RE: FW: My worst nightmare - ORA-8103
Date: Thu, 12 Jun 2014 15:43:30 -0400
Message-ID: <03de01cf8676$97b92570$c72b7050$_at_rsiz.com>
To be specific, if a particular query can have all its column references satisfied by an index, that can happen if the CBO plans it that way without reference to the table. (Typically a fast full scan or an index range scan). The index does contain the rowid, so this can be useful for getting the rowids that you have for blocks that are corrupt to confirm the 480 missing rows you’ve been referencing.
WHEN it is the case that some index or other has all the essential columns you need, then you can get back those pieces without referencing the table itself. Unindexed columns are in the proverbial bit bucket in this case.
Usually folks do not index scalar numbers in tables, so this is typically not a complete solution, but knowing exactly the KEYs for which you are missing some data is very helpful if you are trying to reconstruct data.
Incomplete transactions that failed to be rolled back may affect the ability to read indexes as well (I’m really not sure), so +1 on Gerwin’s concerns below.
mwf
From: oracle-l-bounce_at_freelists.org [mailto:oracle-l-bounce_at_freelists.org] On Behalf Of Gerwin Hendriksen
Sent: Thursday, June 12, 2014 1:11 AM
To: Maureen English
Cc: mwf_at_rsiz.com; oracle-l
Subject: Re: FW: My worst nightmare - ORA-8103
Hello Maureen,
I think Marc might be right with index suggestion. What worries me is the fact you had a corrupted undo tablespace. I guess you used parameters like '_corrupted_rollback_segments' and/or '_offline_rollback_segments' to startup your database before able to create a new undo tablespace. Can you explain how this was done?
Can you also state what kind of errors you encountered in your database at the time you recognized a corrupted undo tablespace? The problem is that most of this problems are going with ORA-600 errors in the 4000 range (transaction layer in the RDBMS code). So the problem you see with the table involved giving ORA-8103 might be just one of the problems in your database. Maybe some others have simply not been touched yet (don't you see any errors in your alert file?).
If the database was openened with one of the underscore parameters I mentioned the only valid thing to do is recreating the database by export / import. This is probably a very big job, but in my opinion the only official Oracle supported way to run the database.
With kind regards,
Gerwin Hendriksen
2014-06-12 6:09 GMT+02:00 Maureen English <maureen.english_at_alaska.edu>:
Thanks! I updated our ticket with Oracle asking for info on this bug. I can't find anything about it on Metalink.
Regarding Mark's post, does this mean that the reason users aren't seeing any problems is because they are able to access the data because the application uses the indexes when querying?
We did have to create a new undo tablespace due to corruption in the undo tablespace.
At least one of the indexes got an ORA-8103 when running analyze on it. I didn't check the others. Now I'm starting to panic again. Oracle says that the data is lost, but now it might not be???
Interestingly enough, we were able to fix the other issues that we've found by recreating indexes, or stepping through a table using a value from a unique index and then find the corrupt or duplicate rows and delete them. This table doesn't have a unique index...and it has 553M+ rows. It takes 20 minutes just to get a rowcount back :-(
I ended up following Oracle's suggestion of copying the rows that didn't return the ORA-8102 error to another table, exported that table, truncated my application table, dropped the indexes on the application table and imported the data back in. I'm currently rebuilding the indexes. Tomorrow will be when the 2 main Finance users will check things out and let me know the status.
I'm so dreading the possibility of someone telling me that all of a sudden nothing balances...especially when it all seems to be in balance now....
- Maureen
On 6/11/2014 5:59 AM, Gerwin Hendriksen wrote:
Hello Maureen,
The ORA-8103 can occur when the block contains a vector to an undo segment and for some reason the change can't be found in the undo segment. Years ago I logged a bug at Oracle (at that time working for Oracle Support) describing a situation when recovering a database ending in an ORA-8103 or ORA-600 [4146], when selecting a table. Because the issue described here is in a reporting database, this bug although a very old bug could be underneath. The bug is caused by the way redo is applied and a small issue there might give the same situation as described in the bug. The bug was filed 31 march 2001 and has number 1714688, the bug is probably never really fixed because fixing it would mean a major design change in the redo apply. Further more the bug was encountered due to human error in the way the recovery was done. Just ask support that this bug might give an idea why you encounter the ORA-8103.
Good luck and kind regards,
Gerwin Hendriksen
2014-06-11 14:01 GMT+02:00 Mark W. Farnham <mwf_at_rsiz.com <mailto:mwf_at_rsiz.com>>:
Sorry, missed the list.
-----Original Message-----
From: Mark W. Farnham [mailto:mwf_at_rsiz.com <mailto:mwf_at_rsiz.com>] Sent: Tuesday, June 10, 2014 2:43 PM
To: 'maureen.english_at_alaska.edu <mailto:maureen.english_at_alaska.edu>' Subject: RE: My worst nightmare - ORA-8103
If a unique index survives, you should be able to retrieve the rowids for the rows not in your recovery target table with a minus. Then, using the rowids and column sets for indexes, you can retrieve at least those column values from the indexes. This at least should provide you with a complete set of keys for the missing rows.
I presume you lack the ability to do individual block patching via RMAN, etc.
good luck,
mwf
-----Original Message-----
From: oracle-l-bounce_at_freelists.org <mailto:oracle-l-bounce_at_freelists.org> [mailto:oracle-l-bounce_at_freelists.org
<mailto:oracle-l-bounce_at_freelists.org>]
On Behalf Of Maureen English
Sent: Tuesday, June 10, 2014 2:29 PM
To: oracle-l_at_freelists.org <mailto:oracle-l_at_freelists.org> Subject: My worst nightmare - ORA-8103
Hi,
Original post to BOracle list...apologies to those who are seeing this again.
On 5/31, the refresh of a materialized view in our reporting instance failed with an ORA-8103 error, as did a gather stats job for the same table in the production database. We had an application upgrade done on 6/1 and copied our production database to a preprod version on 6/2, so the errors weren't caught immediately.
Users complained that the current data wasn't available in the reporting instance but my attempts to refresh and recreate all failed with the same ORA-8103 error.
I've been working with Oracle Support since early last week and keep hitting brick walls. At the moment, we've managed to copy 'uncorrupted' rows out of the table in our preproduction database and are 480 rows short in a table with 550M rows in it. Oracle is currently working to identify the rowids of the corrupt blocks based on the output in a trace file generated by a failing query.
I'm looking at Document 336133.1 while I wait for more info from Oracle.
Anyone have any comments/suggestions/other info that might help identify and fix the problem? We really can't afford to lose that much data. We're working on recovering our database to a different location to try to get back any data that we lose, but since I don't have any idea what caused the corruption, I'm lost.
- Maureen
--
http://www.freelists.org/webpage/oracle-l
--
http://www.freelists.org/webpage/oracle-l
-- Maureen English Lead Database Administrator University of Alaska Fairbanks, AK (907) 450-8329 <tel:%28907%29%20450-8329> -- http://www.freelists.org/webpage/oracle-lReceived on Thu Jun 12 2014 - 21:43:30 CEST