CAP & PACELC

What it is

CAP says that during a network partition, a distributed system must choose between consistency (every read sees the latest write) and availability (every request gets a non-error answer). Networks fail, so partition tolerance is not a choice — the real decision is CP versus AP, made per feature, not per system. PACELC extends it to the 99.9% of the time the network is healthy: if Partition, choose A or C; Else, choose Latency or Consistency. That second half matters more day-to-day, because partitions are rare but the latency-versus-consistency trade is paid on every single write.

In an interview this is the framing device for non-functional requirements. The interviewer wants to hear me map the domain: money movement, seat inventory, unique usernames and permissions get strong consistency; likes, view counts, feeds and profile edits go eventual. Defaulting to availability unless the domain demands otherwise — and saying so — is the expected move.

Say this out loud

I choose per feature, not per system: checkout and inventory are CP, browsing and feeds are AP — and since partitions are rare, the PACELC question of latency versus consistency is the one I'm actually paying for every day.

The picture

the partition is not optional — the network will split. your only real choice is the bottom edge: refuse writes (CP) or serve stale (AP), per feature.

The variants and when to pick each

CP — refuse rather than lie

A CP system stops answering rather than answer wrong. Single-leader databases with synchronous replication, and consensus systems like ZooKeeper and etcd, behave this way: under Raft, the minority side of a partition simply refuses requests. Spanner is the famous refinement — technically CP, but engineered to five nines so the A it sacrifices is almost never visible.

I reach for CP where a stale or conflicting answer is a business incident: double-selling a seat, double-spending a balance, two users claiming one username.

Trade-offCorrectness costs uptime and latency — the minority partition goes dark, and every write pays a coordination round trip, 30–150 ms if it crosses regions.

AP — answer, possibly stale

An AP system keeps serving on both sides of a partition and reconciles afterwards. Cassandra, DynamoDB in its default eventually-consistent mode, Riak and DNS all take this bet: the value of an answer now beats the risk of it being seconds old.

This is the right default for read-heavy consumer features. Nobody files a ticket because a like count was stale for 800 ms, and healthy systems converge in well under a second.

Trade-offYou trade truth for uptime — readers can see old data during the window, and concurrent writes on both sides of a split need a conflict story (last-write-wins or vector clocks) when it heals.

PACELC — the version that matters when nothing is broken

Classifying systems the PACELC way is the fluency signal: DynamoDB and Cassandra are PA/EL — available under partition, latency-first in normal operation. MongoDB is commonly cited PA/EC. Spanner, BigTable and HBase are PC/EC — consistency both times, which is exactly why cross-region writes on them are slow.

The practical reading: even with a perfect network, I still choose between fast writes and consistent reads on every operation. That is the trade I should be defending in the interview, not the once-a-year partition.

Trade-offPC/EC systems pay coordination latency on every write for guarantees most features don't need; PA/EL systems are fast and available but push staleness handling onto the application.

The nuances that separate levels

Three things worth saying explicitly. First, the C in CAP is linearizability, not the C in ACID — conflating them is a common tell. Second, CAP applies per operation on a single object, not to a whole product — one system can be CP for checkout and AP for browsing, and the best answers mix deliberately. Third, "pick two of three" is a misleading slogan: partition tolerance is mandatory, so you pick one of two, and only while partitioned.

Trade-offPrecision here costs nothing and buys credibility — the sloppy version of CAP is so common that interviewers probe it deliberately.

Numbers worth memorising

Cross-region sync coordination	~30–150 ms per write — why PC/EC is slow globally
Same-region consensus round trip	~1–5 ms to a majority
Eventual-consistency convergence	typically <1 s in a healthy system
Spanner TrueTime uncertainty	~1–7 ms, waited out on commit
PACELC to quote	DynamoDB/Cassandra = PA/EL · Spanner/HBase = PC/EC

Where the interviewer pushes

The push is always to make it concrete: "what does your system literally do during a partition — error, stale read, or block?" The strong answer walks a specific component: the minority side of my etcd cluster refuses writes, so config changes stall but nothing lies; my Cassandra tier keeps accepting writes on both sides and repairs on heal, so a follower might serve a feed that is a second old. Naming the user-visible behaviour — not the acronym — is what scores.

Second push: "point at one feature in your design that should be CP and one that should be AP." This tests whether I treat CAP as a per-feature dial. In a booking system: the seat hold is CP — I'd rather show an error than sell a seat twice — while browsing listings is AP, served stale from cache without apology. If I've claimed everything needs strong consistency, I've also silently signed up for coordination latency on every write, and a good interviewer will bill me for it.