Transactions

TiDB supports distributed transactions using either pessimistic or optimistic transaction mode. Starting from TiDB 3.0.8, TiDB uses the pessimistic transaction mode by default.

This document introduces commonly used transaction-related statements, explicit and implicit transactions, isolation levels, lazy check for constraints, and transaction sizes.

The common variables include autocommit, tidb_disable_txn_auto_retry, tidb_retry_limit, and tidb_txn_mode.

Overview - 图1

Note

The tidb_disable_txn_auto_retry and tidb_retry_limit variables only apply to optimistic transactions, not to pessimistic transactions.

Common statements

Starting a transaction

The statements BEGIN and START TRANSACTION can be used interchangeably to explicitly start a new transaction.

Syntax:

  1. BEGIN;
  1. START TRANSACTION;
  1. START TRANSACTION WITH CONSISTENT SNAPSHOT;
  1. START TRANSACTION WITH CAUSAL CONSISTENCY ONLY;

If the current session is in the process of a transaction when one of these statements is executed, TiDB automatically commits the current transaction before starting a new transaction.

Overview - 图2

Note

Unlike MySQL, TiDB takes a snapshot of the current database after executing the statements above. MySQL’s BEGIN and START TRANSACTION take a snapshot after executing the first SELECT statement (not SELECT FOR UPDATE) that reads data from InnoDB after a transaction is started. START TRANSACTION WITH CONSISTENT SNAPSHOT takes a snapshot during the execution of the statement. As a result, BEGIN, START TRANSACTION, and START TRANSACTION WITH CONSISTENT SNAPSHOT are equivalent to START TRANSACTION WITH CONSISTENT SNAPSHOT in MySQL.

Committing a transaction

The statement COMMIT instructs TiDB to apply all changes made in the current transaction.

Syntax:

  1. COMMIT;

Overview - 图3

Tip

Make sure that your application correctly handles that a COMMIT statement could return an error before enabling optimistic transactions. If you are unsure of how your application handles this, it is recommended to instead use the default of pessimistic transactions.

Rolling back a transaction

The statement ROLLBACK rolls back and cancels all changes in the current transaction.

Syntax:

  1. ROLLBACK;

Transactions are also automatically rolled back if the client connection is aborted or closed.

Autocommit

As required for MySQL compatibility, TiDB will by default autocommit statements immediately following their execution.

For example:

  1. mysql> CREATE TABLE t1 (
  2. id INT NOT NULL PRIMARY KEY auto_increment,
  3. pad1 VARCHAR(100)
  4. );
  5. Query OK, 0 rows affected (0.09 sec)
  6. mysql> SELECT @@autocommit;
  7. +--------------+
  8. | @@autocommit |
  9. +--------------+
  10. | 1 |
  11. +--------------+
  12. 1 row in set (0.00 sec)
  13. mysql> INSERT INTO t1 VALUES (1, 'test');
  14. Query OK, 1 row affected (0.02 sec)
  15. mysql> ROLLBACK;
  16. Query OK, 0 rows affected (0.01 sec)
  17. mysql> SELECT * FROM t1;
  18. +----+------+
  19. | id | pad1 |
  20. +----+------+
  21. | 1 | test |
  22. +----+------+
  23. 1 row in set (0.00 sec)

In the above example, the ROLLBACK statement has no effect. This is because the INSERT statement is executed in autocommit. That is, it was the equivalent of the following single-statement transaction:

  1. START TRANSACTION;
  2. INSERT INTO t1 VALUES (1, 'test');
  3. COMMIT;

Autocommit will not apply if a transaction has been explicitly started. In the following example, the ROLLBACK statement successfully reverts the INSERT statement:

  1. mysql> CREATE TABLE t2 (
  2. id INT NOT NULL PRIMARY KEY auto_increment,
  3. pad1 VARCHAR(100)
  4. );
  5. Query OK, 0 rows affected (0.10 sec)
  6. mysql> SELECT @@autocommit;
  7. +--------------+
  8. | @@autocommit |
  9. +--------------+
  10. | 1 |
  11. +--------------+
  12. 1 row in set (0.00 sec)
  13. mysql> START TRANSACTION;
  14. Query OK, 0 rows affected (0.00 sec)
  15. mysql> INSERT INTO t2 VALUES (1, 'test');
  16. Query OK, 1 row affected (0.02 sec)
  17. mysql> ROLLBACK;
  18. Query OK, 0 rows affected (0.00 sec)
  19. mysql> SELECT * FROM t2;
  20. Empty set (0.00 sec)

The autocommit system variable can be changed on either a global or session basis.

For example:

  1. SET autocommit = 0;
  1. SET GLOBAL autocommit = 0;

Explicit and implicit transaction

Overview - 图4

Note

Some statements are committed implicitly. For example, executing [BEGIN|START TRANSACTION] implicitly commits the last transaction and starts a new transaction. This behavior is required for MySQL compatibility. Refer to implicit commit for more details.

TiDB supports explicit transactions (use [BEGIN|START TRANSACTION] and COMMIT to define the start and end of the transaction) and implicit transactions (SET autocommit = 1).

If you set the value of autocommit to 1 and start a new transaction through the [BEGIN|START TRANSACTION] statement, the autocommit is disabled before COMMIT or ROLLBACK which makes the transaction becomes explicit.

For DDL statements, the transaction is committed automatically and does not support rollback. If you run the DDL statement while the current session is in the process of a transaction, the DDL statement is executed after the current transaction is committed.

Lazy check of constraints

By default, optimistic transactions will not check the primary key or unique constraints when a DML statement is executed. These checks are instead performed on transaction COMMIT.

For example:

  1. CREATE TABLE t1 (id INT NOT NULL PRIMARY KEY);
  2. INSERT INTO t1 VALUES (1);
  3. BEGIN OPTIMISTIC;
  4. INSERT INTO t1 VALUES (1); -- MySQL returns an error; TiDB returns success.
  5. INSERT INTO t1 VALUES (2);
  6. COMMIT; -- It is successfully committed in MySQL; TiDB returns an error and the transaction rolls back.
  7. SELECT * FROM t1; -- MySQL returns 1 2; TiDB returns 1.
  1. mysql> CREATE TABLE t1 (id INT NOT NULL PRIMARY KEY);
  2. Query OK, 0 rows affected (0.10 sec)
  3. mysql> INSERT INTO t1 VALUES (1);
  4. Query OK, 1 row affected (0.02 sec)
  5. mysql> BEGIN OPTIMISTIC;
  6. Query OK, 0 rows affected (0.00 sec)
  7. mysql> INSERT INTO t1 VALUES (1); -- MySQL returns an error; TiDB returns success.
  8. Query OK, 1 row affected (0.00 sec)
  9. mysql> INSERT INTO t1 VALUES (2);
  10. Query OK, 1 row affected (0.00 sec)
  11. mysql> COMMIT; -- It is successfully committed in MySQL; TiDB returns an error and the transaction rolls back.
  12. ERROR 1062 (23000): Duplicate entry '1' for key 't1.PRIMARY'
  13. mysql> SELECT * FROM t1; -- MySQL returns 1 2; TiDB returns 1.
  14. +----+
  15. | id |
  16. +----+
  17. | 1 |
  18. +----+
  19. 1 row in set (0.01 sec)

The lazy check optimization improves performance by batching constraint checks and reducing network communication. The behavior can be disabled by setting tidb_constraint_check_in_place=ON.

Overview - 图5

Note

  • This optimization only applies to optimistic transactions.
  • This optimization does not take effect for INSERT IGNORE and INSERT ON DUPLICATE KEY UPDATE, but only for normal INSERT statements.

Statement rollback

TiDB supports atomic rollback after statement execution failure. If a statement results in an error, the changes it made will not take effect. The transaction will remain open, and additional changes can be made before issuing a COMMIT or ROLLBACK statement.

  1. CREATE TABLE test (id INT NOT NULL PRIMARY KEY);
  2. BEGIN;
  3. INSERT INTO test VALUES (1);
  4. INSERT INTO tset VALUES (2); -- Statement does not take effect because "test" is misspelled as "tset".
  5. INSERT INTO test VALUES (1),(2); -- Entire statement does not take effect because it violates a PRIMARY KEY constraint
  6. INSERT INTO test VALUES (3);
  7. COMMIT;
  8. SELECT * FROM test;
  1. mysql> CREATE TABLE test (id INT NOT NULL PRIMARY KEY);
  2. Query OK, 0 rows affected (0.09 sec)
  3. mysql> BEGIN;
  4. Query OK, 0 rows affected (0.00 sec)
  5. mysql> INSERT INTO test VALUES (1);
  6. Query OK, 1 row affected (0.02 sec)
  7. mysql> INSERT INTO tset VALUES (2); -- Statement does not take effect because "test" is misspelled as "tset".
  8. ERROR 1146 (42S02): Table 'test.tset' doesn't exist
  9. mysql> INSERT INTO test VALUES (1),(2); -- Entire statement does not take effect because it violates a PRIMARY KEY constraint
  10. ERROR 1062 (23000): Duplicate entry '1' for key 'test.PRIMARY'
  11. mysql> INSERT INTO test VALUES (3);
  12. Query OK, 1 row affected (0.00 sec)
  13. mysql> COMMIT;
  14. Query OK, 0 rows affected (0.01 sec)
  15. mysql> SELECT * FROM test;
  16. +----+
  17. | id |
  18. +----+
  19. | 1 |
  20. | 3 |
  21. +----+
  22. 2 rows in set (0.00 sec)

In the above example, the transaction remains open after the failed INSERT statements. The final insert statement is then successful and changes are committed.

Transaction size limit

Due to the limitations of the underlying storage engine, TiDB requires a single row to be no more than 6 MB. All columns of a row are converted to bytes according to their data types and summed up to estimate the size of a single row.

TiDB supports both optimistic and pessimistic transactions, and optimistic transactions are the basis for pessimistic transactions. Because optimistic transactions first cache the changes in private memory, TiDB limits the size of a single transaction.

By default, TiDB sets the total size of a single transaction to no more than 100 MB. You can modify this default value via txn-total-size-limit in the configuration file. The maximum value of txn-total-size-limit is 1 TB. The individual transaction size limit also depends on the size of remaining memory available in the server. This is because when a transaction is executed, the memory usage of the TiDB process is scaled up comparing with the transaction size, up to two to three times or more of the transaction size.

TiDB previously limited the total number of key-value pairs for a single transaction to 300,000. This restriction was removed in TiDB v4.0.

Overview - 图6

Note

Usually, TiDB Binlog is enabled to replicate data to the downstream. In some scenarios, message middleware such as Kafka is used to consume binlogs that are replicated to the downstream.

Taking Kafka as an example, the upper limit of Kafka’s single message processing capability is 1 GB. Therefore, when txn-total-size-limit is set to more than 1 GB, it might happen that the transaction is successfully executed in TiDB, but the downstream Kafka reports an error. To avoid this situation, you need to decide the actual value of txn-total-size-limit according to the limit of the end consumer. For example, if Kafka is used downstream, txn-total-size-limit must not exceed 1 GB.

Causal consistency

Overview - 图7

Note

Transactions with causal consistency take effect only when the async commit and one-phase commit features are enabled. For details of the two features, see tidb_enable_async_commit and tidb_enable_1pc.

TiDB supports enabling causal consistency for transactions. Transactions with causal consistency, when committed, do not need to get timestamp from PD and have lower commit latency. The syntax to enable causal consistency is as follows:

  1. START TRANSACTION WITH CAUSAL CONSISTENCY ONLY;

By default, TiDB guarantees linear consistency. In the case of linear consistency, if transaction 2 is committed after transaction 1 is committed, logically, transaction 2 should occur after transaction 1. Causal consistency is weaker than linear consistency. In the case of causal consistency, the commit order and occurrence order of two transactions can be guaranteed consistent only when the data locked or written by transaction 1 and transaction 2 have an intersection, which means that the two transactions have a causal relationship known to the database. Currently, TiDB does not support passing in external causal relationship.

Two transactions with causal consistency enabled have the following characteristics:

Transactions with potential causal relationship have the consistent logical order and physical commit order

Assume that both transaction 1 and transaction 2 adopt causal consistency and have the following statements executed:

Transaction 1Transaction 2
START TRANSACTION WITH CAUSAL CONSISTENCY ONLYSTART TRANSACTION WITH CAUSAL CONSISTENCY ONLY
x = SELECT v FROM t WHERE id = 1 FOR UPDATE
UPDATE t set v = $(x + 1) WHERE id = 2
COMMIT
UPDATE t SET v = 2 WHERE id = 1
COMMIT

In the example above, transaction 1 locks the id = 1 record and transaction 2 modifies the id = 1 record. Therefore, transaction 1 and transaction 2 have a potential causal relationship. Even with the causal consistency enabled, as long as transaction 2 is committed after transaction 1 is successfully committed, logically, transaction 2 must occur after transaction 1. Therefore, it is impossible that a transaction reads transaction 2’s modification on the id = 1 record without reading transaction 1’s modification on the id = 2 record.

Transactions with no causal relationship do not guarantee consistent logical order and physical commit order

Assume that the initial values of id = 1 and id = 2 are both 0. Assume that both transaction 1 and transaction 2 adopt causal consistency and have the following statements executed:

Transaction 1Transaction 2Transaction 3
START TRANSACTION WITH CAUSAL CONSISTENCY ONLYSTART TRANSACTION WITH CAUSAL CONSISTENCY ONLY
UPDATE t set v = 3 WHERE id = 2
UPDATE t SET v = 2 WHERE id = 1
BEGIN
COMMIT
COMMIT
SELECT v FROM t WHERE id IN (1, 2)

In the example above, transaction 1 does not read the id = 1 record, so transaction 1 and transaction 2 have no causal relationship known to the database. With causal consistency enabled for the transactions, even if transaction 2 is committed after transaction 1 is committed in terms of physical time order, TiDB does not guarantee that transaction 2 logically occurs after transaction 1.

If transaction 3 begins before transaction 1 is committed, and if transaction 3 reads the id = 1 and id = 2 records after transaction 2 is committed, transaction 3 might read the value of id = 1 to be 2 but the value of id = 2 to be 0.

Reads without lock do not create causal relationship

Assume that both transaction 1 and transaction 2 adopt causal consistency and have the following statements executed:

Transaction 1Transaction 2
START TRANSACTION WITH CAUSAL CONSISTENCY ONLYSTART TRANSACTION WITH CAUSAL CONSISTENCY ONLY
UPDATE t SET v = 2 WHERE id = 1
SELECT v FROM t WHERE id = 1
UPDATE t set v = 3 WHERE id = 2
COMMIT
COMMIT

In the example above, reads without lock do not create causal relationship. Transaction 1 and transaction 2 have created write skew. In this case, it would have been unreasonable if the two transactions still had causal relationship. Therefore, the two transactions with causal consistency enabled have no definite logical order.