You Don't Get to Keep All Three
Published
"CAP theorem: consistency, availability, partition tolerance, pick two" is one of the most recited lines in system design prep, and it's also slightly wrong. Most developers have heard of the "iron triangle" of delivery; the relationship between cost, quality and time but CAP describes something different. It makes partition tolerance sound like a feature you can choose to skip, in the same way you might choose to skip caching but that isn't the case. Networks partition whether you want them to or not, and the real question the theorem is asking is much narrower than "pick two of three."
What the three letters actually mean
Consistency - every read gets the most recent write, or an error. There's no such thing as a node quietly serving you data from five minutes ago.
Availability - every request gets a response. Not necessarily the latest data, just a response, rather than the system refusing to answer.
Partition tolerance - the system keeps working when some nodes can't talk to others. A network cable gets cut, a region drops offline, a load balancer misroutes traffic for thirty seconds - the system doesn't get a vote on whether this happens.
The part that gets glossed over
Because partitions are a fact of operating anything distributed, not a design choice, "pick two of three" is really "partition tolerance is assumed, so pick one of two." When a partition happens, you have exactly two options: keep answering and risk giving out stale or conflicting data (favouring availability), or stop answering until you can be sure the data is correct (favouring consistency). There is no door marked "neither" and no door marked "both" - a partition forces the choice whether you've planned for it or not. It's possibly worth mentioning that some people
| Approach | During a partition | Example systems | Good fit for |
|---|---|---|---|
| CP (consistency over availability) | Refuses to answer, or errors, rather than risk an incorrect answer | ZooKeeper, etcd, most relational databases with strict quorum | Config stores, leader election, anything where a wrong answer is worse than no answer - bank balances, inventory counts |
| AP (availability over consistency) | Keeps answering with whatever data it has, and reconciles once the partition heals | Cassandra, DynamoDB, DNS | Feeds, view/like counts, shopping carts - anywhere a slightly stale answer beats no answer |
If there were a valid CA state, it's probably a single node machine that's entirely self contained, but given that the definition is about distributed systems, it's not relevant in this conversation.
The interview signal
Reciting "pick two of three" tells the interviewer you've memorised the acronym. Saying "well, partition tolerance isn't really optional, so the actual decision is what happens to this specific system when a partition hits - do we serve stale data or refuse to serve" tells them you understand what's being traded, and that you can apply it to the system in front of you rather than to the theorem in the abstract. It's the same signal as knowing p99 matters more than the average, or knowing that durability and availability aren't the same axis - the interview isn't testing recall, it's testing whether the concept means anything to you under pressure.
One layer deeper
CAP only describes behaviour during a partition. Most of the time there isn't one, and yet systems still make consistency/latency trade-offs on every single request - that's a related but separate trade-off, usually called PACELC, and it's a good thing to bring up if an interview goes deep enough to want it. But CAP itself is narrower than it's given credit for: it's not a general theory of distributed systems trade-offs, it's specifically about what you do the moment the network breaks.
So the next time the question is "how does this system handle a network partition," resist reaching for the acronym. Say what actually happens to a real request when a real node goes unreachable - and whether the answer it gets back is stale, or absent.
