◀ THE FOLD0ROOT.AI // WORLD II · BOSS · THE GAUNTLET◆ .dlw.fold
THE FOLD / BOSS / THE GAUNTLET / THE CORRELATED FAILURE

THE CORRELATED FAILURE

an availability figure is a belief about shared fate
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
Three replicas at 99% give six nines — if they fail independently. They share a rack, a power feed, a kernel version and a deploy pipeline, so they do not.

LIT verified live. With a 1% single-replica failure rate and perfect independence, all three fail together with probability 0.000001: 6.00 nines. Fully correlated, they fail together with probability 0.012.00 nines, the same as one replica, and 10,000× worse. The collapse is not gradual: at a correlation of just 0.2 the figure is already 2.70 nines. Four of the six nines are gone by the time the replicas are one fifth alike.
2 HOW IT WAS WEAVED · AI + HUMAN
Correlated failure is why availability targets are written per failure domain, and why “multi-AZ” means something specific rather than “three copies”.

AVAN (AI) swept the correlation rather than contrasting the endpoints, because the endpoints suggest a trade-off and the sweep shows a cliff. Going from independent to 0.2 correlated costs more nines than going from 0.2 to fully correlated. The damage is done by the first small amount of shared fate.
3 ONE DIMENSION
Correlation against nines. The cliff is at the left.
4 TWO DIMENSIONS · INTERACTIVE
Add replicas, or admit they share something.
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object: three copies of one fate.
AVAN’s addition (the inverse-companion): the forward reading is that correlation destroys the benefit of replication. The inverse is that the independence was never measured, only assumed. Nobody computes the correlation between their replicas; the six nines come from multiplying a number by itself three times, which is an arithmetic operation performed on an assumption. Read backwards, the availability figure in the design document is not a prediction about the system — it is a restatement of a belief about how much the replicas have in common, and that belief is usually never written down at all.
LIT a 1% single-replica failure rate with perfect independence gives all three failing together at probability 0.000001, which is 6.00 nines, while fully correlated they fail together at 0.01 - 2.00 nines, the same as one replica and 10,000 times worse - and the collapse is not gradual, since at a correlation of only 0.2 the figure is already 2.70 nines

FIG Correlated failure is why availability targets are written per failure domain and why multi-AZ means something specific rather than three copies. AVAN swept the correlation rather than contrasting the endpoints, because the endpoints suggest a trade-off and the sweep shows a cliff: going from independent to 0.2 correlated costs more nines than going from 0.2 to fully correlated. The damage is done by the first small amount of shared fate.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of THE GAUNTLET · David Lee Wise (ROOT0), with AVAN