I like what we've done with SlateDB.. all writes go to object storage, but we cache heavily on disk and in memory. You can use local disk, in which case, failover will cause higher latency as the cache warms. EBS will keep the cache warm even on failover.