# Spatial and durable-write report extracts

Prepared 2026-09-13. These are source-grounded extracts of internal measurement reports, not new runs. The original reports were reviewed through Google Drive. The SHA-256 values below bind the exact UTF-8 source text returned by that connector, not independently downloaded original upload bytes. No raw log for these two comparisons is included in this attachment.

## Historical spatial Q1, May 7, 2026

Original report: `PERF-SPATIALBENCH-RESULT-SF1-2026-05-07.md`.
Original report source-text SHA-256: `01b461fb6db3aa06bbe4ea6fb7f466bc107acce707bfa2c721ed6cfdb5af8dd1`.
Located copy modified: 2026-05-13T21:32:29.296Z.

| Q1 fixed-center radius filter | Mean of 3 runs | Reported output rows |
|---|---:|---:|
| 8DB spatial prototype | 14.7 ms | 94 |
| SedonaDB 0.3.0 | 470 ms | 94 |
| DuckDB 1.5.2 with spatial | 140 ms | 94 |

The rounded timings imply approximately 32× versus SedonaDB and 9.5× versus DuckDB for this operation. Hardware was AWS `m7i.2xlarge` in Canada Central, 8 vCPU Intel Sapphire Rapids at 3.2 GHz, 32 GB RAM, x86-64. The deterministic synthetic SpatialBench SF1 dataset contained six million trips, 20,000 buildings and 12,000 zones.

Q1 tested distance at most 0.45 in the dataset's planar coordinate units from a fixed target. 8DB used a handwritten Rust loop over pickup coordinates with bounding-box prefiltering and a squared-distance test. Competitors ran through SQL engine paths. This is a historical spatial operator/kernel measurement; it is not a claim of equivalent whole-database query stacks. Timing excludes loading and setup. Equal row counts do not independently prove equal output record IDs.

The source names branch `feat/spatialbench-stream-a`, but does not unambiguously pin the original baseline 8DB binary commit. Its bootstrap SHA is not substituted for that missing pin. The report names `sf1_trustdb_final.json`, `sf1_competitor_baseline.json` and `sf1_sedonadb_full.json`; those raw JSON files were not located during this audit. They and the exact source revision should be recovered before treating this as a fully replayable comparison.

Only Q1 is presented here. The same report identifies a Q8 output-count mismatch and leaves recall validation open for a later Q12 optimization. Accordingly, this extract does not claim a complete SpatialBench sweep, an official leaderboard rank or a correctness-certified suite result. No peak-memory, cost-per-query or sustained-concurrency claim is made. The upstream benchmark project is [Apache SpatialBench](https://github.com/apache/sedona-spatialbench).

## RocksDB single-write comparison, September 8, 2026

Original report: `M1-CYC32-ROCKSDB-MATCHED-DURABILITY.md`.
Original report source-text SHA-256: `0dabf2cda34128008484cdc1dabb9e6a929185abd31954253c388687c72f8ae8`.
Located copy modified: 2026-09-08T16:42:30.089Z.

| Single write, 8-byte key, 100-byte value, batch 1 | Reported p50 range |
|---|---:|
| 8DB SyncNow | 5.54–5.64 ms |
| RocksDB 11.8.1, `sync=1`, WAL on | 4.67–4.74 ms |

RocksDB was approximately 1.19× faster in this recorded configuration. Hardware was a MacBook Pro M1, 4P+4E cores, 16 GB RAM, macOS and internal SSD. The report identifies 8DB fix commit `b42c075a` and a measured binary on the `2b107249` lineage, not one unambiguous complete measurement commit. RocksDB was the Homebrew 11.8.1 bottle through its C API with stock options and WAL enabled. The report describes two interleaved rounds, with RocksDB measured between the two 8DB runs.

The source labels these configurations power-safe and discusses `F_FULLFSYNC` barrier measurements. This attachment reports the latency result and the source's intended contract; it does not independently certify power-loss survival, inspect every comparator durability setting or prove an equivalent end-to-end recovery contract. The source's separate weaker-barrier mode is not compared with a stronger-durability RocksDB mode as a speedup.

The original report names `w13-r1.log` and `w13-r2.log`. Those raw logs were not retrieved for this extract. This remains report-only evidence, without a fresh run or independent durability validation. The source names `rocksdb_sync_bench.c`, `barrier_classes.c` and the 8DB W13 workload as its measurement paths; they are not included here.

## Publication treatment

Both sections are edited report extracts. Internal machine/workspace paths, operational setup instructions, private workflow references and speculative future gains are omitted. No selected metric value or material known limitation was changed. Historical source naming is rendered as 8DB in the explanatory prose. The original source titles and exact source-text hashes remain visible for provenance. These attachments disclose existing measurements and do not include the private database engine or grant a new blanket license.