← All insights
RESEARCHZcash Forum · evan-forbes

25-second blocks confirmed safe on a Zebra-only devnet

Valar Group measured the stale-block and reorg rates of 25-second blocks on 100 Zebra nodes across 19 regions and judged them to be within a safe operating range.

Evan Forbes and Dev Ojha of Valar Group shared the results of using a Zebra-only devnet to check what the stale (orphan) block rate looks like when block time is cut to 25 seconds. Shorter block times make small payments confirm faster, but propagation delay becomes a larger share of the block time, so stale blocks increase. They measured block propagation delay, block time, stale heights and reorg rate on 100 Zebra nodes spread across 19 regions and several clouds.

The theoretical model treats the stale-block rate as tau_eff divided by the target interval T. tau_eff is the work-weighted time spent mining on the old tip after a new block is found, and the probability is computed with a Poisson approximation. In a healthy 25-second chain the median block time is about 17.3 seconds, and a median below the target is normal. When there is no propagation data, 0.55 times a node's 90% reach time (reach_90) is used as the tau_eff approximation; under this assumption, a reach_90 of 2.5 seconds gives a stale rate of 5.5% and 1 second gives 2.2%.

The experiment used 100 nodes with 8 vCPUs and 16 GB each, each running one CPU miner, with TCP settings applied that greatly improve overall block propagation. The 90% reach time for blocks averaged 1.48 seconds, and the theoretical approximation gave a stale rate of 3.26%. The observed value was 4.86%, with competing blocks appearing at 52 of 1,069 heights, and the rate of switching to a branch with more cumulative work was 0.37%. Block time averaged 26.2 seconds with a median of 18.0 seconds, consistent with a Poisson process at the 25-second target.

On mainnet under similar conditions, the stale rate was about 0.1%, lower than the devnet's roughly 1%. The reasons given were that mining is concentrated in fewer than 20 pools so propagation delay is low, and that mainnet measurement is a lower bound because some stale blocks go unobserved. In conclusion, even in the worst conditions of blocks full of Orchard transactions and a highly geographically distributed network, the stale rate was under 5%. It therefore concluded that reducing the target interval to 25 seconds in NU7 is safe if the TCP setting changes are made across the network.

This summary is based on the original; check the original for exact details.