4 ms·
There's an interesting comment on the blog post: > Does handling all range's consensus traffic in bulk would be the most effective way to handle this? Inste
by ccleve 11y ago
There's an interesting comment on the blog post:
> Does handling all range's consensus traffic in bulk
would be the most effective way to handle this?
Instead could all the ranges from a node be represented
as a single state machine to be handled by plain etcd
raft implementation, like the approach taken in spanner paper?
Do you have any comment on this? I've been trying to wrap my head around the implications. What's the downside?
- bdarnell 11y agoI replied on the blog post. The downside is basically that if you have one giant consensus group, then to add a new replica you have to replicate that entire consensus group. This hurts load balancing and recovery times, since you can't scatter a dead node's ranges across all the other nodes in the cluster.