Zakura 1.2.0 cuts block processing by 87%
By removing duplicated proof verification during block acceptance, it cut processing time for a full Ironwood block from 2.5 seconds to 0.32 seconds. Mining pools benefit most.
Zakura 1.2.0 is a performance release. The network's orphan block rate rose after zcashd support ended, and this release removes the duplicate proof verification during block acceptance that was identified as the biggest cause. Processing time for a full block of ordinary Ironwood transactions fell from 2.5 seconds to 0.32 seconds on benchmark equipment, an 87% reduction. All node operators are advised to update, and mining pools benefit the most.
As background, measured on Zebra 4.3.0, the Mainnet orphan rate was about 0.1%. After zcashd support ended, blocks taking over 10 minutes became more frequent than statistics would expect, and on July 31 the orphan rate neared 1%. Connection fixes and other changes reduced the initial spike, but the baseline was around 0.3%, and on days with many quarter-full blocks it rose to 1%. The cause was that every proof was verified twice, once when propagating through the mempool and again when arriving in a block. This is related to the proof-verification cache having been removed in Zebra 4.3.1 as part of a critical security fix.
Zakura 1.2.0 caches the stateless proof verification itself, not the verdict. Verifying a single Halo2 item is a pure function of its input, so a cache hit gives a result that is bit-for-bit identical to the original computation. Even on a hit, transaction verification still runs to the end against the block's height, time and spent outputs, and because the cache key covers the whole transaction, a tampered transaction cannot reuse it. In an internal testnet of 77-80 nodes, the orphan rate in the 25-second interval was measured at about 30% falling to 3.6%, and 1.57% in the 75-second interval. The absolute values are for that environment, not Mainnet.
In addition, GetBlockRange, proposed and contributed by @maxdesalle, was added to the indexer gRPC, streaming a range of finalized blocks as stored raw bytes. Mutual TLS is now mandatory for indexer RPC on non-loopback addresses, connections and streams are limited, and protocol services enforce a declared frame limit. Chain work is stored at the full 256-bit width, and the maximum block time activation height became a per-network parameter. The end-of-support halt keeps its 40-day window and stops at Mainnet height 3,495,707 (around September 23, 2026).
This summary is based on the original; check the original for exact details.