At the hardware level, one thing that has changed over the years is the exponential growth in network speeds and feeds – within a datacenter and across datacenters.

Layer-3 IP Clos networks are pretty much a standard within a datacenter, making multiple redundant paths available. Across datacenters, optical circuit switched networks with lots of redundant paths are very common.

Modern network fabrics also have end-to-end mechanisms for bandwidth engineering based on per message QoS/priority set by applications as well as very accurate time-keeping.

Observability mechanisms and operational practices have also tremendously improved. A single node becoming unreachable from all other nodes is quite possible. But a single node being reachable to some subset of other nodes in the cluster and not others is nearly impossible. Only way a network partition can happen is due to misconfiguration. Purely layer-3 IP clos networks with BGP advertised multiple ECMP routes are a lot simpler than layer-2 networks of the past. With modern switch operating systems, modern devops practices, Infrastructure-as-Code etc. misconfiguration is nearly impossible.

At a higher layer, disaggregated storage and compute architectures are standard for any distributed database system. In this model, storage fabric is quite simple and much more distributed and redundant. Today, it is a lot easier to guarantee a consistent and highly available transactional write to a quorum of simple storage nodes spread across multiple datacenters in a region. On top of that, building higher-level isolation guaranteeing database transactions in a tier of ephemeral compute nodes is relatively much simpler.

This kind of system is extremely resilient to failures at all levels – node failures are detected and lost processing capacity recovered in seconds; failed transactions are retried in milliseconds; loss of redundancy due to failed storage nodes are rebuilt in seconds; loss of network capacity due to failed circuits or intermediate switches are routed around in microseconds.

So, in today's context, we build databases that provide flexible transaction isolation guarantees that is selected appropriately on a per transaction basis by the application. Application also specifies the latency budget it has available per transaction. With all these improvements, thankfully, application developers don't have worry about CAP theorem like they had to in the heydays of NoSQL databases in 2012.