Re: ORA-01555: snapshot too old

"Thomas Kyte" <tkyte_at_us.oracle.com> wrote in message news:9jbt920j5v_at_drn.newsguy.com...
> In article <3b58fa12_at_usenet.per.paradox.net.au>, "Howard says...
> >
> >No, I'm not kidding.
> >
> >Not for dictionary managed tablespace, anyway (as I carefully pointed out
> >originally).
> >
>
> I saw that. Doesn't matter, the link and papers I pointed you to contain
 much
> information the pre-dates LMTs alltoghter ("HOW TO STOP DEFRAGMENTING AND
 START
> LIVING: THE DEFINITIVE WORD ON FRAGMENTATION" for example)
>
> >The clusters used to record the extents is sized for around half a dozen
> >extents. Any more than that, and you introduce chaining on the data
> >dictionary tables.
> >
>
> again, I'd love to see the test case where we see material proof that this
> nominally affects performance in real life.
>
> I'd really like to see that, never have though.
>
> Chaining in itself is not a horrific thing. Especially in a cluster which
 are
> typically designed to not be full scanned (index access). Here we have a
 linked
> list of all of the data we need.
>
> Also, most of the extent info for the table is read from the segment
 header
> anyhow. Given the dictionary caching that goes on....
>
> The reason I'm hot on this is cause I see people get all paranoid, export
 data,
> reorg, going nuts all of the time to get the extents down.... And their
 systems
> run no faster, no slower - they just run with less availability cause they
 keep
> scheduling all of this down time.

Well, all that downtime's just plain daft, and I agree with you in that regard. 6 is good. 12 won't kill you. Neither will 120. But the clusters are sized for around 6, and that's what I'd be aiming for. Feel free to chain your own data dictionary, but I'd rather not.

>
> Give me the conclusive example that shows that hundreds of extents
 materially
> affects performance in the real world and I'll not bring it up again.
>
> >And, secifically in regard to rollback segments, since transactions can
> >share extents, can you tell me why extension would be more likely with 6
> >extents of 1000 blocks each, or 20 extents of 300 blocks each?
> >
>
> you have one transaction that modified 1 row. That guy went to lunch --
 he'll
> be back soon but until he does, that 1,000 block extent is wedged. You're
 RBS
> will start extend, 1,000 blocks at a time.

No it won't... it will start to extend, 1000 blocks at a time, only when the other 5000 blocks have been filled up, and we start wishing to move back into the first extent.

Given an appropriately sized rollback segment, the number of extents issue is a dead one, and there is no need for 20 (unless you are doing OCP, in which case 20 is the "right" answer).

Guys going to lunch and leaving blocking transactions floating around will cause extension problems however many extents you go for, if you wait long enough.

HJR
>
> >It makes no difference at all to the probability of having to extend the
> >rollback segment.
> >
>
> Not true. It depends whether that little transaction was at the
 "beginnging of
> the 1000 block extent or the end -- if it was at the "end", there is a
 good
> chance he would commit before you got through the first two 300 block
 pieces of
> the extent and then the rbs would not extend.
>
> But, even if they do extend, one grows as a slower rate.
>
> Consider the suggestions that has always been true (cut from various
 support
> notes):
>
> ...
> MINEXTENTS:
> -----------
>
> Set MINEXTENTS to 20, this will make it unlikely that the rollback segment
> needs to grab another extent because the extent that should move into is
 still
> being used by an active transaction
> .........
>
> Why size a rollback segment with a 'minimum' of twenty extents?
>
> Rollback segments dynamically allocate space when required and deallocate
 space
> when no longer needed (if the OPTIMAL parameter is used). The fewer
 extents
> that a rollback segment consists of, the larger the less granular these sp
 ace
> allocations and deallocations are. For example, consider a 200 megabyte
> rollback segment which consists of only two 100-megabyte extents. If this
> segment were to require additional space, it would allocate another 100M
> extent. This immediately increases the size of the rollback segment by
 50% and
> potentially acquires more space than is really needed. By contrast, if
 the
> rollback segment consisted of twenty 10-megabyte extents, any additional
 space
> required would be allocated in 10-megabyte pieces. When a rollback
 segment
> consists of twenty or more extents, any single change in the number of
 extents
> will not move the total size of the rollback segment by more than 5%,
 resulting
> in a much smoother allocation and deallocation of space.
> .....
>
>
>
>
> >HJR
> >
> >
> >"Thomas Kyte" <tkyte_at_us.oracle.com> wrote in message
> >news:9j6rk902rpg_at_drn.newsguy.com...
> >> In article <3b566412_at_usenet.per.paradox.net.au>, "Howard says...
> >> >
> >> >Oh go on, I'll add my 2 cents'-worth...
> >> >
> >> >Dave's right in saying that increasing minextents will help cure
 1555's.
> >> >But that's frankly a daft way to increase the size of rollback
 segments
> >> >(well, OK, not daft, but not entirely wholesome, either).
> >> >
> >> >If you're using dictionary managed tablespace, then no segment should
> >> >(ideally) have more than around half a dozen extents or so, and that
 goes
> >> >for rollback segments, too.
> >> >
> >>
> >> your kidding. What's the basis in reality for that comment about not
 have
 more
> >> the 6 extents?
> >>
> >> Most all rbs's should have more then 6 extents.
> >>
> >> Having a couple of hundred extents is no big deal. Where is the
 scientific
> >> proof, case study, example even that shows otherwise?
> >>
> >> See
> >>
>
>http://asktom.oracle.com/pls/ask/f?p=4950:8:::::F4950_P8_DISPLAYID:73028925
 9
> >844
> >>
> >>
> >>
> >> >Why not simply recreate your rollback segments with the same number of
> >> >extents, but make the extent sizes bigger? In other words, play
 around
 with
> >> >INITIAL and NEXT, not MINEXTENTS.
> >> >
> >> >Regards
> >> >HJR
> >> >
> >> >"Dave Wotton" <Dave.Wotton_at_dwotton.nospam.clara.co.uk> wrote in
 message
> >> >news:8Qa57.85596$Do6.3950598_at_nnrp4.clara.net...
> >> >>
> >> >> "Vincent Ventrone" <vav_at_brandeis.edu> wrote in message
 news:9ivkms$qni$1_at_new-news.cc.brandeis.edu...
> >> >> >>"Guang Mei" <gmei_at_proteome.com> wrote in message
 news:kcI47.30B6.4620_at_news.shore.net...
> >> >> >> We have a cron job every night to exp some schemas(oracle 8.05 on
 Sun).
 We
> >> >> >> got the following error during last night's exp for the first
 time:
> >> >> >>
> >> >> >> EXP-00008: ORACLE error 1555 encountered
> >> >> >>
> >> >> > I'm figuring that you did, indeed, have active transactions going
 on
 in
 the
> >> >> > database during the export & one or more of these transactions
 issued
 a
> >> >> > COMMIT before the export ended. When a transaction COMMITs, Oracle
 makes
 the
> >> >> > undo records available to be overwritten by other transactions
 even
 though
> >> >> > it also considers these undo records to be "inactive, in-use" if
 some
 other
> >> >> > operation needs them for read-consistency. The text of the error
 message --
> >> >> > "rollback segment too small" -- is misleading. Once a transaction
 commits,
> >> >> > the space it was using in the rollback segment is now up for
 rabs --
 size
> >> >> > is not the issue. n other words it's a scheduling problem. You
 have
 two
> >> >> > options I think:
> >> >> >
> >> >> > 1. Find a different time for the export or figure out some way to
 ensure
> >> >> > that it has exclusive use of the database (ould be hard to do.)
> >> >> >
> >> >> > 2. Run the export with the parameter CONSISTENT=Y. This
 parameter
 sets
 a
> >> >> > consistency point for the *entire* export operation and it will
 then
 have
> >> >> > its own undo records to maintain consistency as of the begining of
 the
> >> >> > export operation regardless of what else is going on in the
 database.
 I
> >> >> > haven't tested this proposition myself, but I think this would
 solve
 your
> >> >> > rpoblem. Just make sure that your rollback segments can grow
 becuase
 the
> >> >> > export itself will end up generating a lot of undo if there is a
 lot
 of
 DML
> >> >> > going on while it is running, since the entire export operation is
 now
 one
> >> >> > big transaction.
> >> >>
> >> >> Vincent's explanation of why snapshot too old has occurred is
 correct,
 but
> >> >> his second recommendation is wrong.
> >> >>
> >> >> Using the consistent=y option will indeed make the whole of the
 export
> >> >> internally consistent, but it won't avoid the snapshot too old
 message -
> >> >> in fact it will make it more likely. The export is basically only
 running
> >> >> selects, it isn't doing any DML and so doesn't generate any undo
 records
 to
> >> >> maintain consistency - it is relying on the undo records created by
 other
> >> >> DML statements not being overwritten, just as it does when the
 consistent=y
> >> >> option is not used. Since consistent=y means the database has to
 maintain
 a consistent view of the entire database from
> >> >> the time the export starts
> >> >> (rather than just a consistent view of each table, from the time
 each
> >> >> table is exported), it is much more likely that a required undo
 record
 will
> >> >> be overwritten in that time. Using consistent=y is a good idea,
 particularly
> >> >> if you hope to be able to import more than one table from the export
 and
> >> >> have them consistent, but it won't help your snapshot too old
 problem.
> >> >>
> >> >> As you know, increasing max-extents for your rollback segments won't
 help,
> >> >> because they're not extending anyway, for the reasons Vincente
 pointed
 out.
> >> >> However, recreating your rollback segments with *min-extents*
 significantly
> >> >> bigger *will* help because you are pre-allocating them, so there is
 more
> >> >> space in the rollback segment to use for undo records before they
 are
> >> >> overwritten. If there is sufficient space in the rollback segment to
> >> >> accommodate all the undo generated by any DML running during the
 export,
> >> >> no undo records will be overwritten, so no snapshot too old will
 occur.
> >> >>
> >> >> Dave.
> >> >> --
> >> >> If you reply to this newsgroup posting by email, remove the "nospam"
> >> >> from my email address first.
> >> >>
> >> >>
> >> >>
> >> >>
> >> >>
> >> >
> >> >
> >>
> >> --
> >> Thomas Kyte (tkyte@us.oracle.com) http://asktom.oracle.com/
> >> Expert one on one Oracle, programming techniques and solutions for
 Oracle.
> >> http://www.amazon.com/exec/obidos/ASIN/1861004826/
> >> Opinions are mine and do not necessarily reflect those of Oracle Corp
> >>
> >
> >
>
> --
> Thomas Kyte (tkyte@us.oracle.com) http://asktom.oracle.com/
> Expert one on one Oracle, programming techniques and solutions for Oracle.
> http://www.amazon.com/exec/obidos/ASIN/1861004826/
> Opinions are mine and do not necessarily reflect those of Oracle Corp
>
Received on Sat Jul 21 2001 - 22:50:52 CDT