Re: index node split autonomy

On Tue, 27 May 2003 05:39:58 +1000, Howard J. Rogers wrote:

> Bloody good question!
>
> The split is NOT rolled back, I guess because it would be an awful lot of
> reverse-work, and if you just do one more insert, we'll probably have to
> repeat it all over again.

Actually, according to my information, the whole thing is happening simultaneously. All the involved index blocks are pinned in SGA and, if there is a need for a split, the whole subtree is locked so that it cannot be concurrently modified. The index blocks are modified in SGA and their pprevious values are stored in, surprise, surprise, rollback segment blocks. The only actual difference is that in case of a split, the whole subtree is locked. That can really kill concurrency if you have a lot of those, but the largest index that I've ever seen has had degree 11 and it was humongous. That means that in a 17 GB index, a split has occurred only 10 times. Pretty good, isn't it? Of course, this table is now partitioned, indexes are local to partitions and they are not as big any more. That was explained long time ago, when Oracle has released version 6 with TPO. I heard the information from a guy who was with Ken Jacobs, after Ken's presentation on EOUG in Cannes, France. I believe that the year was 1992 or alike. Boy, am I old!

-- 
Mladen Gogala
Software is like sex, it is better when it is free.
Linus Torvalds