SQL Server 2008 R2 : Locking and Performance - Transaction Isolation Levels in SQL Server

Isolation levels determine the extent to which data being accessed or modified in one transaction is protected from changes to the data by other transactions. In theory, each transaction should be fully isolated from other transactions. However, in practice, for practical and performance reasons, this might not always be the case. In a concurrent environment in the absence of locking and isolation, the following four scenarios can happen:

• Lost update— In this scenario, no isolation is provided to a transaction from other transactions. Multiple transactions can read the same copy of data and modify it. The last transaction to modify the data set prevails, and the changes by all other transactions are lost.

• Dirty reads— In this scenario, one transaction can read data that is being modified by other transactions. The data read by the first transaction is inconsistent because the other transaction might choose to roll back the changes.

• Nonrepeatable reads— In this scenario, which is somewhat similar to zero isolation, a transaction reads the data twice, but before the second read occurs, another transaction modifies the data; therefore, the values read by the first read are different from those of the second read. Because the reads are not guaranteed to be repeatable each time, this scenario is called nonrepeatable reads.

• Phantom reads— This scenario is similar to nonrepeatable reads. However, instead of the actual rows that were read changing before the transaction is complete, additional rows are added to the table, resulting in a different set of rows being read the second time. Consider a scenario in which Transaction A reads rows with key values within the range of 1 through 5 and returns three rows with key values 1, 3, and 5. Before Transaction A reads the data again within the transaction, Transaction B adds two more rows with the key values 2 and 4 and commits the changes. Assuming that Transaction A and Transaction B both can run independently without blocking each other, when Transaction A runs the query a second time, it now gets five rows with key values 1, 2, 3, 4, and 5. This phenomenon is called phantom reads because in the second pass, you get records you did not expect to retrieve.

Ideally, a DBMS must provide levels of isolation to prevent these types of scenarios. Sometimes, for practical and performance reasons, databases relax some of the rules. The American National Standards Institute (ANSI) has defined four transaction isolation levels, each providing a different degree of isolation to cover the previous scenarios. ANSI SQL-92 defines the following four standards for transaction isolation:

• Read Uncommitted (Level 0)

• Read Committed (Level 1)

• Repeatable Read (Level 2)

• Serializable (Level 3)

SQL Server 2008 supports all the ANSI isolation levels; in addition, SQL Server 2008 also supports two additional transaction isolation levels that use row versioning. One is an alternative implementation of Read Committed isolation called Read Committed Snapshot, and the other is the Snapshot transaction isolation level.

You can set the default transaction isolation for a user session by using the SET TRANSACTION ISOLATION LEVEL T-SQL command, or for individual SQL statements, you can specify table-level isolation hints within the query. 

Read Uncommitted Isolation

If you set the Read Uncommitted mode for a session, no isolation is provided to the SELECT queries in that session. A transaction that is running with this isolation level is not immune to dirty reads, nonrepeatable reads, or phantom reads.

To set the Read Uncommitted mode for a session, you run the following statements from the client:

• T-SQL— Use SET TRANSACTION ISOLATION LEVEL READ UNCOMMITTED.

• ODBC— Use the function call SQLSetConnectAttr with Attribute set to SQL_ATTR_TXN_ISOLATION and ValuePtr set to SQL_TXN_READ_UNCOMMITTED.

• OLE DB— Use the function call ITransactionLocal::StartTransaction with the isoLevel set to ISOLATIONLEVEL_READUNCOMMITTED.

• ADO— Set the IsolationLevel property of the Connection object to adXactReadUncommitted.

• ADO.NET— For applications using the System.Data.SqlClient managed namespace, call the SqlConnection.BeginTransaction method and set the IsolationLevel option to ReadUncommitted.

You need to be careful when running queries at Read Uncommitted isolation; it is possible to read changes that have been made to data that are subsequently rolled back. In essence, the accuracy of the results cannot be guaranteed. You should use this mode only when you need to get information quickly from an online transaction processing (OLTP) database, without affecting or being affected by the ongoing updates and when the accuracy of the results is not critical.

Read Committed Isolation

The Read Committed mode is the default locking-isolation mode for SQL Server. With Read Committed as the transaction isolation level, read operations can read pages only for transactions that have already been committed. No “dirty reads” are allowed. Locks acquired by update transactions are held for the duration of the transaction. However, in this mode, read requests within the transaction release locks as soon as the query finishes reading the data. Although this improves concurrent access to the data for updates, it does not prevent nonrepeatable reads or phantom reads. For example, within a transaction, a process could read one set of rows early in the transaction and then, before reading the information again, another process could modify the result set, resulting in a different result set being read the second time.

Because Read Committed is the default isolation level for SQL Server, you do not need to do anything to set this mode. If you need to set the isolation level back to Read Committed mode for a session, you run the following statements from the client:

• T-SQL— Use SET TRANSACTION ISOLATION LEVEL READ COMMITTED.

• ODBC— Use the function call SQLSetConnectAttr with Attribute set to SQL_ATTR_TXN_ISOLATION and ValuePtr set to SQL_TXN_READ_COMMITTED.

• OLE DB— Use the function call ITransactionLocal::StartTransaction with isoLevel set to ISOLATIONLEVEL_READCOMMITTED.

• ADO— Set the IsolationLevel property of the Connection object to adXactReadcommitted.

• ADO.NET— For applications using the System.Data.SqlClient managed namespace, call the SqlConnection.BeginTransaction method and set the IsolationLevel option to ReadCommitted.

Read Committed Snapshot Isolation

When the READ_COMMITTED_SNAPSHOT database option is set to ON, sessions running with the Read Committed isolation mode use row versioning to provide statement-level read consistency. When this database option is enabled and a transaction runs at the Read Committed isolation level, all statements within the transaction see a snapshot of the data as it exists at the start of the statement.

When the READ_COMMITTED_SNAPSHOT option is enabled for a database, SQL Server maintains versions of each row that is modified. Whenever a transaction modifies a row, an image of the row before modification is copied into a page in the version store, which is a collection of data pages in tempdb. If multiple transactions modify a row, multiple versions of the row are linked in a version chain. Queries running with Read Committed Snapshot isolation retrieve the last version of each row that had been committed when the statement started, providing a statement-level snapshot of the data.

In the Read Committed Snapshot isolation mode, read operations do not acquire shared page or row locks on the data. Therefore, readers using row versioning do not block other processes modifying the same data, and, similarly, processes modifying the data do not block the readers. In addition, because the read operations do not acquire locks, locking overhead is reduced. However, processes modifying data still block other processes modifying data because two operations cannot modify the same data at the same time. Exclusive locks on modified data are still acquired and held until the end of the transaction.

While locking overhead is reduced for read operations when using Read Committed Snapshot isolation, it does introduce overhead to maintain the row versions in tempdb. In addition, tempdb must have sufficient space to hold the row versions in addition to the space required for normal tempdb operations.

You might want to consider enabling the READ_COMMITTED_SNAPSHOT database option when blocking that occurs between read and write operations affects performance to the point that the overhead of creating and managing row versions is offset by the concurrency benefits. You may also consider using Read Committed Snapshot isolation when an application requires absolute accuracy for long-running aggregations or queries where data values must be consistent to the point in time that the query starts.

Repeatable Read Isolation

In Repeatable Read mode, SQL Server provides the same level of isolation for updates as in Read Committed mode, but it also allows the data to be read many times within the same transaction and guarantees that the same values will be read each time. Repeatable Read isolation mode prevents other users from updating data that has been read within the transaction until the transaction in which it was read is committed or rolled back. This way, the reading transaction does not pick up changes to the rows it read previously within the transaction. However, this isolation mode does not prevent additional rows (that is, phantom reads) from appearing in the subsequent reads.

Although preventing nonrepeatable reads is desirable for certain transactions, it requires holding locks on the data that has been read until the transaction is completed. This reduces concurrent access for multiple update operations and causes performance degradation due to lock waits and locking contention between transactions. It can also potentially lead to deadlocks.

To set Repeatable Read mode for a session, you run the following statements from the client:

• T-SQL— Use SET TRANSACTION ISOLATION LEVEL REPEATABLE READ.

• ODBC— Use the function call SQLSetConnectAttr with Attribute set to SQL_ATTR_TXN_ISOLATION and ValuePtr set to SQL_TXN_REPEATABLEREAD.

• OLE DB— Use the function call ITransactionLocal::StartTransaction with isoLevel set to ISOLATIONLEVEL_REPEATABLEREAD.

• ADO— Set the IsolationLevel property of the Connection object to adXact REPEATABLEREAD.

• ADO.NET— For applications using the System.Data.SqlClient managed namespace, call the SqlConnection.BeginTransaction method and set the IsolationLevel option to RepeatableRead.

Serializable Read Isolation

Serializable Read mode is similar to repeatable reads but adds to it the restriction that rows cannot be added to a result set that was read previously within a transaction. This prevents phantom reads. In other words, Serializable Read locks the existing data being read as well as rows that do not yet exist. It accomplishes this by locking the data being read. In addition, SQL Server puts locks on the range of values being read so that additional rows cannot be added to the range.

For example, say you run a query in a transaction that retrieves all records for the Sales table in the bigpubs2008 database for a store with the stor_id of 7066. To prevent additional sales records from being added to the Sales table for this store, SQL Server locks the range of values with stor_id of 7066. 

Although preventing phantom reads is desirable for certain transactions, Serializable Read mode, like Repeatable Read, reduces concurrent access for multiple update operations and can cause performance degradation due to lock waits and locking contention between transactions, and it can potentially lead to deadlocks.

To set Serializable Read mode for a session, you run the following statements from the client:

• T-SQL— Use SET TRANSACTION ISOLATION LEVEL SERIALIZABLE.

• ODBC— Use the function call SQLSetConnectAttr with Attribute set to SQL_ATTR_TXN_ISOLATION and ValuePtr set to SQL_TXN_SERIALIZABLE.

• OLE DB— Use the function call ITransactionLocal::StartTransaction with isoLevel set to ISOLATIONLEVEL_SERIALIZABLE.

• ADO— Set the IsolationLevel property of the Connection object to adXact SERIALIZABLE.

• ADO.NET— For applications using the System.Data.SqlClient managed namespace, call the SqlConnection.BeginTransaction method and set the IsolationLevel option to Serializable.

Snapshot Isolation

Snapshot Isolation is an additional isolation level available in SQL Server 2008. Similar to Read Committed Snapshot, Snapshot Isolation mode uses row versioning to take a point-in-time snapshot of the data. However, unlike Read Committed Snapshot isolation, which provides a statement-level snapshot of the data, Snapshot Isolation maintains a snapshot of the data for the duration of the entire transaction. A data snapshot is taken when the transaction starts and the snapshot remains consistent for the duration of the transaction.

Snapshot Isolation mode provides the benefit of repeatable reads without acquiring and holding shared locks on the data that is read. This can help minimize locking and blocking problems between read operations and update operations. Read operations do not have to wait for write operations and writes don’t have to wait for reads.

To set the Snapshot Isolation mode for a session, you run the following statement:

SET TRANSACTION ISOLATION LEVEL SNAPSHOT

In addition, to be able to request the Snapshot Isolation mode in a session, you must enable the database option ALLOW_SNAPSHOT_ISOLATION with the ALTER DATABASE command:

ALTER DATABASE dbname SET ALLOW_SNAPSHOT_ISOLATION ON

When Snapshot Isolation mode is enabled, SQL Server assigns a transaction sequence number to each transaction that manipulates data using row versioning. When either the READ_COMMITTED_SNAPSHOT or ALLOW_SNAPSHOT_ISOLATION database option is set to ON, SQL Server stores a version of the previously committed image of the data row in tempdb whenever the row is modified by a transaction. Each of these versions is marked with the transaction sequence number of the transaction that made the change. The versions of the modified rows are linked together in a chain, with the most recent version of the row always stored in the current database and the versioned rows stored in tempdb.

When a transaction requests a read of data, it searches the version chain to locate the last committed version of the data row with a lower transaction sequence number than the current transaction. Row versions are kept in tempdb only long enough to satisfy the requirements of any transactions running under row versioning–based isolation levels. SQL Server keeps track of the sequence number of the oldest outstanding transaction and periodically deletes all row versions stamped with transaction sequence numbers lower than that.

You might consider using snapshot isolation in the following instances:

• When you want optimistic concurrency control

• When it is unlikely that your transaction would have to be rolled back because of an update conflict

• When an application generates reports based on long-running, multistatement queries that must have point-in-time consistency

• With systems that are incurring a high number of deadlocks because of read/write contention

There is a risk to using snapshot isolation, however. If two client applications both retrieve the same data and then both attempt to write changes to the data back to the database, the second application could potentially overwrite the first application’s changes. This is called a lost update error. Fortunately, SQL Server 2008 resolves this problem by blocking the second transaction’s writes. So, although snapshot isolation provides benefits for resolving conflicts between read and write operations, there can still be conflicts between multiple write operations. For systems with heavy read and insert activity and with little concurrent updating of the same resource, snapshot isolation can provide a solution for concurrency issues.

Another cost of snapshot isolation is that it can make heavy use of tempdb. For this reason, you should locate tempdb on its own high-performance drive system.

DBCC USEROPTIONS
go

Set Option                          Value
---------------------------------- --------------
textsize                           2147483647
language                           us_english
dateformat                         mdy
datefirst                          7
lock_timeout                       -1
quoted_identifier                  SET
arithabort                         SET
ansi_null_dflt_on                  SET
ansi_warnings                      SET
ansi_padding                       SET
ansi_nulls                         SET
concat_null_yields_null            SET
isolation level                 snapshot