There are 2 SQL-Transaction Statements:
- COMMIT Statement -- commit (make persistent) all changes for the current transaction
- ROLLBACK Statement -- roll back (rescind) all changes for the current transaction
Transaction OverviewA database transaction is a larger unit that frames multiple SQL statements. A transaction ensures that the action of the framed statements is atomic with respect to recovery.
A SQL Modification Statement has limited effect. A given statement can only directly modify the contents of a single table (Referential Integrity effects may cause indirect modification of other tables.) The upshot is that operations which require modification of several tables must involve multiple modification statements. A classic example is a bank operation that transfers funds from one type of account to another, requiring updates to 2 tables. Transactions provide a way to group these multiple statements in one atomic unit.
In SQL92, there is no BEGIN TRANSACTION statement. A transaction begins with the execution of a SQL-Data statement when there is no current transaction. All subsequent SQL-Data statements until COMMIT or ROLLBACK become part of the transaction. Execution of a COMMIT Statement or ROLLBACK Statement completes the current transaction. A subsequent SQL-Data statement starts a new transaction.
In terms of direct effect on the database, it is the SQL Modification Statements that are the main consideration since they change data. The total set of changes to the database by the modification statements in a transaction are treated as an atomic unit through the actions of the transaction. The set of changes either:
- Is made fully persistent in the database through the action of the COMMIT Statement, or
- Has no persistent effect whatever on the database, through:
- the action of the ROLLBACK Statement,
- abnormal termination of the client requesting the transaction, or
- abnormal termination of the transaction by the DBMS. This may be an action by the system (deadlock resolution) or by an administrative agent, or it may be an abnormal termination of the DBMS itself. In the latter case, the DBMS must roll back any active transactions during recovery.
In the relational model, each transaction is completely isolated from other active transactions. After initiation, a transaction can only see changes to the database made by transactions committed prior to starting the new transaction. Changes made by concurrent transactions are not seen by SQL DML query and modification statements. This is known as full isolation or Serializable transactions.
SQL92 defines Serializable for transactions. However, fully serialized transactions can impact performance. For this reason, SQL92 allows additional isolation modes that reduce the isolation between concurrent transactions. SQL92 defines 3 other isolation modes, but support by existing DBMSs is often incomplete and doesn't always match the SQL92 modes. Check the documentation of your DBMS for more details.
Transaction isolation controls the visibility of changes between transactions in different sessions (connections). It determines if queries in one session can see changes made by a transaction in another session. There are 4 levels of transaction isolation. The level providing the greatest isolation from other transactions is Serializable.
At transaction isolation level Serializable, a transaction is fully isolated from changes made by other sessions. Queries issued under Serializable transactions cannot see any subsequent changes, committed or not, from other transactions. The effect is the same as if transactions were serial, that is, each transaction completing before another one is begun.
At the opposite end of the spectrum is Read Uncommitted. It is the lowest level of isolation. With Read Uncommitted, a session can read (query) subsequent changes made by other sessions, either committed or uncommitted. Read uncommitted transactions have the following characteristics:
- Dirty Read -- a session can read rows changed by transactions in other sessions that have not been committed. If the other session then rolls back its transaction, subsequent reads of the same rows will find column values returned to previous values, deleted rows reappearing and rows inserted by the other transaction missing.
- Non-repeatable Read -- a session can read a row in a transaction. Another session then changes the row (UPDATE or DELETE) and commits its transaction. If the first session subsequently re-reads the row in the same transaction, it will see the change.
- Phantoms -- a session can read a set of rows in a transaction that satisfies a search condition (which might be all rows). Another session then generates a row (INSERT) that satisfies the search condition and commits its transaction. If the first session subsequently repeats the search in the same transaction, it will see the new row.
In Read Committed isolation level, Dirty Reads are not possible, but Non-repeatable Reads and Phantoms are possible. In Repeatable Read isolation level, Dirty Reads and Non-repeatable Reads are not possible but Phantoms are. In Serializable, Dirty Reads, Non-repeatable Reads, and Phantoms are not possible.
The isolation provided by each transaction isolation level is summarized below:
|Dirty Reads||Non-repeatable Reads||Phantoms|
Note: SQL92 defines the SET TRANSACTION statement to set the transaction isolation level for a session, but most DBMSs support a function/method in the Client API as an alternative.
SQL-Schema Statements in TransactionsThe 3rd type of SQL Statements - SQL-Schema Statements, may participate in the transaction mechanism. SQL-Schema statements can either be:
- included in a transaction along with SQL-Data statements,
- required to be in separate transactions, or
- ignored by the transaction mechanism (can't be rolled back).
COMMIT [WORK]WORK is an optional keyword that does not change the semantics of COMMIT.
ROLLBACK [WORK]WORK is an optional keyword that does not change the semantics of ROLLBACK.