Master-slave replication is eventually consistent, because clients making reads of the same “key” at the same time to the primary or the replicas can get different values (if a write took place, and hasn’t been replicated yet).

Reading Your Own Writes

Fix:

• Client always reads from the leader when it’s a read of something that they can have written to.
• Client keeps track of the logical timestamp they wrote at, and sends the timestamp with the read request (logical timestamp as in sequence number; real timestamp would depend on trusting clocks). If replica hasn’t arrived to it, wait/redirect/fail.

Monotonic Reads

Strong consistency > Monotonic Reads > Eventual consistency

Fix? Each user (e.g. via userID hash) does reads always to the same replica. Works unless replica goes offline 🤷‍♂️.

Consistent Prefix Reads

Fix?

Writes go to the same machine.