notion of ripping out that logic and instead
> building an in-memory hashtable of already-returned TIDs. This could
> theoretically run out of memory if the multiple indexscan returns enough
> tuples, but I think in practice that wouldn't happen because the planner
> wouldn't choose an indexscan when very large numbers of tuples are being
> selected.
Don't forget all of the tyros who tune their queries through "set
enable_seqscan=false". I think we'd need some kind of memory safety valve on
this, like sandboxing it in sort_mem.
--
-Josh BerkusAglio Database SolutionsSan Francisco