My problem is different

[Alon Lubin]

Hi Rick,
Thanks for your reply.

My problem is with the decision to work in row-based logging for "read committed" isolation level while "repeatable read" works in statement based replications.
The problem is not with the amount of rows, but with translation into full rows delete in the binary log.

I don't think read-committed is less safe than repeatable read for this issue.

In addition, the delete statement in the log contains all the columns of the deleted table and not the primary key columns, which leads to even bigger log waste.

We hit a major IO problem on our slaves for simple delete statements.
The delete statement if to be executed on the slave as delete should not take more than few seconds, while replicating the specific deletes takes forever.

Thanks, Alon.