THE FOLD / CO-OP / THE SYNC / THE CLOCK SKEW
THE CLOCK SKEW
one millisecond, 189 wrong orders
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
Two machines timestamp two events. The second event genuinely happened after the first. Whether the timestamps agree depends on how well the two clocks are set, and nothing in the data says how well that was.
LIT verified live. ten thousand pairs of events, the second always genuinely later by up to 20 ms. At 0 skew the timestamps invert 0 times. At 1 ms of skew they invert 189 times, at 5 ms 848, at 10 ms 1,687, at 25 ms 3,205 and at 50 ms 3,954 — 39.5%, approaching the half you would get from a coin. The inversion count rises monotonically with skew across every step, and no inverted pair is distinguishable from a correctly ordered one by looking at it.
LIT verified live. ten thousand pairs of events, the second always genuinely later by up to 20 ms. At 0 skew the timestamps invert 0 times. At 1 ms of skew they invert 189 times, at 5 ms 848, at 10 ms 1,687, at 25 ms 3,205 and at 50 ms 3,954 — 39.5%, approaching the half you would get from a coin. The inversion count rises monotonically with skew across every step, and no inverted pair is distinguishable from a correctly ordered one by looking at it.
2 HOW IT WAS WEAVED · AI + HUMAN
Clock skew is why distributed systems use logical clocks — Lamport (1978) and vector clocks — rather than trusting wall-clock ordering, and why Spanner buys atomic clocks to bound it instead.
AVAN (AI) swept the skew rather than asserting that ordering is unsafe, because the curve is the argument. The failure does not appear at some threshold; it is already 1.4% at a single millisecond and degrades smoothly to a coin flip. There is no skew small enough to make the ordering sound, only skew small enough to make the errors rare — which is a different property, and not the one anyone thinks they are relying on.
AVAN (AI) swept the skew rather than asserting that ordering is unsafe, because the curve is the argument. The failure does not appear at some threshold; it is already 1.4% at a single millisecond and degrades smoothly to a coin flip. There is no skew small enough to make the ordering sound, only skew small enough to make the errors rare — which is a different property, and not the one anyone thinks they are relying on.
3 ONE DIMENSION
One millisecond of skew. 189 events in the wrong order.
4 TWO DIMENSIONS · INTERACTIVE
Turn up the skew and watch causality dissolve.
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object.
AVAN’s addition (the inverse-companion): the forward reading is that you cannot order distributed events by wall clock. The inverse is that the timestamps were never claims about order — they are claims about the reading of a local dial, and order is something a reader inferred. Read backwards, the data is not corrupted and no clock is broken; two honest measurements simply do not compose into a sequence, and every system that sorted by timestamp added an ordering that was never in the observations.
LIT ten thousand pairs of events, the second always genuinely later by up to 20 ms: at 0 skew the timestamps invert 0 times, at 1 ms of skew 189 times, at 5 ms 848, at 10 ms 1,687, at 25 ms 3,205 and at 50 ms 3,954 - 39.5%, approaching the half you would get from a coin - with the inversion count rising monotonically across every step and no inverted pair distinguishable from a correctly ordered one
FIG Clock skew is why distributed systems use logical clocks - Lamport (1978) and vector clocks - rather than trusting wall-clock ordering, and why Spanner buys atomic clocks to bound it instead. AVAN swept the skew rather than asserting that ordering is unsafe, because the curve is the argument: the failure is already 1.9% at a single millisecond and degrades smoothly to a coin flip. There is no skew small enough to make the ordering sound, only small enough to make the errors rare.
FIG Clock skew is why distributed systems use logical clocks - Lamport (1978) and vector clocks - rather than trusting wall-clock ordering, and why Spanner buys atomic clocks to bound it instead. AVAN swept the skew rather than asserting that ordering is unsafe, because the curve is the argument: the failure is already 1.9% at a single millisecond and degrades smoothly to a coin flip. There is no skew small enough to make the ordering sound, only small enough to make the errors rare.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of THE SYNC · David Lee Wise (ROOT0), with AVAN