13 Data Concurrency and Consistency

Set the Isolation Level

Application designers, application developers, and database administrators can choose appropriate isolation levels for different transactions, depending on the application and workload. You can set the isolation level of a transaction by using one of these statements at the beginning of a transaction:

SET TRANSACTION ISOLATION LEVEL READ COMMITTED; 

SET TRANSACTION ISOLATION LEVEL SERIALIZABLE; 

SET TRANSACTION READ ONLY; 

To save the networking and processing cost of beginning each transaction with a SET TRANSACTION statement, you can use the ALTER SESSION statement to set the transaction isolation level for all subsequent transactions:

ALTER SESSION SET ISOLATION_LEVEL SERIALIZABLE; 

ALTER SESSION SET ISOLATION_LEVEL READ COMMITTED; 

Read Committed Isolation

The default isolation level for Oracle is read committed. This degree of isolation is appropriate for environments where few transactions are likely to conflict. Oracle causes each query to run with respect to its own materialized view time, thereby permitting nonrepeatable reads and phantoms for multiple executions of a query, but providing higher potential throughput. Read committed isolation is the appropriate level of isolation for environments where few transactions are likely to conflict.

Serializable Isolation

Serializable isolation is suitable for environments:

• With large databases and short transactions that update only a few rows

• Where the chance that two concurrent transactions will modify the same rows is relatively low

• Where relatively long-running transactions are primarily read only

Serializable isolation permits concurrent transactions to make only those database changes they could have made if the transactions had been scheduled to run one after another. Specifically, Oracle permits a serializable transaction to modify a data row only if it can determine that prior changes to the row were made by transactions that had committed when the serializable transaction began.

To make this determination efficiently, Oracle uses control information stored in the data block that indicates which rows in the block contain committed and uncommitted changes. In a sense, the block contains a recent history of transactions that affected each row in the block. The amount of history that is retained is controlled by the INITRANS parameter of CREATE TABLE and ALTER TABLE.

Under some circumstances, Oracle can have insufficient history information to determine whether a row has been updated by a "too recent" transaction. This can occur when many transactions concurrently modify the same data block, or do so in a very short period. You can avoid this situation by setting higher values of INITRANS for tables that will experience many transactions updating the same blocks. Doing so enables Oracle to allocate sufficient storage in each block to record the history of recent transactions that accessed the block.

Oracle generates an error when a serializable transaction tries to update or delete data modified by a transaction that commits after the serializable transaction began:

ORA-08177: Cannot serialize access for this transaction 

When a serializable transaction fails with the "Cannot serialize access" error, the application can take any of several actions:

• Commit the work executed to that point

• Execute additional (but different) statements (perhaps after rolling back to a savepoint established earlier in the transaction)

• Undo the entire transaction

Figure 13-2 shows an example of an application that rolls back and retries the transaction after it fails with the "Cannot serialize access" error:

Figure 13-2 Serializable Transaction Failure

Description of the illustration cncpt103.gif

Transaction Set Consistency

A useful way to view the read committed and serializable isolation levels in Oracle is to consider the following scenario: Assume you have a collection of database tables (or any set of data), a particular sequence of reads of rows in those tables, and the set of transactions committed at any particular time. An operation (a query or a transaction) is transaction set consistent if all its reads return data written by the same set of committed transactions. An operation is not transaction set consistent if some reads reflect the changes of one set of transactions and other reads reflect changes made by other transactions. An operation that is not transaction set consistent in effect sees the database in a state that reflects no single set of committed transactions.

Oracle provides transactions executing in read committed mode with transaction set consistency for each statement. Serializable mode provides transaction set consistency for each transaction.

Table 13-2 summarizes key differences between read committed and serializable transactions in Oracle.

Row-Level Locking

Both read committed and serializable transactions use row-level locking, and both will wait if they try to change a row updated by an uncommitted concurrent transaction. The second transaction that tries to update a given row waits for the other transaction to commit or undo and release its lock. If that other transaction rolls back, the waiting transaction, regardless of its isolation mode, can proceed to change the previously locked row as if the other transaction had not existed.

However, if the other blocking transaction commits and releases its locks, a read committed transaction proceeds with its intended update. A serializable transaction, however, fails with the error "Cannot serialize access", because the other transaction has committed a change that was made since the serializable transaction began.

Referential Integrity

Because Oracle does not use read locks in either read-consistent or serializable transactions, data read by one transaction can be overwritten by another. Transactions that perform database consistency checks at the application level cannot assume that the data they read will remain unchanged during the execution of the transaction even though such changes are not visible to the transaction. Database inconsistencies can result unless such application-level consistency checks are coded with this in mind, even when using serializable transactions.

Distributed Transactions

In a distributed database environment, a given transaction updates data in multiple physical databases protected by two-phase commit to ensure all nodes or none commit. In such an environment, all servers, whether Oracle or non-Oracle, that participate in a serializable transaction are required to support serializable isolation mode.

If a serializable transaction tries to update data in a database managed by a server that does not support serializable transactions, the transaction receives an error. The transaction can undo and retry only when the remote server does support serializable transactions.

In contrast, read committed transactions can perform distributed transactions with servers that do not support serializable transactions.

Read Committed Isolation

For many applications, read committed is the most appropriate isolation level. Read committed isolation can provide considerably more concurrency with a somewhat increased risk of inconsistent results due to phantoms and non-repeatable reads for some transactions.

Many high-performance environments with high transaction arrival rates require more throughput and faster response times than can be achieved with serializable isolation. Other environments that supports users with a very low transaction arrival rate also face very low risk of incorrect results due to phantoms and nonrepeatable reads. Read committed isolation is suitable for both of these environments.

Oracle read committed isolation provides transaction set consistency for every query. That is, every query sees data in a consistent state. Therefore, read committed isolation will suffice for many applications that might require a higher degree of isolation if run on other database management systems that do not use multiversion concurrency control.

Read committed isolation mode does not require application logic to trap the "Cannot serialize access" error and loop back to restart a transaction. In most applications, few transactions have a functional need to issue the same query twice, so for many applications protection against phantoms and non-repeatable reads is not important. Therefore many developers choose read committed to avoid the need to write such error checking and retry code in each transaction.

Serializable Isolation

Oracle's serializable isolation is suitable for environments where there is a relatively low chance that two concurrent transactions will modify the same rows and the long-running transactions are primarily read only. It is most suitable for environments with large databases and short transactions that update only a few rows.

Serializable isolation mode provides somewhat more consistency by protecting against phantoms and nonrepeatable reads and can be important where a read/write transaction runs a query more than once.

Unlike other implementations of serializable isolation, which lock blocks for read as well as write, Oracle provides nonblocking queries and the fine granularity of row-level locking, both of which reduce write/write contention. For applications that experience mostly read/write contention, Oracle serializable isolation can provide significantly more throughput than other systems. Therefore, some applications might be suitable for serializable isolation on Oracle but not on other systems.

All queries in an Oracle serializable transaction see the database as of a single point in time, so this isolation level is suitable where multiple consistent queries must be issued in a read/write transaction. A report-writing application that generates summary data and stores it in the database might use serializable mode because it provides the consistency that a READ ONLY transaction provides, but also allows INSERT, UPDATE, and DELETE.

Coding serializable transactions requires extra work by the application developer to check for the "Cannot serialize access" error and to undo and retry the transaction. Similar extra coding is needed in other database management systems to manage deadlocks. For adherence to corporate standards or for applications that are run on multiple database management systems, it may be necessary to design transactions for serializable mode. Transactions that check for serializability failures and retry can be used with Oracle read committed mode, which does not generate serializability errors.

Serializable mode is probably not the best choice in an environment with relatively long transactions that must update the same rows accessed by a high volume of short update transactions. Because a longer running transaction is unlikely to be the first to modify a given row, it will repeatedly need to roll back, wasting work. Note that a conventional read-locking, pessimistic implementation of serializable mode would not be suitable for this environment either, because long-running transactions—even read transactions—would block the progress of short update transactions and vice versa.

Application developers should take into account the cost of rolling back and retrying transactions when using serializable mode. As with read-locking systems, where deadlocks occur frequently, use of serializable mode requires rolling back the work done by terminated transactions and retrying them. In a high contention environment, this activity can use significant resources.

In most environments, a transaction that restarts after receiving the "Cannot serialize access" error is unlikely to encounter a second conflict with another transaction. For this reason, it can help to run those statements most likely to contend with other transactions as early as possible in a serializable transaction. However, there is no guarantee that the transaction will complete successfully, so the application should be coded to limit the number of retries.

Although Oracle serializable mode is compatible with SQL92 and offers many benefits compared with read-locking implementations, it does not provide semantics identical to such systems. Application designers must take into account the fact that reads in Oracle do not block writes as they do in other systems. Transactions that check for database consistency at the application level can require coding techniques such as the use of SELECT FOR UPDATE. This issue should be considered when applications using serializable mode are ported to Oracle from other environments.

Quiesce Database

You can put the system into quiesced state. The system is in quiesced state if there are no active sessions, other than SYS and SYSTEM. An active session is defined as a session that is currently inside a transaction, a query, a fetch or a PL/SQL procedure, or a session that is currently holding any shared resources (for example, enqueues--enqueues are shared memory structures that serialize access to database resources and are associated with a session or transaction). Database administrators are the only users who can proceed when the system is in quiesced state.

Database administrators can perform certain actions in the quiesced state that cannot be safely done when the system is not quiesced. These actions include:

• Actions that might fail if there are concurrent user transactions or queries. For example, changing the schema of a database table will fail if a concurrent transaction is accessing the same table.

• Actions whose intermediate effect could be detrimental to concurrent user transactions or queries. For example, suppose there is a big table T and a PL/SQL package that operates on it. You can split table T into two tables T1 and T2, and change the PL/SQL package to make it refer to the new tables T1 and T2, instead of the old table T.

  When the database is in quiesced state, you can do the following:

  CREATE TABLE T1 AS SELECT ... FROM T; 
  CREATE TABLE T2 AS SELECT ... FROM T; 
  DROP TABLE T; 
  
  

  You can then drop the old PL/SQL package and re-create it.

For systems that must operate continuously, the ability to perform such actions without shutting down the database is critical.

The Database Resource Manager blocks all actions that were initiated by a user other than SYS or SYSTEM while the system is quiesced. Such actions are allowed to proceed when the system goes back to normal (unquiesced) state. Users do not get any additional error messages from the quiesced state.

Modes of Locking

Oracle uses two modes of locking in a multiuser database:

• Exclusive lock mode prevents the associates resource from being shared. This lock mode is obtained to modify data. The first transaction to lock a resource exclusively is the only transaction that can alter the resource until the exclusive lock is released.

• Share lock mode allows the associated resource to be shared, depending on the operations involved. Multiple users reading data can share the data, holding share locks to prevent concurrent access by a writer (who needs an exclusive lock). Several transactions can acquire share locks on the same resource.

Lock Duration

All locks acquired by statements within a transaction are held for the duration of the transaction, preventing destructive interference including dirty reads, lost updates, and destructive DDL operations from concurrent transactions. The changes made by the SQL statements of one transaction become visible only to other transactions that start after the first transaction is committed.

Oracle releases all locks acquired by the statements within a transaction when you either commit or undo the transaction. Oracle also releases locks acquired after a savepoint when rolling back to the savepoint. However, only transactions not waiting for the previously locked resources can acquire locks on the now available resources. Waiting transactions will continue to wait until after the original transaction commits or rolls back completely.

Data Lock Conversion Versus Lock Escalation

A transaction holds exclusive row locks for all rows inserted, updated, or deleted within the transaction. Because row locks are acquired at the highest degree of restrictiveness, no lock conversion is required or performed.

Oracle automatically converts a table lock of lower restrictiveness to one of higher restrictiveness as appropriate. For example, assume that a transaction uses a SELECT statement with the FOR UPDATE clause to lock rows of a table. As a result, it acquires the exclusive row locks and a row share table lock for the table. If the transaction later updates one or more of the locked rows, the row share table lock is automatically converted to a row exclusive table lock.

Lock escalation occurs when numerous locks are held at one level of granularity (for example, rows) and a database raises the locks to a higher level of granularity (for example, table). For example, if a single user locks many rows in a table, some databases automatically escalate the user's row locks to a single table. The number of locks is reduced, but the restrictiveness of what is being locked is increased.

Oracle never escalates locks. Lock escalation greatly increases the likelihood of deadlocks. Imagine the situation where the system is trying to escalate locks on behalf of transaction T1 but cannot because of the locks held by transaction T2. A deadlock is created if transaction T2 also requires lock escalation of the same data before it can proceed.

Deadlock Detection

Oracle automatically detects deadlock situations and resolves them by rolling back one of the statements involved in the deadlock, thereby releasing one set of the conflicting row locks. A corresponding message also is returned to the transaction that undergoes statement-level rollback. The statement rolled back is the one belonging to the transaction that detects the deadlock. Usually, the signalled transaction should be rolled back explicitly, but it can retry the rolled-back statement after waiting. 

Deadlocks most often occur when transactions explicitly override the default locking of Oracle. Because Oracle itself does no lock escalation and does not use read locks for queries, but does use row-level locking (rather than page-level locking), deadlocks occur infrequently in Oracle.

Avoid Deadlocks

Multitable deadlocks can usually be avoided if transactions accessing the same tables lock those tables in the same order, either through implicit or explicit locks. For example, all application developers might follow the rule that when both a master and detail table are updated, the master table is locked first and then the detail table. If such rules are properly designed and then followed in all applications, deadlocks are very unlikely to occur.

When you know you will require a sequence of locks for one transaction, consider acquiring the most exclusive (least compatible) lock first.

Row Locks (TX)

The only DML locks Oracle acquires automatically are row-level locks. There is no limit to the number of row locks held by a statement or transaction, and Oracle does not escalate locks from the row level to a coarser granularity. Row locking provides the finest grain locking possible and so provides the best possible concurrency and throughput.

The combination of multiversion concurrency control and row-level locking means that users contend for data only when accessing the same rows, specifically:

• Readers of data do not wait for writers of the same data rows.

• Writers of data do not wait for readers of the same data rows unless SELECT ... FOR UPDATE is used, which specifically requests a lock for the reader.

• Writers only wait for other writers if they attempt to update the same rows at the same time.

  
  

  Note: Readers of data may have to wait for writers of the same data blocks in some very special cases of pending distributed transactions.

  
  

A transaction acquires an exclusive DML lock for each individual row modified by one of the following statements: INSERT, UPDATE, DELETE, and SELECT with the FOR UPDATE clause.

A modified row is always locked exclusively so that other users cannot modify the row until the transaction holding the lock is committed or rolled back. However, if the transaction dies due to instance failure, block-level recovery makes a row available before the entire transaction is recovered. Row locks are always acquired automatically by Oracle as a result of the statements listed previously.

If a transaction obtains a row lock for a row, the transaction also acquires a table lock for the corresponding table. The table lock prevents conflicting DDL operations that would override data changes in a current transaction.

Table Locks (TM)

A transaction acquires a table lock when a table is modified in the following DML statements: INSERT, UPDATE, DELETE, SELECT with the FOR UPDATE clause, and LOCK TABLE. These DML operations require table locks for two purposes: to reserve DML access to the table on behalf of a transaction and to prevent DDL operations that would conflict with the transaction. Any table lock prevents the acquisition of an exclusive DDL lock on the same table and thereby prevents DDL operations that require such locks. For example, a table cannot be altered or dropped if an uncommitted transaction holds a table lock for it.

A table lock can be held in any of several modes: row share (RS), row exclusive (RX), share (S), share row exclusive (SRX), and exclusive (X). The restrictiveness of a table lock's mode determines the modes in which other table locks on the same table can be obtained and held.

Table 13-3 shows the table lock modes that statements acquire and operations that those locks permit and prohibit.

RS: row share

RX: row exclusive

S: share

SRX: share row exclusive

X: exclusive

*Yes, if no conflicting row locks are held by another transaction. Otherwise, waits occur.

The following sections explain each mode of table lock, from least restrictive to most restrictive. They also describe the actions that cause the transaction to acquire a table lock in that mode and which actions are permitted and prohibited in other transactions by a lock in that mode.

SELECT ... FROM table ... FOR UPDATE OF ... ;  

LOCK TABLE table IN ROW SHARE MODE; 

LOCK TABLE table IN EXCLUSIVE MODE;

INSERT INTO table ... ; 

UPDATE table ... ; 

DELETE FROM table ... ;  

LOCK TABLE table IN ROW EXCLUSIVE MODE; 

LOCK TABLE table IN SHARE MODE;  

LOCK TABLE table IN SHARE EXCLUSIVE MODE; 

LOCK TABLE table IN EXCLUSIVE MODE; 

LOCK TABLE table IN SHARE MODE; 

LOCK TABLE table IN SHARE ROW EXCLUSIVE MODE; 

LOCK TABLE table IN EXCLUSIVE MODE; 

LOCK TABLE table IN ROW EXCLUSIVE MODE; 

LOCK TABLE table IN SHARE ROW EXCLUSIVE MODE; 

LOCK TABLE table IN SHARE MODE; 

LOCK TABLE table IN SHARE ROW EXCLUSIVE MODE; 

LOCK TABLE table IN ROW EXCLUSIVE MODE; 

LOCK TABLE table IN EXCLUSIVE MODE; 

LOCK TABLE table IN EXCLUSIVE MODE; 

DML Locks Automatically Acquired for DML Statements

The previous sections explained the different types of data locks, the modes in which they can be held, when they can be obtained, when they are obtained, and what they prohibit. The following sections summarize how Oracle automatically locks data on behalf of different DML operations.

Table 13-4 summarizes the information in the following sections.

X: exclusive

RX: row exclusive

RS: row share

S: share

SRX: share row exclusive

SELECT 

INSERT ... SELECT ... ; 

UPDATE ... ; 

DELETE ... ; 

SELECT ... FOR UPDATE OF ... ; 

Exclusive DDL Locks

Most DDL operations, except for those listed in the next section, "Share DDL Locks", require exclusive DDL locks for a resource to prevent destructive interference with other DDL operations that might modify or reference the same schema object. For example, a DROP TABLE operation is not allowed to drop a table while an ALTER TABLE operation is adding a column to it, and vice versa.

During the acquisition of an exclusive DDL lock, if another DDL lock is already held on the schema object by another operation, the acquisition waits until the older DDL lock is released and then proceeds.

DDL operations also acquire DML locks (data locks) on the schema object to be modified.

Share DDL Locks

Some DDL operations require share DDL locks for a resource to prevent destructive interference with conflicting DDL operations, but allow data concurrency for similar DDL operations. For example, when a CREATE PROCEDURE statement is run, the containing transaction acquires share DDL locks for all referenced tables. Other transactions can concurrently create procedures that reference the same tables and therefore acquire concurrent share DDL locks on the same tables, but no transaction can acquire an exclusive DDL lock on any referenced table. No transaction can alter or drop a referenced table. As a result, a transaction that holds a share DDL lock is guaranteed that the definition of the referenced schema object will remain constant for the duration of the transaction.

A share DDL lock is acquired on a schema object for DDL statements that include the following statements: AUDIT, NOAUDIT, COMMENT, CREATE [OR REPLACE] VIEW/ PROCEDURE/PACKAGE/PACKAGE BODY/FUNCTION/ TRIGGER, CREATE SYNONYM, and CREATE TABLE (when the CLUSTER parameter is not included).

Breakable Parse Locks

A SQL statement (or PL/SQL program unit) in the shared pool holds a parse lock for each schema object it references. Parse locks are acquired so that the associated shared SQL area can be invalidated if a referenced object is altered or dropped. A parse lock does not disallow any DDL operation and can be broken to allow conflicting DDL operations, hence the name breakable parse lock.

A parse lock is acquired during the parse phase of SQL statement execution and held as long as the shared SQL area for that statement remains in the shared pool.

Duration of DDL Locks

The duration of a DDL lock depends on its type. Exclusive and share DDL locks last for the duration of DDL statement execution and automatic commit. A parse lock persists as long as the associated SQL statement remains in the shared pool.

DDL Locks and Clusters

A DDL operation on a cluster acquires exclusive DDL locks on the cluster and on all tables and materialized views in the cluster. A DDL operation on a table or materialized view in a cluster acquires a share lock on the cluster, in addition to a share or exclusive DDL lock on the table or materialized view. The share DDL lock on the cluster prevents another operation from dropping the cluster while the first operation proceeds.

Latches

Latches are simple, low-level serialization mechanisms to protect shared data structures in the system global area (SGA). For example, latches protect the list of users currently accessing the database and protect the data structures describing the blocks in the buffer cache. A server or background process acquires a latch for a very short time while manipulating or looking at one of these structures. The implementation of latches is operating system dependent, particularly in regard to whether and how long a process will wait for a latch.

Internal Locks

Internal locks are higher-level, more complex mechanisms than latches and serve a variety of purposes.