This means we can get into infinite retry loops if our sharding data gets out of sync.

First step is a hard limit, second step is to improve detection of when writes are "making progress" in a cluster.