the second hundred · metaphor 182
A family, a company, a couple — any pair of people who have to keep acting as one — wants three things at once: to agree on the truth, to stay responsive to whoever needs them, and to keep working even when contact between them breaks. The hard news is that once contact does break, you can have any two of the three, never all three.
While the line is open it costs nothing: you agree, you respond, life goes on, and it feels as though all three are simply yours. Then the rift arrives — a silence, a distance, a fight that stops the two of you talking — and suddenly you're forced to choose. Do you freeze and refuse to act alone, so that nothing you both decide can ever drift out of sync? Or do you each keep answering the people in front of you, knowing that the two of you are quietly deciding different things?
There is no third door. You cannot stay perfectly in agreement and stay fully responsive and survive the break — the moment the break is real, one of those three has to give. Distributed computing proved this about machines and called it the CAP theorem; it is just as true of anyone who has tried to stay one team through a season of not speaking.
The trilemma
Two replicas hold a copy of the same truth — a number, a plan, an account balance. While the link between them is up, a write to one is copied to the other before it counts, so they never disagree; a read from either returns the same fresh answer; every request is served. All three wishes are satisfied and it costs you nothing to have them. This is the comfortable state most relationships live in most of the time, and it is exactly why the theorem is so easy to forget.
Cut the link and the arithmetic bites. A write now arrives at one replica that cannot reach the other. It has two choices, and only two. It can refuse — hold the line, serve nothing it can't confirm, keep the two copies identical by simply not changing anything it can't propagate. That preserves consistency at the cost of availability: the refused writes are the price. Or it can accept locally and let the two copies drift — staying fully available at the cost of consistency: now the two replicas quietly hold different truths, and reads start coming back stale.
What you cannot do is refuse nothing and drift nowhere while the link is down. That would require the two copies to coordinate — which is precisely what the rift forbids. And you cannot escape by dropping the third wish either: giving up partition-tolerance just means pretending rifts never happen, which works right up until one does and takes both other guarantees down with it. So: keep the rift, and pick two.
What to try
Every meter is computed from the live request stream — availability is the fraction of recent requests actually served, consistency is the fraction of recent reads that returned the freshest value written anywhere, and both update as the stream flows. In peace, both sit at 100% and the P meter reads untested — you're enjoying C and A precisely because the rift hasn't come. Leave it on Consistency (CP) and press Cut the link: writes start bouncing back refused, availability slides down in real time, but consistency stays pinned at 100% — the two replicas never disagree because neither is allowed to drift.
Now flip to Availability (AP) and cut again. This time nothing is refused — availability holds at 100% — but the two replicas take writes independently and pull apart on the canvas, and consistency crumbles as more and more reads come back stale, showing the value from a replica that missed the latest write. Turn up write pressure to make either wound deeper. Then press Heal the link: under AP the panel has to reconcile the divergence, and one replica's writes are overwritten to force agreement again — the consistency you spent, made visible as the cost of having stayed available.
The mapping
Co-parents in different houses. Two founders in a silence. A couple weathering a stretch where they can't really talk. Each pair is a two-replica system holding a shared truth — the rules for the kids, the direction of the company, the state of the relationship — and each faces the same three wishes: agree, stay responsive to the people who need you, and keep functioning even when the two of you are out of contact. The theorem says the third wish is the one that changes everything: when contact holds, the tension is invisible; when it breaks, it is forced into the open.
So the wisdom isn't avoid the tradeoff — you can't. It's choose it on purpose, in advance, per decision. Some truths must never diverge, so you agree to freeze them when you're out of touch: nothing irreversible happens unilaterally, even at the cost of responsiveness (a CP marriage on the big questions). Other truths are better kept moving even if the two of you drift, because a stale answer beats a frozen one and you can reconcile later (an AP arrangement on the daily ones). Mature partnerships are not the ones that beat CAP. They're the ones that decided, before the rift, which guarantee they were willing to lose.
Read as life lessons
While contact holds you seem to have all three, so you never decide which to keep. The rift doesn't create the tradeoff — it reveals one that was always there, and now demands an answer.
Freezing — declining to act on what you can't confirm together — isn't failure; it's the price of agreement. Every unanswered request is consistency, bought. The question is only whether it's worth it here.
Staying responsive by acting alone doesn't avoid the cost — it postpones it to reconciliation, when someone's version gets overwritten. Availability today is a consistency bill that comes due at the reunion.
In the wild
Money is CP: rather than risk two branches spending the same balance, a partitioned system refuses the transaction. You'd rather be told “try later” than have your account silently disagree with itself.
Shopping carts and social feeds are AP: better to stay responsive and reconcile later than to go dark. If two devices edit a cart during an outage, the system merges the mess afterward.
Couples and co-founders quietly pick per topic — freeze the irreversible decisions until you can talk (CP), keep the reversible ones moving and sort out conflicts on reunion (AP).
The mapping, exactly
| Distributed system | Life |
|---|---|
| two replicas | Two people holding a copy of the same shared truth — the plan, the rules, the state of things. |
| the partition | The rift: a silence, a distance, a fight that stops the two of you from staying in contact. |
| consistency (C) | Never disagreeing — both of you always on the same version of the truth, no surprises on reunion. |
| availability (A) | Always able to respond — answering the people in front of you without waiting for the other. |
| partition-tolerance (P) | Continuing to function at all while out of contact, rather than the whole thing simply halting. |
| reconciliation on heal | The reunion, where drift is resolved — and someone's version, made in good faith while apart, gets overwritten. |
The honest model
Two replicas each hold a value stamped with a logical clock. A stream of requests arrives, each a write (with probability set by the slider) or a read, aimed at a random replica. With the link up, a write bumps a global clock and copies to both replicas at once, so their stamps stay equal; every read returns that single fresh value. The availability meter counts served-over-total requests; the consistency meter counts, among reads, how many returned the value carrying the maximum clock anywhere — the true latest write. In peace both are exactly one.
Cut the link and the branch splits by strategy. Under CP, a write that can't reach its peer is refused — nothing is served that would let the copies diverge, so consistency stays at one while availability falls by roughly the write share. Under AP, the write commits to its local replica only, advancing one stamp and not the other; the copies drift, and every read aimed at the replica lacking the newest write is counted stale — availability stays at one while consistency decays. Healing an AP divergence runs last-writer-wins: the higher clock overwrites the lower, and the panel tallies the writes it just erased. Nothing here is scripted; the meters are bookkeeping over the actual request stream, and they can never all read green during a cut.
Where the metaphor tears
CAP only forces the tradeoff while a partition is happening. The rest of the time a system — or a couple — can have consistency and availability both. Reading CAP as “you can only ever have two of these” is the classic misread; the honest statement is “when contact breaks, you must sacrifice one.” Most of life is not a partition.
Real systems don't choose C or A once and for all; they tune consistency per operation, tolerate bounded staleness, degrade gracefully. Likewise people rarely go fully rigid or fully independent — they negotiate how much drift is acceptable on which questions. The triangle's crisp corners are a teaching cartoon of a landscape with a thousand intermediate points.
The model resolves conflicts by discarding the loser — but the best human reunions don't. Two people who each grew while apart may find their divergence was generative: not a conflict to resolve down to one version, but two truths to hold. Last-writer-wins is a machine's mercy, and a poor model of a partnership that gets richer for having been briefly two.