Transaction isolation is one of the foundations of database transaction processing. Isolation is one of the four key properties of a transaction (commonly referred as ACID).
The SQL-92 standard defines four levels of transaction isolation: Read Uncommitted, Read Committed, Repeatable Read, and Serializable. See the following table for details:
|Isolation Level||Dirty Write||Dirty Read||Fuzzy Read||Phantom|
|READ UNCOMMITTED||Not Possible||Possible||Possible||Possible|
|READ COMMITTED||Not Possible||Not possible||Possible||Possible|
|REPEATABLE READ||Not Possible||Not possible||Not possible||Possible|
|SERIALIZABLE||Not Possible||Not possible||Not possible||Not possible|
TiDB implements Snapshot Isolation (SI) consistency, which it advertises as
REPEATABLE-READ for compatibility with MySQL. This differs from the ANSI Repeatable Read isolation level and the MySQL Repeatable Read level.
The Repeatable Read isolation level only sees data committed before the transaction begins, and it never sees either uncommitted data or changes committed during transaction execution by concurrent transactions. However, the transaction statement does see the effects of previous updates executed within its own transaction, even though they are not yet committed.
For transactions running on different nodes, the start and commit order depends on the order that the timestamp is obtained from PD.
Transactions of the Repeatable Read isolation level cannot concurrently update a same row. When committing, if the transaction finds that the row has been updated by another transaction after it starts, then the transaction rolls back. For example:
create table t1(id int); insert into t1 values(0); start transaction; | start transaction; select * from t1; | select * from t1; update t1 set id=id+1; | update t1 set id=id+1; -- In pessimistic transactions, the `update` statement executed later waits for the lock until the transaction holding the lock commits or rolls back and releases the row lock. commit; | | commit; -- The transaction commit fails and rolls back. Pessimistic transactions can commit successfully.
The Repeatable Read isolation level in TiDB differs from ANSI Repeatable Read isolation level, though they sharing the same name. According to the standard described in the A Critique of ANSI SQL Isolation Levels paper, TiDB implements the Snapshot Isolation level. This isolation level does not allow strict phantoms (A3) but allows broad phantoms (P3) and write skews. In contrast, the ANSI Repeatable Read isolation level allows phantom reads but does not allow write skews.
The Repeatable Read isolation level in TiDB differs from that in MySQL. The MySQL Repeatable Read isolation level does not check whether the current version is visible when updating, which means it can continue to update even if the row has been updated after the transaction starts. In contrast, if the row has been updated after the transaction starts, the TiDB optimistic transaction is rolled back and retried. Transaction retries in TiDB's optimistic concurrency control might fail, leading to a final failure of the transaction, while in TiDB's pessimistic concurrency control and MySQL, the updating transaction can be successful.
Starting from TiDB v4.0.0-beta, TiDB supports the Read Committed isolation level.
For historical reasons, the Read Committed isolation level of current mainstream databases is essentially the Consistent Read isolation level defined by Oracle. In order to adapt to this situation, the Read Committed isolation level in TiDB pessimistic transactions is also a consistent read behavior in essence.
The MySQL Read Committed isolation level is in line with the Consistent Read features in most cases. There are also exceptions, such as semi-consistent read. This special behavior is not supported in TiDB.