TROUBLESHOOTING GUIDE (TSG) - ORA-1555 [ID 467872.1] | ||
|
||
修改时间 05-OCT-2011 类型 HOWTO 状态 PUBLISHED |
In this Document
Goal
Solution
References
There are many good notes on ORA-1555 causes and solutions. In order to be able to resolve a problem, we must first understand the problem.
We will examine an extended example of the most common cause of the ORA-1555.
EXAMPLE:
Time 1: Session #1 starts a query on table A
Time 2: Session #2 updates row X in table A
Time 3: The query from session #1 arrives at row X and discovers the last updated time
(based on the SCN number) of the row to be later than Time 1 ... so the UNDO or
LOB segments are used to get the read consistent view (before image) of the row
and then the query proceeds
Time 4: Session #2 updates row Y in table A and then COMMITs (thus making it possible
for this transaction slot to be overwritten)
Time 5: Session #2 updates row Z in table B and COMMITs ... either due to space pressure
(using AUM) or bad luck (using rollback) in the read consistent view for the update
of row Y in table A at time 4 ... is overwritten (we wont examine why the overwrite
occurs at this point)
Time 6: The query from session #1 arrives at row Y and and discovers the last updated time
(based on the SCN number) of the row to be later than Time 1 ... so the UNDO or
LOB segments are examined to find the read consistent view ... BUT ... the
transaction slot containing the view ... was overwritten at time 5 ... so no read
consistent view is available ... so an ORA-1555 occurs
NOTE ... UNDO is a generic term that can refer to either UNDO (AUM) or Rollback segments
==============================================================================================
How can AUM allow a read consistent view of a transaction slot for a committed row to be overwritten?
1) The transaction slot for the row has expired.
This means that the current time minus the commit time of the row is greater than the
UNDO_RETENTION. Once a transaction slot of a committed row is 'expired', it is available for
reuse. Often we get customers who ask:
Why do my queries sometimes run so much longer than my UNDO_RETENTION before
encountering an ORA-1555 and other times it occurs soon after?
The answer is all left up to chance and how much pressure the undo tablespace is under.
2) The transaction slot for the row has NOT expired and yet was still overwritten.
This condition occurs under two conditions:
a) The undo tablespace has become full and must 'steal' transaction slots from committed, unexpired
rows
b) A bug
3) The LOB segment read consistent copy of the LOB is no longer available
Investigate how the LOB column is configured, in-row or out-of-row. In-row LOBs should be utilizing
normal UNDO algorithms in the UNDO tablespace. Out-of-row LOBs use the information below.
Read consistency with LOBs is controlled by 2 methods
a) Use of PCTVERSION (old method)
Specify the maximum percentage of overall LOB storage space used for maintaining old versions of
the LOB. The default value is 10, meaning that older versions of the LOB data are not overwritten
until they consume 10% of the overall LOB storage space.
The downfall to this method is customers who do frequent updates or deletes of tables with LOBs
often exceed even 100% PCTVERSION ...
b) Use of RETENTION (current method for use with AUM)
Use this clause to indicate that Oracle Database should retain old versions of this LOB column.
Oracle Database uses the value of the UNDO_RETENTION initialization parameter to determine
the amount of committed undo data to retain in the database.
This method uses the same expiration as undo segments
If an ORA-1555 occurs with a LOB segment. then this means that either:
PCTVERSION was exceeded and the read consistent view of the LOB was overwritten
OR
the LOB has exceeded RETENTION and the row LOB was overwritten sometime during
the query that generated the ORA-1555
New information
LOB problems with ORA-1555 can be very difficult to fix. Updates to out-of-row LOBs
should be minimized. Deletes and inserts have much less impact to UNDO operations than
updates. PCTVERSION is the preferred approach with 10.2.x and AUM. However,
100% will mean that lots of disk space is required to keep up with UNDO "copies" of LOB
information in environments with frequent LOB updates.
In some cases it has helped to move the LOB column to a tablespace where the Segment
Space Management is manual.
NOTE: transaction slots for UNCOMMITTED rows cannot be overwritten
Note 422826.1 How To Identify LOB Segment Use PCTVERSION Or RETENTION
Note 563470.1 Lob retention not changing when undo_retention is changed
Note 800386.1 ORA-1555 - UNDO_RETENTION is silently ignored if the LOB
resides in a MSSM tablespace
=======================================================================================
Now that we understand why the ORA-1555 occurs and some aspects about how they can occur we need to examine the following:
How can we logically determine and resolve what has occurred to cause the ORA-1555?
1) Determine if UNDO_MANAGEMENT is MANUAL or AUTO
If set to MANUAL, it is best to move to AUM. If it is not feasible to switch to AUM see
Note 69464.1 Rollback Segment Configuration & Tips
to attempt to tune around the ORA-1555 using V$ROLLSTAT
If set to AUTO, proceed to #2
2) Gather the basic data
a) Acquire both the error message from the user / client ... and the message in the alert log
User / Client session example:
ORA-01555: snapshot too old: rollback segment number 9 with name "_SYSSMU1$" too small
Alert log example
ORA-01555 caused by SQL statement below (Query Duration=9999 sec, SCN:0x000.008a7c2d)
b) Determine the QUERY DURATION from the message in the alert log
From our example above ... this would be 9999
c) Determine the undo segment name from the user / client message
From our example above ... this would be _SYSSMU1$
d) Determine the UNDO_RETENTION of the undo tablespace
show parameter undo_retention
3) Determine if the ORA-1555 is occurring with an UNDO or a LOB segment
If the undo segment name is null ...
ORA-01555: snapshot too old: rollback segment number with name "" too small
or the undo segment is unknown
ORA-01555: snapshot too old: rollback segment number # with name "???" too small
then this means this is a read consistent failure on a LOB segment
If the segment_name or the undo segment is known the error is occurring with an UNDO segment.
==============================================================================================
What to do if an ORA-1555 is occurring with an UNDO segment
-------------------------------------------------------------------------------------------------
1) QUERY DURATION > UNDO_RETENTION
There are no guarantees that read consistency can be maintained after the transaction slot for the
committed row has expired (exceeded UNDO_RETENTION)
Why would one think that the transaction slot's time has exceeded UNDO_RETENTION?
Lets answer this with an example
If UNDO_RETENTION = 900 seconds ... but our QUERY DURATION is 2000 seconds ...
This says that our query has most likely encountered a row that was committed more than 900
seconds ago ... and has been overwritten as we KNOW that the transaction slot being examined
no longer matches the row we are looking for
The reason we say "most likely" is that it is possible that an unexpired committed transaction slot was
overwritten due to either space pressure on the undo segment or this is a bug
SOLUTION:
The best solution is to tune the query can to reduce its duration. If that cannot be done then increase
UNDO_RETENTION based on QUERY DURATION to allow it to protect the committed
transaction slots for a longer period of time
NOTE: Increasing UNDO_RETENTION requires additional space in the UNDO tablespace. Make
sure to accommodate for this space. One method of doing this is to set AUTOEXTEND on one or
more of the UNDO tablespace datafiles for a period of time to allow for the increased space. Once the
size has stabilized, AUTOEXTEND can be removed.
See the solution for #2 below for more options
2) QUERY DURATION <= UNDO_RETENTION
This case is most often due to the UNDO tablespace becoming full sometime during the time when the
query was running
How do we tell if the UNDO tablespace has become full during the query?
Examine V$UNDOSTAT.UNXPSTEALCNT for the period while the query that generated the
ORA-1555 occurred.
This column shows committed transaction slots that have not exceeded UNDO_RETENTION but
were overwritten due to space pressure on the undo tablespace (IE became full).
If UNEXPSTEACNT > 0 for the time period during which the query was running then this shows
that the undo tablespace was too small to be able to maintain UNDO_RETENTION. Unexpired
blocks were over written, thus ending read consistency for those blocks for that time period.
EXAMPLE:
Tue May 26 16:16:57 2009
ORA-01555 caused by SQL statement below (SQL ID: 54yn3n36w24ft, Query Duration=922 sec, SCN: 0x0007.8a55f4e3)
922 seconds is 15 min 22 seconds
So the start time of the query would be May 26 16:01:35
If after researching, you find no existing bugs seem to match your case, then additional info will be
needed to file a bug.
Starting with release 9.2.0.4 an event, 10442, can be set to dump all the necessary diagnostics
information to one trace file. NOTE: Event 10442 does not trigger diagnostics in cases where the
ORA-1555 error is associated with out-of-row LOB undo.
The event can be enabled using the following settting in init.ora
event="10442 trace name context forever, level 10"
or
Alter system set events '10442 trace name context forever , level 10';
Reproduce the ORA-1555 error with a new connection to trigger the trace diagnostics file.
NOTE: There have been issues with 10g and 11g on RAC using 10442. The event can contribute to factors leading to a hang on one or more nodes on the RAC. Refer to Bug 10051817 for more fixes for these problems on RAC.
Another thing to consider is whether Oracle is behaving as designed as far as "steal logic" is concerned
Here is the logic for allocating space for a new transaction--based on Bug:4070480
* Allocate new extent in the undo tablespace
* If failed, steal an expired extent from another undo segment. This involves shrinking from the
other segment and adding to the current undo segment.
* If failed, autoextend a datafile (if enabled)
* If failed, reuse an unexpired extent from the current undo segment.
* If failed, steal an unexpired extent from another undo segment.
* If failed, report an "Out-Of-Space" error.
What to do if an ORA-1555 is occurring with a LOB segment
---------------------------------------------------------------------------------------------
1) Often the ORA-1555 with the null undo segment is accompanied by an ORA-22924
Error: ORA-22924
Text: snapshot too old
---------------------------------------------------------------------------
Cause: The version of the LOB value needed for the consistent read was already overwritten by another
writer.
2) Determine if PCTVERSION or RETENTION is used
Examine DBA_LOBS.PCTVERSION and DBA_LOBS.RETENTION
(or investigate through the data dictionary--see below)
NOTE: Only one of these may be set, so this will determine which method of undo management is
used for this LOB
There is a bug with these two columns prior to 10.2 and as such the data dictionary must be
examined to determine which of these is being used
Note 422826.1 How To Identify LOB Segment Use PCTVERSION Or RETENTION From
Data Dictionary
If PCTVERSION is used
------------------------------------
This method of LOB undo management is akin to use of rollback segments and as such there is little
control to tune undo usage.
The only parameter we have control over with this method is PCTVERSION itself.
SOLUTION:
Increase PCTVERSION