When a replication source server shuts down and restarts, itsMEMORY
tables become empty. To replicate this effect to replicas, the first time that the source uses a givenMEMORY
table after startup, it logs an event that notifies replicas that the table must be emptied by writing aDELETE
or (from MySQL 8.0.22)TRUNCATE TABLE
statement for that table to the binary log. This generated event is identifiable by a comment in the binary log, and if GTIDs are in use on the server, it has a GTID assigned. The statement is always logged in statement format, even if the binary logging format is set toROW
, and it is written even ifread_only
orsuper_read_only
mode is set on the server. Note that the replica still has outdated data in aMEMORY
table during the interval between the source's restart and its first use of the table. To avoid this interval when a direct query to the replica could return stale data, you can set theinit_file
system variable to name a file containing statements that populate theMEMORY
table on the source at startup.
When a replica server shuts down and restarts, itsMEMORY
tables become empty. This causes the replica to be out of synchrony with the source and may lead to other failures or cause the replica to stop:
Row-format updates and deletes received from the source may fail with
Can't find record in '
.memory_table
'年代tatements such as
INSERT INTO ... SELECT FROM
may insert a different set of rows on the source and replica.memory_table
The replica also writes aDELETE
or (from MySQL 8.0.22)TRUNCATE TABLE
statement to its own binary log, which is passed on to any downstream replicas, causing them to empty their ownMEMORY
tables.
The safe way to restart a replica that is replicatingMEMORY
tables is to first drop or delete all rows from theMEMORY
tables on the source and wait until those changes have replicated to the replica. Then it is safe to restart the replica.
An alternative restart method may apply in some cases. Whenbinlog_format=ROW
, you can prevent the replica from stopping if you setreplica_exec_mode=IDEMPOTENT
(from MySQL 8.0.26) orslave_exec_mode=IDEMPOTENT
(before MySQL 8.0.26) before you start the replica again. This allows the replica to continue to replicate, but itsMEMORY
tables still differ from those on the source. This is acceptable if the application logic is such that the contents ofMEMORY
tables can be safely lost (for example, if theMEMORY
tables are used for caching).replica_exec_mode=IDEMPOTENT
orslave_exec_mode=IDEMPOTENT
applies globally to all tables, so it may hide other replication errors in non-MEMORY
tables.
(The method just described is not applicable in NDB Cluster, wherereplica_exec_mode
orslave_exec_mode
is alwaysIDEMPOTENT
, and cannot be changed.)
的大小MEMORY
tables is limited by the value of themax_heap_table_size
system variable, which is not replicated (see年代ection 17.5.1.39, “Replication and Variables”). A change inmax_heap_table_size
takes effect forMEMORY
tables that are created or updated usingALTER TABLE ... ENGINE = MEMORY
orTRUNCATE TABLE
following the change, or for allMEMORY
tables following a server restart. If you increase the value of this variable on the source without doing so on the replica, it becomes possible for a table on the source to grow larger than its counterpart on the replica, leading to inserts that succeed on the source but fail on the replica withTable is fullerrors. This is a known issue (Bug #48666). In such cases, you must set the global value ofmax_heap_table_size
on the replica as well as on the source, then restart replication. It is also recommended that you restart both the source and replica MySQL servers, to ensure that the new value takes complete (global) effect on each of them.
年代ee年代ection 16.3, “The MEMORY Storage Engine”, for more information aboutMEMORY
tables.