Eventual/Strong Simulation

[bbengfort]

Implement the eventual combined with strong simulation.

Goal

Determine the behavior in the case of conflicting writes given some small # of users (where a “user” might be an automatic process). Feed writes into different devices.

Input

• degree of "bounciness" of input (# consecutive writes at one device before bouncing to another)
• varying connectivity
• different topologies, i.e. more relaxed devices vs cluster devices
• varying anti-entropy periods
• varying rumor-mongering policies

  Metrics

• overall performance
• overall #forks
• % time that objects are converged across system
• latency observed overall, per device, per type of consistency % time system (the whole thing) "fully" replicated, i.e. visible namespace, all forks, all old version.

  Variant 2: Single user

Is there anything to do here? As long as user moves between devices slower than consensus+anti-entropy, it's all good, right?

[bbengfort]

Note we'll have to implement the following:

1. Raft clusters do anti-entropy with their eventual counterparts
2. Raft clusters can "reject" writes if they are not consistent