Re: Pluggable database in 12C

From: joel garry <>
Date: Thu, 18 Oct 2012 09:14:56 -0700 (PDT)
Message-ID: <>

On Oct 18, 2:21 am, "Jonathan Lewis" <> wrote:
> "Noons" <> wrote in message
> news:k5ogqf$bds$
> |
> | Last bit of info I have indicates they got it completely wrong: all PDBs
> share
> | the same container redo log set! Talk about contention...
> | One of the *biggest* advantages MSSQL has in this field is that it is
> possible
> | to optimize I/O for each PDB log set.
> | With 12c if it stays with a global redo log?  Ah yes, of course: buy
> Exadata!
> | That's gonna work really well...
> |
> That was such an obvious design flaw that I raised it at (I think) one of
> the Engineered Systems breakfast seminars.
> The point I made was in reply to the "you only need one of each background
> process for the whole system rather than one of each for each database."
> The follow-up answer to this was that you are able to define multiple log
> writers (not just I/O slaves for a single log writer). At that point I
> should have asked whether these multiple writers would behave like the
> multiple log writers you get in RAC, viz: separate log file groups that
> have to be resynchronized on recovery - but I didn't ask that question
> because it was such an obvious implementation detail that I didn't even
> think about thinking about it.
> Regards
> Jonathan Lewis
> Author: Oracle Core (Apress 2011)

My (admittedly muddled) thoughts on reading Noon's comments were: Maybe this could really be an extension, conceptually anyways, of private redo log threads (including the part about falling back to "old ways" when the "new ways" are inappropriate); and not to discount the Exadata ability to write the same thing to different devices including non-volatile memory - why assume that architecture would stay specific to Exadata (aside from the obvious milking maximum money)? Which kind of begs the question of how to deal with redo when you have no spinning rust at all (as already kind-of asked on the forums, and there was a Linus Q&A session on /. where he mentioned his personal machine).

The stock market discounts 6-12 months ahead. Physical DB architecture has to think 5-10 years ahead.


-- is bogus.
"I think they're Iranians" -
Received on Thu Oct 18 2012 - 18:14:56 CEST

Original text of this message