5.1.3 Determination of Safe and Unsafe Statements in Binary Logging

The “safeness” of a statement in MySQL replication refers to whether the statement and its effects can be replicated correctly using statement-based format. If this is true of the statement, we refer to the statement as safe; otherwise, we refer to it as unsafe.

In general, a statement is safe if it deterministic, and unsafe if it is not. However, certain nondeterministic functions are not considered unsafe (see Nondeterministic functions not considered unsafe, later in this section). In addition, statements using results from floating-point math functions—which are hardware-dependent—are always considered unsafe (see Section 4.1.12, “Replication and Floating-Point Values”).

Handling of safe and unsafe statements.  A statement is treated differently depending on whether the statement is considered safe, and with respect to the binary logging format (that is, the current value of binlog_format).

• When using row-based logging, no distinction is made in the treatment of safe and unsafe statements.

• When using mixed-format logging, statements flagged as unsafe are logged using the row-based format; statements regarded as safe are logged using the statement-based format.

• When using statement-based logging, statements flagged as being unsafe generate a warning to this effect. Safe statements are logged normally.

Each statement flagged as unsafe generates a warning. If a large number of such statements were executed on the source, this could lead to excessively large error log files. To prevent this, MySQL has a warning suppression mechanism. Whenever the 50 most recent ER_BINLOG_UNSAFE_STATEMENT warnings have been generated more than 50 times in any 50-second period, warning suppression is enabled. When activated, this causes such warnings not to be written to the error log; instead, for each 50 warnings of this type, a note The last warning was repeated N times in last S seconds is written to the error log. This continues as long as the 50 most recent such warnings were issued in 50 seconds or less; once the rate has decreased below this threshold, the warnings are once again logged normally. Warning suppression has no effect on how the safety of statements for statement-based logging is determined, nor on how warnings are sent to the client. MySQL clients still receive one warning for each such statement.

For more information, see Section 5.1, “Replication Formats”.

Statements considered unsafe.  Statements with the following characteristics are considered unsafe:

• Statements containing system functions that may return a different value on the replica.  These functions include FOUND_ROWS(), GET_LOCK(), IS_FREE_LOCK(), IS_USED_LOCK(), LOAD_FILE(), MASTER_POS_WAIT(), RAND(), RELEASE_LOCK(), ROW_COUNT(), SESSION_USER(), SLEEP(), SOURCE_POS_WAIT(), SYSDATE(), SYSTEM_USER(), USER(), UUID(), and UUID_SHORT().

  Nondeterministic functions not considered unsafe.  Although these functions are not deterministic, they are treated as safe for purposes of logging and replication: CONNECTION_ID(), CURDATE(), CURRENT_DATE(), CURRENT_TIME(), CURRENT_TIMESTAMP(), CURTIME(), LAST_INSERT_ID(), LOCALTIME(), LOCALTIMESTAMP(), NOW(), UNIX_TIMESTAMP(), UTC_DATE(), UTC_TIME(), and UTC_TIMESTAMP().

  For more information, see Section 4.1.14, “Replication and System Functions”.

• References to system variables.  Most system variables are not replicated correctly using the statement-based format. See Section 4.1.39, “Replication and Variables”. For exceptions, see Mixed Binary Logging Format.

• Loadable Functions.  Since we have no control over what a loadable function does, we must assume that it is executing unsafe statements.

• Fulltext plugin.  This plugin may behave differently on different MySQL servers; therefore, statements depending on it could have different results. For this reason, all statements relying on the fulltext plugin are treated as unsafe in MySQL.

• Trigger or stored program updates a table having an AUTO_INCREMENT column.  This is unsafe because the order in which the rows are updated may differ on the source and the replica.

  In addition, an INSERT into a table that has a composite primary key containing an AUTO_INCREMENT column that is not the first column of this composite key is unsafe.

  For more information, see Section 4.1.1, “Replication and AUTO_INCREMENT”.

• INSERT ... ON DUPLICATE KEY UPDATE statements on tables with multiple primary or unique keys.  When executed against a table that contains more than one primary or unique key, this statement is considered unsafe, being sensitive to the order in which the storage engine checks the keys, which is not deterministic, and on which the choice of rows updated by the MySQL Server depends.

  An INSERT ... ON DUPLICATE KEY UPDATE statement against a table having more than one unique or primary key is marked as unsafe for statement-based replication. (Bug #11765650, Bug #58637)

• Updates using LIMIT.  The order in which rows are retrieved is not specified, and is therefore considered unsafe. See Section 4.1.18, “Replication and LIMIT”.

• Accesses or references log tables.  The contents of the system log table may differ between source and replica.

• Nontransactional operations after transactional operations.  Within a transaction, allowing any nontransactional reads or writes to execute after any transactional reads or writes is considered unsafe.

  For more information, see Section 4.1.35, “Replication and Transactions”.

• Accesses or references self-logging tables.  All reads and writes to self-logging tables are considered unsafe. Within a transaction, any statement following a read or write to self-logging tables is also considered unsafe.

• LOAD DATA statements.  LOAD DATA is treated as unsafe and when binlog_format=MIXED the statement is logged in row-based format. When binlog_format=STATEMENT LOAD DATA does not generate a warning, unlike other unsafe statements.

• XA transactions.  If two XA transactions committed in parallel on the source are being prepared on the replica in the inverse order, locking dependencies can occur with statement-based replication that cannot be safely resolved, and it is possible for replication to fail with deadlock on the replica. When binlog_format=STATEMENT is set, DML statements inside XA transactions are flagged as being unsafe and generate a warning. When binlog_format=MIXED or binlog_format=ROW is set, DML statements inside XA transactions are logged using row-based replication, and the potential issue is not present.

• DEFAULT clause that refers to a nondeterministic function.  If an expression default value refers to a nondeterministic function, any statement that causes the expression to be evaluated is unsafe for statement-based replication. This includes statements such as INSERT, UPDATE, and ALTER TABLE. Unlike most other unsafe statements, this category of statement cannot be replicated safely in row-based format. When binlog_format is set to STATEMENT, the statement is logged and executed but a warning message is written to the error log. When binlog_format is set to MIXED or ROW, the statement is not executed and an error message is written to the error log. For more information on the handling of explicit defaults, see Explicit Default Handling as of MySQL 8.0.13.

For additional information, see Section 4.1, “Replication Features and Issues”.