17.4.1.20 Replication and max_allowed_packet

max_allowed_packet sets an upper limit on the size of any single message between the MySQL server and clients, including replication slaves. If you are replicating large column values (such as might be found in TEXT or BLOB columns) and max_allowed_packet is too small on the master, the master fails with an error, and the slave shuts down the I/O thread. If max_allowed_packet is too small on the slave, this also causes the slave to stop the I/O thread.

Row-based replication currently sends all columns and column values for updated rows from the master to the slave, including values of columns that were not actually changed by the update. This means that, when you are replicating large column values using row-based replication, you must take care to set max_allowed_packet large enough to accommodate the largest row in any table to be replicated, even if you are replicating updates only, or you are inserting only relatively small values.

On a multi-threaded slave (with slave_parallel_workers > 0), ensure that the slave_pending_jobs_size_max system variable is set to a value equal to or greater than the setting for the max_allowed_packet system variable on the master. The default setting for slave_pending_jobs_size_max, 128M, is twice the default setting for max_allowed_packet, which is 64M. max_allowed_packet limits the packet size that the master will send, but the addition of an event header can produce a binary log event exceeding this size. Also, in row-based replication, a single event can be significantly larger than the max_allowed_packet size, because the value of max_allowed_packet only limits each column of the table.

The replication slave actually accepts packets up to the limit set by its slave_max_allowed_packet setting, which defaults to the maximum setting of 1GB, to prevent a replication failure due to a large packet. However, the value of slave_pending_jobs_size_max controls the memory that is made available on the slave to hold incoming packets. The specified memory is shared among all the slave worker queues.

The value of slave_pending_jobs_size_max is a soft limit, and if an unusually large event (consisting of one or multiple packets) exceeds this size, the transaction is held until all the slave workers have empty queues, and then processed. All subsequent transactions are held until the large transaction has been completed. So although unusual events larger than slave_pending_jobs_size_max can be processed, the delay to clear the queues of all the slave workers and the wait to queue subsequent transactions can cause lag on the replication slave and decreased concurrency of the slave workers. slave_pending_jobs_size_max should therefore be set high enough to accommodate most expected event sizes.