V3 Multi-Datacenter Replication Reference: Cascading Realtime Writes

Introduction

Riak includes a feature that cascades realtime writes across multiple clusters.

Cascading Realtime Writes is enabled by default on new clusters running Riak. It will need to be manually enabled on existing clusters.

Cascading realtime requires the {riak_repl, rtq_meta} capability to function.

Note on cascading tracking

Cascading tracking is a simple list of where an object has been written. This works well for most common configurations. Larger installations, however, may have writes cascade to clusters to which other clusters have already written.

+---+     +---+     +---+
| A | <-> | B | <-> | C |
+---+     +---+     +---+
  ^                   ^
  |                   |
  V                   V
+---+     +---+     +---+
| F | <-> | E | <-> | D |
+---+     +---+     +---+

In the diagram above, a write at cluster A will begin two cascades. One goes to B, C, D, E, and finally F; the other goes to F, E, D, C, and finally B. Each cascade will loop around to A again, sending a replication request even if the same request has already occurred from the opposite direction, creating 3 extra write requests.

This can be mitigated by disabling cascading in a cluster. If cascading were disabled on cluster D, a write at A would begin two cascades. One would go through B, C, and D, the other through F, E, and D. This reduces the number of extraneous write requests to 1.

A different topology can also prevent extra write requests:

+---+                     +---+
| A |                     | E |
+---+                     +---+
 ^  ^                     ^  ^
 |   \  +---+     +---+  /   |
 |    > | C | <-> | D | <    |
 |   /  +---+     +---+  \   |
 V  V                     V  V
+---+                     +---+
| B |                     | F |
+---+                     +---+

A write at A will cascade to C and B. B will not cascade to C because A will have already added C to the list of clusters where the write has occurred. C will then cascade to D. D then cascades to E and F. E and F see that the other was sent a write request (by D), and so they do not cascade.

Usage

Riak Cascading Writes can be enabled and disabled using the riak-repl command. Please see the Version 3 Operations guide for more information.

To show current the settings:

riak-repl realtime cascades

To enable cascading:

riak-repl realtime cascades always

To disable cascading:

riak-repl realtime cascades never