> when using a system "with consensus" there is often no need to actually invoke consensus on the read side of the system, in which case you don't have to pay the throughput and latency penalties.
Doesn't matter if the system is designed for faster reads. There is still coordination, you still pay coordination overhead, including throughput and latency penalties. Without coordination, for example, you can have some part of the database relevant to the client stored directly on the client, not doing remote reads at all.
You're comparing apples to oranges. If read coordination can be avoided (and it often can, even in systems with consensus), whether you stick a cache on the client or not is completely orthogonal to whether the system uses consensus for write operations.
Think about it, if you read from a replica that is partitioned from the rest of the system and there is no coordination, how would the replica or the client know that the value replica returns is too old and therefore breaks strong consistency guarantee?
No, you don't get linearizability of all operations, you might not even read your earlier writes - but the whole point is that you sometimes don't need these guarantees for reads. You get a consistent snapshot read, and that's often good enough. You can get an idea of how recent the snapshot is based on timestamps, but "recent" is hard to define in a distributed system.
> but the whole point is that you sometimes don't need these guarantees for reads.
If that was your point, than sure. If you drop consistency, you can drop coordination too. But typically people expect reads to be consistent in consensus based systems, which requires coordination.
Doesn't matter if the system is designed for faster reads. There is still coordination, you still pay coordination overhead, including throughput and latency penalties. Without coordination, for example, you can have some part of the database relevant to the client stored directly on the client, not doing remote reads at all.