Avoiding ORA-1555 Errors

One of the most vexing problems that Oracle DBAs around the world face every day is this:

ORA-1555: snapshot too old: rollback segment number 9 with name "R07" too small

To most DBAs it is far from clear what might have caused the error, and even more perplexing as to how they can prevent it from occurring again. But perhaps the most exasperating thing about this error is that queries are most prone to it when they have been running for a long time, and thus many hours of processing can be lost.

The good news is that it is easy to prevent this error entirely and absolutely.

What does the error mean?

The ORA-1555 error means that a consistent get on a particular database block has failed.

When a transaction or query begins, the current SCN is recorded. That SCN serves as the snapshot SCN for the query or transaction. This term is derived from the requirement that the transaction or query must see a consistent snapshot of the database at that time.

Every block used to select rows for the query or transaction must reflect the state of the database at the snapshot SCN. This applies to the selection of rows to be updated or deleted, as much as it does to the selection of rows for a query. If a block has to be changed, then those changes will be applied to the current version of that block. However, the selection of the rows to be changed must be based on a version of the block consistent with the snapshot SCN. The temporary reconstruction of a version of the block consistent with the snapshot SCN is called a consistent get.

Why do consistent gets fail?

There are two types of consistent get failure: rollback failure, and cleanout failure.

Rollback failure

If the block has been modified in any way by another transaction since the snapshot SCN, then those changes must be rolled back for the consistent get. To do so, it is necessary to read the rollback segment data blocks to which the rollback information for those changes was written.

However, if any of those changes were made by a discrete transaction, then there will be no rollback information, because discrete transactions do not generate rollback information. If so, an ORA-1555 error will be raised. Similarly, an ORA-1555 error will be raised if the required rollback segment blocks are no longer available because the rollback segment extent containing those blocks has been deallocated in a shrink operation, or reused by subsequent transactions.

Note that the rollback segment blocks required are those that were used by any other transactions that have modified the block after the snapshot SCN. These blocks could be in any rollback segment in the database.

Note further that for those blocks to be unavailable by virtue of extent reuse, all extents in that rollback segment must have been used at least once since the snapshot SCN. This is why the error message suggests that the rollback segment is too small.

Cleanout failure

DBWn often writes a block to disk before the last transaction to modify that block has been committed. If so, the interested transaction list in the block header still shows that transaction as having an open interest in the block, and the row level locks in the row headers of the affected rows remain in force. When the block is read for another query or transaction, block cleanout must be performed. This involves finding out whether the previous transaction has committed, and if so its row level locks are cleaned out and the commit SCN for the transaction is record in the interested transaction list entry in the block header.

For a consistent get, block cleanout is necessary to establish the relative sequence of the commit SCN for the interested transaction and the snapshot SCN for the consistent get. If the interested transaction has not yet committed, or committed after the snapshot SCN, then rollback is required as described above. But if the interested transaction committed before the snapshot SCN, then no rollback of its changes is required.

To determine the commit SCN for an interested transaction, if it is not already recorded in the interested transaction list entry, and if it is no longer active, it is necessary to consult the transaction table in the rollback segment header block that was used by that transaction. The rollback segment number for the interested transaction is encoded in the interested transaction entry as part of the transaction identifier. However, the header block for that rollback segment may no longer contain a record for the interested transaction, because that block is also subject to change (lots of it) and the interested transaction may be ancient.

Fortunately, however, the consistent get does not need to determine the exact commit SCN for the interested transaction - only the relative sequence of the commit SCN and the snapshot SCN. It is therefore sufficient to perform a recursive consistent get on the rollback segment header block of the rollback segment for the interested transaction. If that consistent get is successful, and if the transaction header for the interested transaction is not extant in the consistent version of its rollback segment header block, and if the transaction identifier indicates that the transaction predated the snapshot SCN, then it may be concluded that the interested transaction committed in relative antiquity, and so no roll back is required.

(Incidentally, for current mode block cleanouts, the rollback segment header block is rolled back as far as possible, and the oldest available commit SCN for any transaction in that rollback segment at that time is recorded in the interested transaction list entry as the upper bound for its commit SCN. In other words, the transaction is marked as having committed no later than that SCN.)

However, it is possible for the consistent get on the rollback segment header block for an interested transaction to fail. This may occur if the rollback information for any of the changes to that rollback segment header block since the snapshot SCN are not available. These changes are written to that rollback segment itself, and are therefore subject to unavailability due to extent reuse or deallocation in the same way as other changes. However, there is more latitude in the case of rollback segment transaction table changes. Because slots in the transaction table are reused cyclically, the rollback segment extents themselves may have to be reused many times before the rollback information for the transaction header for an interested transaction will be rendered unavailable thereby.

How can you reduce the risk?

The following simple guidelines should be followed to reduce the risk of snapshot too old errors.

Do not run discrete transactions while sensitive queries or transactions are running, unless you are confident that the data sets required are mutually exclusive.
 
Schedule long running queries and transactions out of hours, so that the consistent gets will not need to rollback changes made since the snapshot SCN. This also reduces the work done by the server, and thus improves performance.
 
Code long running processes as a series of restartable steps.
 
Shrink all rollback segments back to their optimal size manually before running a sensitive query or transaction to reduce risk of consistent get rollback failure due to extent deallocation. This can be done with the APT script shrink_rollback_segs.sql.
 
Use a large optimal value on all rollback segments, to delay extent reuse. For an indication of how long you might have before the problem strikes, the APT script rollback_reuse_time.sql can be used to get the average time before rollback segment extent reuse.
 
Don't fetch across commits. That is, don't fetch on a cursor that was opened prior to the last commit, particularly if the data queried by the cursor is being changed in the current session.
 
Use a large database block size to maximize the number of slots in the rollback segment transaction tables, and thus delay slot reuse.
 
Commit less often in tasks that will run at the same time as the sensitive query, particularly in PL/SQL procedures, to reduce transaction slot reuse.
 
If necessary, add extra rollback segments to make more transaction slots available.

Note that adding extra rollback segments is somewhat in conflict with using a large optimal size, assuming the disk space available for rollback segments is invariant. The choice of a strategy at this point should depend upon the relative risk of consistent get rollback failures, as opposed to consistent get cleanout failures.

What can you do when all else fails?

For particularly sensitive queries and transactions, all this risk reduction is unnecessary. All that is needed is to prevent the deallocation or reuse of any rollback segment extents that have been used by any transaction subsequent to the snapshot SCN.

One simple way of doing that is to ensure that there is only one (large) rollback segment online from the time of the snapshot SCN, and to explicitly use that rollback segment for the sensitive query or transaction. This protects all extents in that rollback segment that may be used thereafter, from extent deallocation and reuse, until the conclusion of the sensitive transaction or query.

A more sophisticated variation on the same theme is to leave an uncommitted transaction in every online rollback segment. Of course, this introduces a risk of running out of space in the rollback segment tablespaces, but that risk is relatively easy to control. The following set of APT scripts can be used to apply this technique on Unix systems.

prevent_1555_setup.sql

This script creates a clustered table in the SYSTEM schema that is used to implement and record the protection of rollback segments from extent deallocation and reuse.

prevent_1555.sql

This is the main script of the set. It is called to ensure protection from ORA-1555 errors for a specified number of seconds. This script calls protect_rbs.sql in the background for each online rollback segment.

protect_rbs.sql

This script first shrinks the specified rollback segment to reduce the risk of running out of space in the rollback segment tablespaces. It then records its protection in the control table, before leaving an uncommitted transaction sleeping for the required number of seconds.

prevent_1555_wait.sql

This script must be run after the dummy transactions have been created in each online rollback segment. It waits for all older active transactions to finish. This is necessary in environments with other long-running transactions that may not yet have completed, because the earlier undo for those transactions is not protected and might be required by the critical report unless those transactions are allowed to finish before the critical report starts.

protected_rollback_segs.sql

This script is used to report the protection status of the rollback segments.

Copyright Ixora Pty Ltd Send Email Home