Two out
of three.

A system that works.

Hold something small in your head: a ticket shop with five tickets left, and three datacenters — Frankfurt, Virginia, Singapore — that must agree on how many remain. Every replica keeps a full copy of the count. Every customer asks the nearest one.

Buy a ticket below and watch it travel: the sale is accepted at one site, then copied to the others. Every replica shows the same number. Every customer gets an answer.

With the cables intact, updates can propagate and replicas eventually agree. The blinking lamp marks an update still in flight; this figure has not promised linearizable reads during that interval. Every node is answering requests here. The stronger consistency promise, and what it costs under a split, comes next.

Remember this moment. It is the part of the “two out of three” posters that everyone forgets: on a good day, you keep everything. The theorem has nothing to say about good days.

The lines are not always there.

That tidy figure makes one quiet assumption: that the cables hold. They don’t. Backhoes find fiber-optic cables with remarkable enthusiasm. Switches install firmware at 3 a.m. and never come back. Google engineers have reported sharks biting undersea fiber. And even with every cable intact, a network can delay a packet so long that slow and dead become indistinguishable — in an asynchronous network, no timeout can tell them apart.

In a network where messages can be lost or delayed forever, no system can promise both linearizable consistency and perfect availability.THE CAP THEOREM — AFTER GILBERT & LYNCH, SIGACT NEWS, 2002

This is why partition tolerance is not a feature you build. It is the world you live in. You don’t get to declare it away:

Since we are about to trade these away, let us be precise about what they are:

Consistency

Linearizability: every read reflects the most recent write — as if there were exactly one copy, touched in one agreed order.

Availability

Every request to a living node returns a real, non-error answer. No refusals, no hanging timeouts.

Partition tolerance

The system survives messages being lost or delayed without limit between its nodes. On any real network, this one is not negotiable.

Now it’s your network.

Click a cable to cut it — and learn the first real lesson: with three sites and three links, one cut changes almost nothing. Frankfurt still reaches Singapore through Virginia. Redundancy works; nobody even notices. To truly split a system, you must isolate it — cut a second line.

When you do, the theorem stops being a diagram and becomes a decision made under pressure about real customers. A sale arrives on the island side, which holds no quorum. Refuse it — or serve it and let the histories drift? Flip the policy switch and run the same partition both ways.

MODEL NOTES — each replica displays a local count. AP merges per-origin sales after a heal and can oversell. CP synchronizes the majority’s check-and-decrement before acknowledging a sale; the packet animation is a teaching aid, not a real consensus protocol. Minority writes are refused, and this toy does not model strongly consistent reads.

What “eventually” actually does.

Eventual consistency sounds gentle — as if the data were merely late, like a train. Watch what really happens when you repair the cable: the islands — a state engineers call split-brain — compare their histories and merge them, sale by sale, counting every confirmed order. If both sides sold the same scarce ticket while apart, the merge does not smooth it over. It reveals it: the shop is oversold, and some specific customer will meet the CAP theorem in person — at the door, without a ticket.

A CP system takes the opposite bargain. While split, it refuses every sale it cannot vouch for. Nobody is oversold; somebody is simply told no. Neither choice is kind. The theorem doesn’t prefer either — it only removes the third option, the one where every customer goes home happy.

Without a partition, CAP itself does not force a choice. Replication still has latency and failure costs, and a split makes the consistency and availability conflict unavoidable.THE SHAPE OF BREWER’S RETROSPECTIVE, IEEE COMPUTER, 2012

You pay even on a good day.

Here is what the triangle hides: even with every cable intact, consistency and availability are not simultaneously free. If every replica must agree the instant a write is acknowledged, the write must first reach them — at the speed of light, through a trench, across an ocean. If you don’t wait, you answer instantly from a local copy that is, briefly, wrong.

Daniel Abadi named this PACELC: if Partition, choose Availability or Consistency; Else, choose Latency or Consistency. Feel the cost:

Spanner and ZooKeeper live on the left of that bar — they wait. Cassandra and DynamoDB live on the right — they answer. Nobody lives at the corner where a write is both instant and already everywhere; physics has opinions.

Where real systems sit.

Roughly — and each of these will happily argue with you about its dot at great length. What matters is the shape of the space. Notice that the far corner, the one marked CA, is not actually in the room.

The only way to hold consistency and availability unconditionally is to hold no network at all — one machine, no replicas. The moment a second copy exists across a wire, partitions enter the story, and CAP starts writing your error messages for you.

Three scenarios.

From these figures, you can trace a partitioned sale and distinguish availability from a single coherent stock count. Try changing one assumption and check whether your explanation still holds.