Re: rollback segment too old

"Howard J. Rogers" wrote:
>
> "Knut Talman" <knut.talman_at_mytoys.de> wrote in message
> news:3D130468.826B5A8A_at_mytoys.de...
> > Pinto wrote:
> >
> > > if you can see, the wraps have more than doubled. and as expected i am
> still
> > > getting ora-1555
> >
> > Not only the size of a rollback segment can become a bottleneck also the
> number
> > of rbs is important. Every transaction has to use the header of the rbs to
> enter
> > its information. If there are no slots left (in the header) you'll get
> > ORA-01555, even if there are a lot of free rollback segment blocks.
> >
>
> Er, no... you'd end up with horrendous amounts of buffer busy waits, but
> that's about it. The 1555s you are thinking of is because of the delayed
> block cleanout mechanism. It's when the transaction slot on a *normal data
> block* is re-used, thus preventing retrieval of rollback information needed
> for a consistent read, and hence producing a 1555.

Of course, I mixed it up. What I wanted to say (after rethinking) was:

After a transaction has commited, the rollback block it has used can be released, also the transaction slot can be overwritten. If a read request that started a long time ago finally reaches the data block after this has happened and tries to find the relevant rollback it may fail, because the tranaction table entry can not be rebuilt -> ORA-01555. Even if it can be rebuilt but points to an already overwritten rollback block -> ORA-01555. To reduce the rate at which slots in the segment header transaction tables are recycled you have to have plenty of rollback segments. To reduce the rate at which rollback information is overwritten, the rollback segments have to be large enough.
Conclusion: The only way to reduce or avoid ORA-01555 is to have enough rollback segments, which are large enough. Often it is a trial and error process to find the best values.

Regards,

Knut Received on Fri Jun 21 2002 - 08:56:50 CDT