Re: New challenge - clustering - do I understand this correctly?

On 11 Aug 2003 20:36:24 -0700, Nuno Souto <wizofoz2k_at_yahoo.com.au> wrote:

> Ed Stevens <nospam_at_noway.nohow> wrote in message
> news:<jo4gjvsv17mr0j8jqo2p7a04pbb5c6s7bg_at_4ax.com>...
>
>> With OPS, you have two concurrently instances against the same
>> database. They can both be accessed concurrently for load balancing
>> or whatever. If one fails, SQLNet (if correctly configured) will
>> reconnect everything to the surviving instance.
>
> which will then rollback any interrupted transactions in the failed
> instance, as well as releasing locks, etcetc.
>
>
>>
>> Do I have this right?
>
> In a nut shell, yes.
>
>> If so, are there pros and cons that might not be immediately obvious.
>
> HA won't scale without replacing hardware by faster/more capacity one.
>
> OPS/RAC won't scale linearly for heavy concurrent update activity,
> but this can be compensated for by design. Otherwise, just add
> more of the same boxes at any time.
>
>
>> What I see is that the HA option can result in a few minutes of outage
>> during the switch, but is much easier for the DBA to adminster -- it
>> would be handled just like any other single machine DB and all cluster
>> activity and recovery from node failure is handled by the OS --
>> transparent to Oracle. On the other hand the OPS option is more
>> complicated for the DBA to administer, but can result in near zero
>> down time.
>
> I don't understand how you can think RAC/OPS is harder on the
> administrator. All that happens is that a second instance recovers
> transactions from the failed instance. All else runs along same old same
> old. What in that requires "complication" from the administrator?

Beg to differ. Tuning the beasties is a good deal more complicated than tuning a single instance. Care to knock up a GC_FILES_TO_LOCKS parameter off the top of your head? And where's the white paper to tell me what a good number of multi-block locks is, and when I should group them? And where did Oracle get the idea from that LMS and LMD should have latencies of less than 20 milliseconds? Where'd that number come from?! Where's the scientific proof behind the statistics that the Oracle doco. trots out for these RAC-specific tuning measures?

Regards
HJR Received on Tue Aug 12 2003 - 01:43:20 CDT