Multi-Datacenter Replication Reference: Comparison

This document is a systematic comparison of Version 2 and Version 3 of Riak Enterprise’s Multi-Datacenter Replication capabilities.

Important note on mixing versions

If you are installing Riak Enterprise anew, you should use version 3 replication. Under no circumstances should you mix version 2 and version 3 replication. This comparison is meant only to list improvements introduced in version 3.

Version 2

  • Version 2 replication relies upon the twin concepts of listeners and sites. Listeners are the sources of replication data, while sites are the destination of replication data. Sites and listeners are manually configured on each node in a cluster. This can be a burden to the administrator as clusters become larger.
  • A single connection tied to the cluster leader manages all replication communications. This can cause performance problems on the leader and is a bottleneck for realtime and fullsync replication data.
  • Connections are established from site to listener. This can be confusing for firewall administrators.
  • The realtime replication queue will be lost if the replication connection breaks, even if it’s re-established. Reconciling data in this situation would require manual intervention using either of the following:
    • a fullsync
    • another Riak write to the key/value on the listener, thus re-queueing the object
  • Riak CS MDC proxy_get connections can only request data from a single leader node

When to use version 2 replication

  • If you are running clusters below version 1.3.0 of Riak Enterprise, version 2 replication is the only method of replication available.
  • In the Riak 1.3 series, version 3 replication was provided as a technology preview and did not have feature parity with version 2. This was provided in the Riak 1.4 series.

Version 3

  • Version 3 replication uses the twin concepts of sources and sinks. A source is considered the primary provider of replication data, whereas a sink is the destination of replication data.
  • Establishing replication connections between clusters has been greatly simplified. A single riak-repl connect command needs to be issued from a source cluster to a sink cluster. IP and port information of all nodes that can participate in replication on both source and sink clusters are exchanged by the replication cluster manager. The replication cluster manager also tracks nodes joining and leaving the cluster dynamically.
  • If the source has M nodes, and the sink has N nodes, there will be M realtime connections. Connections aren’t tied to a leader node as they are with version 2 replication.
  • Communications for realtime, fullsync, and proxy_get operations are multiplexed over the same connection for each node participating in replication. This reduces the amount of firewall configuration on both sources and sinks.
  • A fullsync coordinator runs on a leader of the source cluster. The coordinator assigns work across nodes in the sources cluster in an optimized fashion.
  • Realtime replication establishes a bounded queue on each source node that is shared between all sinks. This queue requires consumers to acknowledge objects when they have been replicated. Dropped TCP connections won’t drop objects from the queue.
  • If a node in the source cluster is shut down via the command line, a realtime replication queue is migrated to other running nodes in the source cluster.
  • Network statistics are kept per socket.
  • Fullsyncs between clusters can be tuned to control the maximum number of workers that will run on a source node, a sink node, and across the entire source cluster. This allows for limiting impact on the cluster and dialing in fullsync performance.
  • Version 3 is able to take advantage of Active Anti-Entropy (AAE) technology, which can greatly improve fullsync performance.
  • Riak CS MDC proxy_get connections will be distributed across the source cluster (as CS blocks are requested from the sink cluster in this scenario).