CODEWITHSUNDEEP
database-replicationdistributed-systemconsensus
pdodwad
pdodwad

Reputation: 11

How does leaderless replication actually work ? Is there really no single co-ordinator / leader node which maintains the replication?

I have been going through Martin Kleppmann's "Designing Data Intensive Applications" as a resource to understand database replication and its relation to consensus algorithm.

Based on my understanding of consensus, there seems to be a single leader / co-ordinator node which drives consensus across a bunch of replica nodes. Raft and Paxos seem to work like this.

So my question is : How does leaderless replication system work ? Is there actually no single leader / co-ordinator node which orchestrates the agreement / consensus among a bunch of nodes ?

Even if read or write requests are sent to bunch of replica nodes, shouldn't there be an implicit co-ordinator / leader who handles the request ?

I did not find much conclusive material to explain how leaderless replication actually achieves consensus among the replica nodes.

Please correct me if my understanding is wrong, or I am missing something.

Upvotes: 1

Views: 881

Answers (1)

AndrewR
AndrewR

Reputation: 1740

My take on the ropic:

So my question is : How does leaderless replication system work ? Is there actually no single leader / co-ordinator node which orchestrates the agreement / consensus among a bunch of nodes ?

I've seen two options:

  • A) a request lands on any node and that node coordinates the operation
  • B) a request must land on a node responsible for the data, so if a wrong node is picked, the request is redirected to a right one

Dynamo being a good example of leaderless system is defined as B from above.

Another interesting question is what is the difference b/n leaderless and multi-leader? My take on this is based on which nodes "own" the data. E.g. I could have a system with two MariaDB databases, allow writes on both and have a two way replication process where I will resolve conflict by picking a record randomly - this would be multileader system.

Even if read or write requests are sent to bunch of replica nodes, shouldn't there be an implicit co-ordinator / leader who handles the request ?

Often the client itself is the coordinator - they send requests to all replicas, read results, and may even do read-repairs.

Upvotes: 0

Related Questions