Maybe it's just my suspicious-of-everything-distributed brain, but that diagram seems to assume a single universal time scale without any discussion of the implications.
You successfully nerdsniped me and I'm having a lot of trouble finding discussion of the formal implications of what they call "strong consistency" here, other than reading that they did in fact formally verify it to some extent. The best that I could find is this other HN thread where people claim it is causally consistent in a discussion about a Deep Dive (frustratingly shallow, as it happens): https://news.ycombinator.com/item?id=26968627
I have never heard strong consistency used to describe such a weak guarantee before - i.e. it's marketing bs. Usually strong consistency refers to linearizability (or at the least sequential consistency). The diagram a few pages in to this paper gives a nice overview: https://arxiv.org/abs/1512.00168
Yes I actually read that paper while I was digging around but it didn’t seem to help in this case because Amazon don’t specify whether reads made after a concurrent write is made are guaranteed to return the same value as each other. If they are I think the system would be linearizable, yes? Either way they don’t say linearizable anywhere and they describe it specifically as “read-after-write” so I think it would be wrong to assume linearizability. What’s missing from this model for linearizability? S3 doesn’t have transactions after all.
I'm not convinced that issues requests from the multiple clients for the same key actually matters. My speculation is that they map a key to their backend via some type of (consistent/rendezvous) hash and then ensure that all requests for said key lands on the same process/server* that contains the state for the key.
This means that for a specific key, you end up on 1 specific process. If you can ensure this, you basically get monotonic reads/writes along with RYW and Writes-Follow-Reads. All this maps to causal consistency so it is believable.
* The request could probably be sent to a read-only replica first but it could then forward it to the leader replica handling writes by examining some logical timestamp.
