Sloppy Quorum And Hinted Handoff -

Strong consistency

e.g. Postgres on one node. If there are concurrent writes & concurrent reads, whether all clients see the same value for the same key depends on the Isolation Level, but assuming a strong isolation level they will.
This 👀 the ideal scenario. Why not use it always? It doesn’t scale, and if machine dies, all is lost.

e.g. a master-slave replication scheme on top of MySQL. Only the master processes writes, and one or more replicas process reads in parallel.
Now reads scale, but writes still don’t, and new issues!
Depending on replication lag, clients connected to different replicas see different values for the same key!
If master dies, system is read-only until new master chosen.
🧠: linearizability is still guaranteed, because only one node processes writes.

In leader-based replication, there is one primary node and some replicas responsible for handling a read/write/delete of a given key.

Defining the w, r and n numbers for replication:

w: number of nodes that must acknowledge a write for the success to be sent to the client (client blocked until then).
r: number of nodes that must be asked for the value of a given key; if they return different values, there will be a strategy to pick winner (e.g. Last Write Wins in Cassandra, let application decide in DynamoDB).
n: total nodes that are concerned with a given key (<= total nodes in cluster).

Typically w + r > n.

All n nodes that are concerned with a given key eventually converge to the same value.

👀: If less than w nodes are online (of those n), that key is read-only!!

⚠️ This is either not supported or disabled by default everywhere except Riak & DynamoDB ⚠️

If less than w nodes are reachable (of those n), but there are other nodes in the cluster to fill in, accept the write anyway (sloppy quorum) and forward them to the right nodes later when they are reachable again.
This strategy maximises availability over consistency, typically on cross-datacenter temporary network issues, but what happens to consistency if the nodes that were unreachable to this client were on for everyone else, serving reads & writes concurrently? 😱