Reads That Ignore Distance

Kleppmann's delay-sensitivity table gives the lowest operation latency any algorithm can achieve at a given consistency level, as a function of network delay $d$ between replicas:

consistency level	write latency	read latency
linearizability	$O(d)$	$O(d)$
sequential consistency	$O(d)$	$O(1)$
causal consistency	$O(1)$	$O(1)$

You are designing a global feature-flag service. Flags are read on every request in every region and written a few times a day. Reads must stay fast no matter how bad the inter-region link gets, and every region must still apply the flag writes in one agreed order. What is the strongest level in the table that meets both requirements, and what does it cost?

None of them. The lower bound $|r| + |w| \ge d$ applies to any total order, so demanding one agreed order forces $O(d)$ on both paths.

Sequential consistency. Reads can be served locally at $O(1)$ , and the price is paid entirely on the write path, which becomes $O(d)$ .

Linearizability. A single agreed order is exactly what linearizability provides, and read replicas placed in each region bring read latency back to $O(1)$ .

Causal consistency. It is the only row with $O(1)$ reads, and a total order emerges anyway once all the causal dependencies have been delivered.