# Previous RadixDB 100M NVMe baseline

[Русский](FINAL_100M_NVME_ACCEPTANCE_REPORT.ru.md)

Validation date: 2026-09-04.

## Candidate and environment

- source/build revision: `cce945773038e150845760e82327d81c5cdb2eae`;
- version `0.5.1`, protocol `13`, release profile;
- Cargo.lock SHA-256:
  `8e93a106d8c2665343b0602d993c8e8fa0147dbb0109d6befc5e1a47b7b1da71`;
- binary SHA-256: server
  `3db6c6d0af92a896f477cf2190423f8caa8f6abc73f018cf0d6a2ca11749bd2b`,
  benchmark
  `729174e885fab5bd7bc01e4eb6b0898c0524067c12d0081c0942bd6e137984c0`;
- AMD Ryzen 9 7950X, Apacer AS2280Q4U 2 TB NVMe, Btrfs;
- Linux `7.1.3-201.fc44.x86_64`, Rust/Cargo `1.97.0`;
- page-cache level `0`, storage workers `auto`, release build;
- the 100M fixture and workload were unchanged from the two preceding baselines.

## Command

```bash
target/release/radixdb-bench \
  --run --verify-existing --scale 100m --participant server \
  --expected-checksum 100000000:49734600639880 \
  --warmup 1 --repeats 5 --allow-large \
  --root /path/to/benchmark-root \
  --run-id radixdb-final-100m-rN-20260904
```

The same binary was used for three consecutive runs. The value published in
the comparison table is the median of the three run-level results; each query
result is itself the median of five measured repetitions after one warm-up.

## Correctness and resources

All three runs confirmed checksum `100000000:49734600639880`, the expected
cardinalities and the absence of errors.

| Metric | Run 1 | Run 2 | Run 3 | Median | Previous baseline | Delta |
|---|---:|---:|---:|---:|---:|---:|
| full verify elapsed | 14.337 s | 13.501 s | 14.274 s | 14.274 s | 14.800 s | -3.6% |
| peak RSS | 1.60 GiB | 1.64 GiB | 1.62 GiB | 1.62 GiB | 1.62 GiB | 0% |
| peak threads | 39 | 39 | 39 | 39 | 39 | 0 |
| peak open FDs | 135 | 134 | 134 | 134 | 135 | -1 |
| process read | 852 MiB | 0 B | 0 B | 0 B | 0 B | cache-hot parity |
| device read | 852 MiB | 16 KiB | 32 KiB | 32 KiB | 16 KiB | cache-hot parity |
| active database footprint | 2.48 GiB | 2.48 GiB | 2.48 GiB | 2.48 GiB | 2.48 GiB | 0% |
| files / directories | 364 / 125 | 364 / 125 | 364 / 125 | 364 / 125 | 364 / 125 | 0 |

The first run is a cold-cache diagnostic and physically read 852 MiB; the next
two are comparable cache-hot measurements. Median scan throughput was about
8.94 million rows/s for the full row and 39.58 million rows/s for the projected
scan.

## Latency

A negative delta means an improvement. The table includes the original
accepted 100M baseline, its immediate successor and this final result. The stop
threshold for a sustained regression is `+25%`.

| Case | Original baseline, ms | Previous baseline, ms | Candidate, ms | Delta vs previous | Delta vs original |
|---|---:|---:|---:|---:|---:|
| `server.cold_start_select_database` | 2757.583 | 3090.832 | 3061.895 | -0.9% | +11.0% |
| `correctness.seed_checksum` | 12.112 | 12.352 | 12.497 | +1.2% | +3.2% |
| `select.pk` | 0.483 | 0.483 | 0.481 | -0.4% | -0.4% |
| `select.range` | 0.480 | 0.495 | 0.485 | -2.0% | +1.0% |
| `scan.full` | 88.238 | 98.787 | 93.247 | -5.6% | +5.7% |
| `scan.projected` | 20.865 | 22.381 | 21.055 | -5.9% | +0.9% |
| `aggregate.group_having` | 10.439 | 10.619 | 10.286 | -3.1% | -1.5% |
| `join.parent` | 20.038 | 19.578 | 17.945 | -8.3% | -10.4% |
| `reference.navigation` | 28.771 | 30.143 | 27.313 | -9.4% | -5.1% |
| `reference.direct.explicit` | 20.287 | 20.167 | 18.233 | -9.6% | -10.1% |
| `reference.fact_dictionary` | 682.622 | 717.161 | 725.155 | +1.1% | +6.2% |
| `reference.fact_first` | 415.482 | 429.069 | 415.987 | -3.0% | +0.1% |
| `reference.target_first` | 382.919 | 406.923 | 378.050 | -7.1% | -1.3% |
| `update.rollback` | 50.887 | 54.332 | 44.532 | -18.0% | -12.5% |
| `delete.rollback` | 3.735 | 2.033 | 1.746 | -14.1% | -53.3% |

No case approached the stop threshold. The only slowdowns against the previous
baseline were `correctness.seed_checksum` (+1.2%) and
`reference.fact_dictionary` (+1.1%); both are within the noise range. The
access-path sections of all three reports were byte-identical, with SHA-256
`256517746bd0166744b5fdf25286fd9e7ebc91b48fca4bb5245e4627432c32a9`.

## Evidence

The complete raw reports are stored outside the public Git repository. Their
local benchmark-host paths are intentionally omitted from this public copy.

Their SHA-256 values, in run order, are:

- `d112fc53aa71bec3ece08bb588b57271282ebf00d66dc4e7d97ed96e63ce478b`;
- `23b7dd23445df28c29cda3f2e4d97b01715237944f42e37594f8984856c2c62e`;
- `5242b929c41387e2cdd23a47553103e10c7ce64b1933d4103c1644da7cfb51cd`.

## Verdict

The result is accepted. On the canonical 100M database, the final crate
structure preserved correctness, readback, access paths, footprint and bounded
resources. No sustained latency or throughput regression was observed against
the two stored baselines.