[rshrin]

For the Eventually Consistent counter, the low millisecond requirement is for reads and writes, not for the convergence of counts. For this category, the convergence is in the order of seconds (user-personalization, feature-gating fall in this category). For the "Best-effort" category, there are some use cases that run experiments in a single-region and need access to current counts at low latencies (they basically add increments and read the value back in the same call i.e. AddAndGet), but are willing to sacrifice "some degree" of accuracy for it. See the table in the 2nd section. There are multiple dimensions in terms of Read/Write Latency, Staleness, Global reads/writes etc. mapped to the two kinds of use cases. Maybe you are conflating a few things. Finally, there is the experimental type of Accurate counters that can get the current count with high degree of accuracy (but the latency there depends on a few things as explained in the article). The last type is more like what can be done using this approach, no current use case for it.