The way that transactions are implemented using a "transaction" message does not look right to me.

Requiring the transaction message to contain all the operations of the transaction does not scale (transactions can be very big!) and does not seem necessary.

A better way to implement this would be to have a "transaction_start" message which the user sends to start a transaction, then the user sends operations (insert, delete etc) as normal, and then a "transaction_commit" or "transaction_rollback" message to end the transaction.