Re: Why is undo backed with redo for GTT

Hi!

> Try considering this...
>
> The undo/rollback segments reside in a tablespace,
> just like tables & indexes. When an INSERT, UPDATE
> or DELETE occurs, one or more blocks in a table
> changes, zero or more blocks in an index change, one
> or more blocks in the undo change AND every change
> MUST be recorded in the redo logfile for safe keeping.
> Otherwise they could not be recovered (rolled forward
> and if uncommitted rolled back) after a crash.

The point was, that the changes do *not* have to be recorded in redo, because we are dealing with temporary tables here (which don't have to survive the crash) but Oracle just hasn't (yet) built a mechanism, which distinguishes whether the undo should to be logged or not (the mechanism works in other areas such as nologging index rebuild or CTAS), but this is block level, not change vector level logging afaik..

So redo is generated for undo even if the relevant undo doesn't need to be reconstucted after a crash.

Actually, rollback segments should just be fine, if temp table DML undo wouldn't be logged, thus leaving few unneeded blocks in RBS to a inconsistent state - rolling back is done walking through the undo chain anyway and if we just pass rolling back transactions on transactions on (non existent) temp tables after crash, we shouldn't encounter any problems.

Sorry, if my post is too confusing, it's quite late here ;) Tanel. Received on Wed Mar 05 2003 - 17:20:15 CST