MIP-12 — North America data point (Montreal testnet node)
Since the 300 ms discussion so far rests on Japan, here is the same measurement from the other side of the world. Same caveat as ColinkaMalinka’s node: this is a testnet node outside the active set, and the gauge is only p50/p90/p99 over a 5-minute rolling window, so what follows is the share of windows whose quantile crossed the line, not a real P(delay > 300 ms).
One release only, v0.16.0, 7–26 Aug, 5,516 windows, no gaps. I kept the versions apart deliberately — folding our 0.15.x history into the same series pulled p99 exceedance from 96 % down to 82 %, and that is the release changing, not the network.
| gauge |
range |
median |
over 300 ms |
| p50 |
65–150 ms |
125 ms |
0.00 % |
| p90 |
176–266 ms |
236 ms |
0.00 % |
| p99 |
251–425 ms |
320 ms |
96.1 % |
The interesting part is that the shape differs from Japan rather than just the magnitude. Our body is tighter — p90 did not cross 300 ms once in 5,516 windows, against 5.7 % there. But our p99 sits above the line 96.1 % of the time, against 26.8 %. So this is not a noisier location; it is a location with a higher floor, and calm network conditions do not move it.
Sweeping the threshold on the same data: p99 exceeded 300 ms in 96.1 % of windows, 325 ms in 10.1 %, 350 ms in 0.49 %, 375 ms in 0.02 %. The 99th percentile of the p99 gauge is 340 ms. Put differently, 350 ms from Montreal buys about the margin that 300 ms buys from Tokyo.
michael-yxchen’s correction upthread is what makes me unsure how to read this, so I would rather ask than assert. If vote pace is minimum spacing rather than a deadline, what actually happens to a node whose readiness sits structurally above the pace? From Japan the picture is a tail that occasionally crosses during variance, which reads as harmless. From here it is not a tail event at all — 320 ms is the steady state. Those feel like different regimes to me, and I cannot tell from the MIP whether the second one costs anything.
The reason I care is geography. A floor calibrated from one vantage point ends up encoding that vantage point’s network position, and operators cannot tune their way out of the speed of light. If being structurally above the pace turns out to cost anything at all, the rational move is to relocate towards wherever the stake already is, which concentrates validators into shared transit, shared power and shared jurisdiction. Both measurements in this thread are OVH nodes, which is already a hint in that direction.
I would not want a decision resting on our numbers without someone inside the active set confirming the same shape from this region — proposal delivery may well look different for a non-voting node, and neither data point here is in consensus.
Happy to share the raw series, widen the window, or add a European vantage point if that helps. A histogram export for this metric would let all of us answer the actual question rather than approximate it with three gauges.