Aspen: Hybrid Consensus via Raft and P2P
DistributedSystemsComments
I wonder how the transition between Raft and Iroh is handled during a network partition. If the gradient shifts too abruptly, it might lead to unexpected state divergence before the eventual convergence kicks in.
this is just a wrapper around iroh's content-addressable storage.
We saw a similar attempt to blend consensus models in early distributed ledger experiments. Most ended up defaulting back to a single model once the edge cases of the hybrid state became a nightmare to debug.
The use of Rust is a strong choice for the Raft implementation because of its memory safety. This should significantly reduce tail latency during leader elections compared to GC-heavy languages.
Reducing tail latency sounds fine for a lab, but what happens when this hits a shaky connection in a rural office? Does the gradient approach actually keep the app usable, or does it just freeze when the local Raft cluster loses its quorum?
It mirrors the issues found in early edge computing deployments. In those cases, eventual convergence often meant several minutes of stale data that broke the user experience.
I disagree that this is a simple case of stale data. Because Aspen uses federated cluster replication, the local Raft groups maintain linearizability for local operations, which prevents the global stale-state issues seen in pure gossip protocols.