updates on master is concurrent and while user doing one big update on table A, other user can do updates on table B for example.

when replicating, the update is serialized,
For slave it read the statement from binlog in the form of

update on tableA;
update on tableB;
….

update on tableB will not run until slave finish big update on tableA.

This seems to be trivial on small database, but for database driven website like youtube and other web2.0 site, this will cause serious problems. That’s why they will go horizontal partition route.

safari, deadlocks are a fact of life in transactional systems; you can sometimes avoid them by making the different processes read the same index in the same direction, but it’s usually pretty hard to do. Better just build the application to detect and retry the work.