Consistency Guarantees eventual consistency eventually data converges different than transaction isolation isolation - avoid race conditions due to concurrent execution consistency - about coordinating replica state levels strongest - linearizability causality and total ordering commit in a distributed system consensus problem Linearizability one replica illusion (one copy of the data) guarantee read is most recent - recency guarantee read A / write C / read A ok read A / begin write C / read B / end write C / read A not ok read is concurrent with the write linearizability, 1 always follows 1 (no flipping) cas(x, vold, vnew) vs Serializability serializability is about transactions guaranteeing a sequential order linearizability is about recency guarantees on read and write importance locking and leader election lock must be linearizable all nodes must agree on who holds the lock zookeeper - used for distributed locking and leader election constraints and uniqueness guarantees enforcing uniquess (username, filename) similar to acquiring a lock similar to cas operation constraints that bank balance >= 0 two people don’t book the same flight requires a single up-to-date value for account balance or seat occupancy uniquess constraints in DBs are linearizable foreigh key and attribute constraints can be implemented without linearizability cross-channel timing dependencies File storage service is not linearizable, that is two requests went into it, one to store the image, and the other to resize the image.