If I understood him, at cache 1 he was faster with his table solution anyway,
so that is what he is using.
But your point is absolutely correct and has been since Oracle introduced
sequences. And that was at least by V6 (1988).
From: oracle-l-bounce@xxxxxxxxxxxxx [mailto:oracle-l-bounce@xxxxxxxxxxxxx] On ;
Behalf Of Mark J. Bobak
Sent: Thursday, December 17, 2020 11:59 AM
To: Powell, Mark
Cc: ORACLE-L
Subject: Re: Doubt related do INITRANS
Cache 1? That doesn't make sense. There is no justifiable reason to force
cache 1. An Oracle sequence can't guarantee gapless sequences, regardless of
the cache setting. If you really *need* a gapless sequence, you can't use a
sequence. Or better, revisit the design to avoid the requirement for the
gapless sequence.
-Mark
On Thu, Dec 17, 2020 at 11:53 AM Powell, Mark <mark.powell2@xxxxxxx> wrote:
Eriovaldo, I hope your design scales. You will want to test scalability as
best you can, as early in development as you can before this hits production.
Mark Powell
Database Administration
(313) 592-5148
_____
From: oracle-l-bounce@xxxxxxxxxxxxx <oracle-l-bounce@xxxxxxxxxxxxx> on behalf
of Eriovaldo Andrietta <ecandrietta@xxxxxxxxx>
Sent: Thursday, December 17, 2020 10:03 AM
To: Jonathan Lewis <jlewisoracle@xxxxxxxxx>
Cc: ORACLE-L <oracle-l@xxxxxxxxxxxxx>
Subject: Re: Doubt related do INITRANS
Jonathan,
Unfortunately I need to use cache equal to 1. This is my strong requirement.
It involves the whole infrastructure dimension and now, events of wait in the
database are being supported by the infrastructure and under control.
Many thanks for attention.
Regards
Eriovaldo
Em qui., 17 de dez. de 2020 às 11:41, Jonathan Lewis <jlewisoracle@xxxxxxxxx>
escreveu:
It seems a little unlikely that you could do better with a home-grown solution
than with an Oracle sequence - but you may be doing something that eliminates
the benefit of Oracle sequences but allows you to get extra mileage from your
own. A critical error that appears frequently with Oracle sequences is that
people forget to increase the cache (which defaults to 20) to something more
appropriate - worst case scenario they set it to "nocache".
In you case, if you're going with a table with a few independent rows you could
benefit by rigging the table so that you get one row per block. Set pctfree =
99 and, if necessary, add a column of varchar2(90) and a value of 90 characters
in it so that the block is full as soon as you insert a row. This would avoid
contention between sequences.
Regards
Jonathan Lewis
On Thu, 17 Dec 2020 at 14:31, Eriovaldo Andrietta <ecandrietta@xxxxxxxxx> wrote:
Hi Jonathan,
I got the message.
The application can also access other lines with less frequency, and only one
line exhaustively.
Based on the concept of the initrans, I will keep the initrans greater than 1
(I am using 10) for this table.
I had already done some tests using an Oracle sequence object and I got better
performance with the table using 1 million lines. I need to review my tests.
Now, everything is clear.
Thanks for answering.
Regards
Eriovaldo
Em qui., 17 de dez. de 2020 às 08:36, Jonathan Lewis <jlewisoracle@xxxxxxxxx>
escreveu:
Of course, if you have 20 rows in this block of your "sequences" table and a
couple of hundred users who are constantly trying to increment one or other of
the sequences then you might need to set initrans to 20 to make sure that every
row in the block can be locked by a different transaction at the same time.
(At that point you might find that you run into problems with "buffer busy
wait" waits - as Oracle can do on its SEQ$ table if someone creates a load of
sequences with a low (or zero) sequence cache size.
Regards
Jonathan Lewis
On Thu, 17 Dec 2020 at 11:30, Jonathan Lewis <jlewisoracle@xxxxxxxxx> wrote:
Eriovaldo,
The purpose of INITRANS (pre-allocating an "interested transaction list") is to
ensure that N different transactions can change (at least) N different rows in
the block at the same time. If you're going to have only one row in the block
then at most one tranaction can change the row at any one moment, any other
transactions that want to change that row will queue up on the "transaction id"
- i.e. the transaction table entry in the undo segment header - of the
transaction that is currently modifying the row and wait (showing a TX lock
waiting (usually) for a mode 6 lock) until the first transaction commits.
This means, among other things, that you won't need to set initrans to N
because you think there will be N transactions trying to update that row,
leaving it to default will be perfectly adequate. Try the experiment - create
the table , insert the row, try to update it from 20 different sessions without
committing. then commit each one in turn as it get to the front of the queue.
Then dump the block - you'll find that the ITL still consists of two entries.
Regards
Jonathan Lewis
