Isolation (database systems)

From Wikipedia, the free encyclopedia

In database systems, isolation is a property that defines how/when the changes made by one operation become visible to other concurrent operations. Isolation is one of the ACID (Atomicity, Consistency, Isolation, Durability) properties.

Contents

[edit] Isolation levels

The isolation property is the most often relaxed ACID property in a DBMS (Database Management System). This is because to maintain the highest level of isolation a DBMS must acquire locks on data, which may result in a loss of concurrency, or else implement multiversion concurrency control, which may require additional application logic to function correctly.

Most DBMSs offer a number of transaction isolation levels which control the degree of locking which occurs when selecting data. For many database applications the majority of database transactions can be constructed in such a way as to not require high isolation levels, thus reducing the locking overhead for the system. The programmer must carefully analyse database access code to ensure that any relaxation of isolation does not cause difficult-to-find software bugs. Conversely, at higher isolation levels the possibility of deadlock is increased, which also requires careful analysis and programming techniques to avoid.

The isolation levels defined by the ANSI/ ISO SQL standard are:

[edit] SERIALIZABLE

This isolation level specifies that all transactions occur in a completely isolated fashion; i.e., as if all transactions in the system had executed serially, one after the other. The DBMS may execute two or more transactions at the same time only if the illusion of serial execution can be maintained. At this isolation level, phantom reads cannot occur.

With a lock-based concurrency control DBMS implementation, serializability requires that range locks are acquired when a query uses a ranged WHERE clause. When using non-lock concurrency control, no lock is acquired; however, if the system detects a concurrent transaction in progress which would violate the serializability illusion, it must force that transaction to roll back, and the application will have to restart the transaction.

[edit] REPEATABLE READ

All data records retrieved by a SELECT statement cannot be changed; however, if the SELECT statement contains any ranged WHERE clauses, phantom reads may occur. In this isolation level the transaction acquires read locks on all retrieved data, but does not acquire range locks.

[edit] READ COMMITTED

Data records retrieved by a query are not prevented from modification by some other transaction. Non-repeatable reads may occur, meaning data retrieved in a SELECT statement may be modified by some other transaction when it commits. In this isolation level, read locks are acquired on selected data but they are released immediately whereas write locks are released at the end of the transaction.

[edit] READ UNCOMMITTED

In this isolation level, dirty reads are allowed. One transaction may see uncommitted changes made by some other transaction.

The default isolation level of different DBMSs varies quite widely. Most databases which feature transactions allow the user to set any isolation level. Some DBMSs also require additional syntax when performing a SELECT statement which is to acquire locks.

However, the definitions above have been criticised in the paper A Critique of ANSI SQL Isolation Levels as being ambiguous, and as not accurately reflecting the isolation provided by many databases:

This paper shows a number of weaknesses in the anomaly approach to defining isolation levels. The three ANSI phenomena are ambiguous. Even their broadest interpretations do not exclude anomalous behavior. This leads to some counter-intuitive results. In particular, lock-based isolation levels have different characteristics than their ANSI equivalents. This is disconcerting because commercial database systems typically use locking. Additionally, the ANSI phenomena do not distinguish among several isolation levels popular in commercial systems.

[edit] Example queries

In these examples two transactions take place. In the first transaction, Query 1 is performed, then Query 2 is performed in the second transaction and the transaction committed, followed by Query 1 being performed again in the first transaction.

The queries use the following data table.

users
id name age
1 Joe 20
2 Jill 25

[edit] Repeatable reads (phantom reads)

A phantom read occurs when, in the course of a transaction, two identical queries are executed, and the collection of rows returned by the second query is different from the first. This can occur when range locks are not acquired on performing a SELECT.

Transaction 1 Transaction 2
/* Query 1 */
SELECT * FROM users
WHERE age BETWEEN 10 AND 30;
/* Query 2 */
INSERT INTO users VALUES ( 3, 'Bob', 27 );
COMMIT;
/* Query 1 */
SELECT * FROM users
WHERE age BETWEEN 10 AND 30;

Note that transaction 1 executed the same query twice. If the highest level of isolation were maintained, the same set of rows should be returned both times, and indeed that is what is mandated to occur in a database operating at the SQL SERIALIZABLE isolation level. However, at the lesser isolation levels, a different set of rows may be returned the second time.

In the SERIALIZABLE isolation mode, Query 1 would result in all records with age in the range 10 to 30 being locked, thus Query 2 would block until the first transaction was committed. In REPEATABLE READ mode, the range would not be locked, allowing the record to be inserted and the second execution of Query 1 to return the new row in its results.

[edit] Non-repeatable reads

In a lock-based concurrency control method, non-repeatable reads may occur when read locks are not acquired when performing a SELECT. Under multiversion concurrency control, non-repeatable reads may occur when the requirement that a transaction affected by a commit conflict must roll back is relaxed.

Transaction 1 Transaction 2
/* Query 1 */
SELECT * FROM users WHERE id = 1;
/* Query 2 */
UPDATE users SET age = 21 WHERE id = 1;
COMMIT; /* in multiversion concurrency
   control, or lock-based READ COMMITTED */
/* Query 1 */
SELECT * FROM users WHERE id = 1;
COMMIT; /* lock-based REPEATABLE READ */

In this example, Transaction 2 commits successfully, which means that its changes to the row with id 1 should become visible. However, Transaction 1 has already seen a different value for age in that row. At the SERIALIZABLE and REPEATABLE READ isolation level, the DBMS must return the old value. At READ COMMITTED and READ UNCOMMITTED, the DBMS may return the updated value; this is a non-repeatable read.

There are two basic strategies used to prevent non-repeatable reads. The first is to delay the execution of Transaction 2 until Transaction 1 has committed or rolled back. This method is used when locking is used, and produces the serial schedule T1, T2. A serial schedule does not exhibit non-repeatable reads.

In the other strategy, which is used in multiversion concurrency control, Transaction 2 is permitted to commit first, which provides for better concurrency. However, Transaction 1, which commenced prior to Transaction 2, must continue to operate on a past version of the database — a snapshot of the moment it was started. When Transaction 1 eventually tries to commit, the DBMS looks to see if the result of committing Transaction 1 would be equivalent to the schedule T1, T2. If it is, then Transaction 1 can succeed. If it cannot be seen to be equivalent, however, Transaction 1 must roll back with a serialization failure.

Using a lock-based concurrency control method, at the REPEATABLE READ isolation mode, the row with ID = 1 would be locked, thus blocking Query 2 until the first transaction was committed or rolled back. In READ COMMITTED mode the second time Query 1 was executed the age would have changed.

Under multiversion concurrency control, at the SERIALIZABLE isolation level, both SELECT queries see a snapshot of the database taken at the start of Transaction 1. Therefore, they return the same data. However, if Transaction 1 were then to attempt to UPDATE that row as well, a serialization failure would occur and Transaction 1 would be forced to roll back.

At the READ COMMITTED isolation level, each query sees a snapshot of the database taken at the start of each query. Therefore, they each see different data for the updated row. No serialization failure is possible in this mode (because no promise of serializability is made) and Transaction 1 will not have to be retried.

[edit] READ UNCOMMITTED (dirty reads)

A dirty read occurs when a transaction reads data from a row that has been modified by another transaction, but not yet committed.

Dirty reads work similarly to non-repeatable reads, however the second transaction would not need to be committed for the first query to return a different result. The only thing prevented in the READ UNCOMMITTED mode is that updates will not appear in the results out of order; that is, earlier updates will always appear in a result set before later updates.[verification needed]

[edit] See also