Mesh Topology

In a star topology, if the central primary dies, replication stops for everyone. In a ring, if one node dies, downstream nodes lose data. In hub-and-spoke, the hub is a bottleneck and single point of failure.

What if there was no single point of failure at all? What if every node could directly replicate to every other node?

That's a mesh topology. In a full mesh with N nodes, every node has a direct replication link to every other node. If any single node fails, the remaining N-1 nodes still have direct connections to each other. No data flow is disrupted. No rerouting needed.

This is how CockroachDB, YugabyteDB, and TiDB work internally, every node can communicate directly with every other node. It's also how some enterprise database clusters operate, particularly in financial services where every millisecond of failover time has a cost.

The trade-off is connection count. A full mesh with 5 nodes needs 10 connections (N*(N-1)/2). With 10 nodes, that's 45 connections. With 50 nodes, it's 1,225. The number of connections grows quadratically, which limits practical mesh size.

Full mesh. Every node connects to every other node. Maximum redundancy, maximum connection overhead. Practical for small clusters (3-10 nodes).

Partial mesh. Each node connects to a subset of other nodes. Reduces connection count while maintaining most of the redundancy. If Node A can't reach Node C directly, it can route through Node B. Cassandra's gossip protocol creates a partial mesh where each node communicates with a random subset of peers.

How mesh replication works:

In a full mesh with multi-master replication:

1. Node A accepts a write
2. Node A sends the change directly to Nodes B, C, D, and E
3. Each node applies the change
4. Nodes B through E do NOT forward the change to each other (they each received it directly from A)

In a partial mesh:

1. Node A accepts a write
2. Node A sends it to its direct peers (say, B and C)
3. Node B forwards it to its peers that haven't seen it (say, D)
4. Node C forwards it to its peers that haven't seen it (say, E)
5. A gossip protocol ensures all nodes eventually receive every change