Final release measurements
These source-bound results compare the complete V5, V7 and final V8 engines on Python 3.12.14 / Linux / overlay. Each trial includes admission, automatic maintenance, persistence and an explicit final checkpoint drain. Setup/open and final value/receipt validation are excluded. All 63 trials passed unchanged-source guards and exact contents/receipt validation. These are POSIX regular-file prototype measurements, not native NTFS results.
Single-key workloads
Each trial has 128 durable single-key transactions with 128-byte values and a 32,768-block arena. There are three trials per workload with rotated backend order. The table shows medians and ratios of medians.
| Initial files | Workload | V5 tx/s | V7 tx/s | V8 tx/s | V8/V7 | V8/V5 |
|---|---|---|---|---|---|---|
| 4,096 | Hot prefix | 2006.66 | 1600.22 | 1908.01 | 1.192x | 0.951x |
| 4,096 | Uniform | 1782.60 | 1418.28 | 1472.93 | 1.039x | 0.826x |
| 16,384 | Hot prefix | 2089.76 | 1413.31 | 1411.53 | 0.999x | 0.675x |
| 16,384 | Uniform | 1265.12 | 794.46 | 977.93 | 1.231x | 0.773x |
| 65,536 | Hot prefix | 1592.08 | 1138.89 | 1370.66 | 1.204x | 0.861x |
| 65,536 | Uniform | 1214.78 | 910.83 | 841.64 | 0.924x | 0.693x |
V8/V7 single-key throughput ranges from 0.924x to 1.231x. The 65,536-file uniform case is 7.6% lower than V7; the 16,384-file hot-prefix case is effectively equal in these three-trial medians. Against V5, V8 ranges from 0.675x to 0.951x, with a worst measured throughput deficit of 32.5%. V5 has the weaker arena-admission contract. These results support neither a universal speedup nor an NTFS-superiority claim.
All three engines request identical host-write bytes and perform 132 fsync calls in each single-key trial. Every V8 single-key trial has 128 fast accepts, zero exact fallbacks, two exact compiler calls and zero full-graph scans.
Separate growing-cell stress workload
This workload starts with an empty published namespace and performs 64 atomic transactions, each adding 128 new keys with 128-byte values: 8,192 logical key updates. It deliberately exposes growth of a pending reservation cell anchored to an empty root. Transaction throughput and logical-key throughput differ by 128 and must not be substituted for the single-key rates above. There are three trials per backend, with the same complete final-drain timing boundary.
| Engine | Durable atomic tx/s | Logical key updates/s | Final checkpoint ms |
|---|---|---|---|
| V5 | 186.96 | 23930.40 | 224.60 |
| V7 | 37.66 | 4820.00 | 247.06 |
| V8 | 135.38 | 17328.43 | 281.08 |
V8 improves 3.595x over V7 on this specific structural stress and reaches 0.724x V5 throughput. All three engines request identical host-write bytes and perform 66 fsync calls. V8 has 64 fast accepts, zero exact fallbacks, one exact compiler call and no full-graph scan. The gain demonstrates this workload's reduced reservation recounting, not a general filesystem frontier.
Interpretation and reproduction
The weighted persistent budget keeps encoded entry weights and leaf sums, copies the affected symbolic path, and caches subtree page counts. This removes repeated serialization and repartitioning of a growing pending cell. The single-key profile already showed much of the remaining time in actual checkpoint construction/validation and fsync, so ordinary-workload gains are more modest and some regressions remain.
Raw trials, latency samples, engine counters, allocation metadata, logical-workload digests and source hashes are in benchmark_final_results.json. Counter names are implementation-scoped: for example, reservation_entry_encodings counts initial published-leaf weights; a zero value for empty genesis does not mean newly offered keys required no encoding.
Run the frozen experiment from the source root:
python3 experiments/benchmark_final.py --namespaces 4096 16384 65536 --operations 128 --trials 3 --pool-blocks 32768
Add --output rerun_benchmark_results.json to preserve the supplied result. No simulated flush delay, direct IO, cache dropping, NAND traffic measurement or physical power-cut qualification is used. All live logical data remain in RAM.
Engine and benchmark hashes:
{
"benchmark": "6bdd994db6ab30d9a30e1e393534d18528218bfe4c2afc3f4a7fb5840f7b2f74",
"v5": "3685cd31ebf282050e26001439f3bd06b9cf3cff765ab7085399786c2f9616bd",
"v7": "a67c93290fd561eff480f84b483dee0b44833c694da9b96c5d830af00616b980",
"v8": "e22e219b85e4b2a777b295ecad616cf4ef400e0189d7d7583d14f432d89ed807"
}