What they're testing
Whether you use CAP to JUSTIFY a design, or just recite three letters.
The short answer~30 seconds
CAP says a distributed system can't simultaneously have consistency, availability and partition tolerance. But partitions aren't something you choose — they happen — so the real choice is between C and A WHEN one occurs. PACELC adds the more important half: even when there's no partition (Else), you still choose between latency and consistency. That's the choice you pay for on every request, not once every few years.
The user saves their profile. The primary commits and returns success immediately.
The long answer
The right way to use CAP in an interview isn't declaring "this system is AP". It's showing that DIFFERENT PARTS of the same system choose differently: an account balance needs CP because double-spending is unacceptable, while a post's view counter is happily AP because being off by a few dozen for three seconds is invisible. Saying that shows you treat it as an analytical tool rather than a label.
PACELC's "Else" branch is where real money is spent. Wanting reads to always see the latest write means either reading from the primary (concentrating load, adding latency for distant users) or waiting for a quorum (an extra network round trip). Accepting a few hundred milliseconds of staleness lets you read the nearest replica. For a service with users on several continents that's 20ms versus 200ms — large enough to move business metrics.
One detail that prevents misusing CAP: "consistency" there means linearizability, considerably stricter than the everyday sense. Plenty of systems calling themselves CP actually offer sequential or causal consistency. Knowing there's a SPECTRUM rather than two options marks someone who read past the headline.
What they'll ask next
?What does the quorum rule R + W > N mean?
With N replicas, if reads plus writes exceed N the two sets must overlap, so a read always touches at least one replica holding the newest write. N=3, W=2, R=2 is the common setting: survives one dead node and still reads correctly. Setting W=1 for fast writes forces R=3 to compensate, and then one dead node makes reads impossible.
These lose points
- Saying "we'll pick CA". No distributed system is CA — dropping P means running on one machine.
- Applying one label to a whole system. Different components choose differently, and pointing that out is the good answer.