Yes, I wrote “If the buffer had to flush along the way before the commit, they
would already be in the redo stream,” and I should have written that also
applies if the dirty current block were written (with respect to the row lock,
as you wrote, the current block changed even if no value changed).
I may be confused, but I think all that could be sorted out in memory in the
usual precedence as long as the data block, redo, and undo have not yet been
written by the time of commit or rollback.
Recalling your cases where the “no change” didn’t turn out as less work, this
might end up being more work in time to save writing anyway.
I won’t argue it’s a win to not lock the row in the first place, but it might
be if you know the before and after value are equal I’m not sure I can think of
a case where that comes out wrong. (From your note, I bet you already have
something in mind.)
The entire thing is so good as a non-leaking bucket I doubt upsetting it ends
up being worthwhile, anyway. I think of the change from block level before
image file through 5 to redo and undo vectors in 6 that unleashed the
“transaction processing option” as epochal. (The only thing I thought they got
wrong was that physical backup probably would have been better at the file
level than at the tablespace level, leaving 40-bazillion things they got right.)
At the query level I wonder if permuting the query when it can be certainly
iso-functional as
update someTable
set someColumn1 = 1.234
becomes
update someTable
set someColumn1 = 1.234
where someColumn1 != 1.234
makes more sense than trying to stuff it down a layer.
mwf
From: oracle-l-bounce@xxxxxxxxxxxxx [mailto:oracle-l-bounce@xxxxxxxxxxxxx] On ;
Behalf Of Jonathan Lewis
Sent: Thursday, June 17, 2021 5:53 PM
To: ORACLE-L
Subject: Re: Original design approach to Oracle REDO Logs
Oracle actually introduced a "no change" optimisation for redo in 12.2
If an update doesn't change the values in the row then Oracle doesn't go
through the normal generation of undo and redo for that row. HOWEVER - and
this is where the "don't change the block" idea fails - it locks the row, which
does generate undo and redo. And if you're unlucky (and if you're updating
every row in the block in a small way you probably will be) the "optimisation"
will generate more undo and redo rather than less because real updates can
generate one undo and redo vector per block but the lock row vectors are (up to
19.3) single row vectors.
I wrote about this a couple of years ago:
https://jonathanlewis.wordpress.com/2019/09/08/quiz-night-34/
The "lock row" is necessary for backwards compatibility (if nothing else) - but
I think it would be hard to argue that it shouldn't happen because if the row
were not locked then it would be possible to create cases that produced wrong
results.
Regards
Jonathan Lewis
On Thu, 17 Jun 2021 at 21:12, Mark W. Farnham <mwf@xxxxxxxx> wrote:
While I agreed with what JL wrote, IF we knew how fast processors are now, then
an update all rows query might compare the block images on the fly and not
write the blocks where all the rows in a block were updated to the same value
so that the actual block image resulting from the update didn’t need to be
written.
Off the top of my head I *think* that would work and be efficient, but I am not
engaged enough to ponder possible edge cases. But I *think* you have all the
information you need in hand at commit time to flag blocks to *not* bother
writing. If the buffer had to flush along the way before the commit, they would
already be in the redo stream, but with the massive memory now available a
private redo thread might be able to handle it both correctly and efficiently.
I really like Oracle’s redo model though. Someone once put it to me thusly:
Everything Bill Bridge starts ends up in checkmate. (Meaning Bill wins.) Making
sure the bucket has no leaks was definitely priority number one.
From: oracle-l-bounce@xxxxxxxxxxxxx [mailto:oracle-l-bounce@xxxxxxxxxxxxx] On ;
Behalf Of Jonathan Lewis
Sent: Thursday, June 17, 2021 10:16 AM
To: ORACLE-L (oracle-l@xxxxxxxxxxxxx)
Subject: Re: Original design approach to Oracle REDO Logs
Mike,
I don't think you can ask about hypothetical strategies for the redo log before
stating what you think the purpose of the database should be.
If you expect the database to be the reference of last resort for the
correctness and (immediate) consistency of the data then you might like a
different strategy from someone who thinks the database is some sort of record
of evolution of the data that allows the current version of the data to be
re-imaged by a client.
Regards
Jonathan Lewis
On Thu, 17 Jun 2021 at 13:55, Michael D O'Shea/Woodward Informatics Ltd
<woodwardinformatics@xxxxxxxxxxxxxxxx> wrote:
Chaps, request for opinions/discussion/feedback .....
Question: If Oracle were written today, would the same strategy behind the
"redo log" be adopted.
An example .... for a table of 10,000,000 rows, and a contrived piece of DML
update someTable
set someColumn1 = 1.234
that "updated" all the rows
but where someColumn1 was only updated to 1.234 for 6 rows as the remaining
10,000,000 - 6 rows were already 1.234
Should (assuming just DML and also just the basic data types, number, varchar2,
date, ... ) the redo log
* record the data "change" for all 10,000,000 "updates"
* record the real data change for just the 6 real updates
* record solely the SQL used to perform the update for some playback purpose
* send a message to some message broker such as Solace, Tibco, .... allowing
subscribers to process the data change that way (might be replication/backup,
some sort of refresh or push notification to other applications, email dispatch
and so on)
* some other approach else
There is a considerable movement to event streaming technology such as Kafka to
(indirectly) drive data change events to downstream and dependent systems (in
Oracle assumably by polling the redo log file, or maybe some LogMiner interface
.. I don't know the detail) in databases that include Oracle, MongoDB (referred
to as Change Streams), and many more.
My "ask" focuses more on "just the database" interoperability with the
remainder of what is often a large tech stack, and not the original design
decision around redo logs for data recovery.
Mike
http://www.strychnine.co.uk
