Please find the attached HOT patch, which I think is now ready for
review.
The only known issue left to-be-done is making the index available
in the read-committed transaction which created it. Right now the index
is made available to the transaction unless one or more rows were
UPDATEd in the same transactions before creating the index OR there were
RECENTLY_DEAD tuples that we did not index while building the index.
Both of these cases do not seem very common to me.The patch can easily
be tweaked to make the index available even in these cases, but I left it
because Tom has raised concerns about transaction using the new index
with some old snapshot. I haven't been able to create such a scenario
so thought would best leave it until someone picks it up for review and
take it up at that time.
One of the regression test fails, but thats just a side-effect of how CIC
works now. I didn't change the expected output just yet because that might
have covered this side-effect. I plan to spend some time on performance
testing and would post those results soon.
I would appreciate if others also give it a shot and see if it gives any
performance jump in their respective setups. I think the best would
come out in IO bound testing.
Thanks,
Pavan
--
EnterpriseDB
http://www.enterprisedb.com
