Replication

~2 mins read

Keeping the same data on multiple nodes

Why distribute data?

• fault tolerance
• lower latency, eg. multiple AZ, closer to users
• scale

A simple concept yet requires a lot of thought to implement well

Sync or Async?

1. sync -> wait for a write to be replicated on all nodes
2. async -> write to the leader and async write followers
3. semi-sync -> write to the leader and some of the followers and async write others

What to replicate?

1. SQL statements

   indeterministic, eg. now()

2. WAL (write ahead log)

   depends on internal bytes layout of a specific db version, so an upgrade requires downtime

3. Row based

   higher level, easier to parse for external systems, nodes might run different versions of db

4. Trigger based

   more flexible, custom application code, more prone to bugs, greater overhead

How to replicate?

Single Leader

Leaderless Replication

Multi-leader

Makes more sense in a multiple datacenter setup

Offline clients and collaborative editing are also similar to a multi leader setup

See also

Partitioning