Zero simulated metrics. Every single datapoint below was generated by automated benchmark runs executed on physical hardware, measuring ULTSQL (Pure Dart & Standalone AOT Native Binary) directly against SQLite3 (native C).
Zero Dart VM, zero JIT, and zero Dart SDK overhead. We benchmarked the standalone native AOT executable
(ultsql.exe / ultsql) directly against the native C library SQLite 3
(sqlite3.dll v3.50.4) on an NVMe SSD using a pure Python 3 harness across 50,000 rows.
| Benchmark Workload | Unit | โก Native UltSQL (ultsql.exe) |
๐ชถ Native SQLite 3 (sqlite3.dll) |
Empirical Result |
|---|---|---|---|---|
| 1. Bulk Ingest (Transaction Batch) | rows/s |
3,357,282 | 1,710,922 | ๐ 1.96x faster |
| 2. CREATE INDEX on age (B+Tree) | ms |
2.739 | 8.197 | ๐ 2.99x faster |
| 3. Point Lookup by PK (Prepared) | ยตs/query |
0.682 | 3.474 | ๐ 5.09x faster |
4. Range Aggregate (age 30โ40) |
ms |
0.222 | 1.475 | ๐ 6.64x faster |
| 5. Full-scan GROUP BY Aggregate | ms |
0.692 | 8.951 | ๐ 12.93x faster |
| 6. Indexed In-Place UPDATE | ms |
0.108 | 3.203 | ๐ 29.66x faster |
| 7. Indexed DELETE | ms |
0.028 | 19.992 | ๐ 714.00x faster |
8a. Auto-Commit (synchronous = NORMAL) |
tx/s |
519,481 | 94,437 | ๐ 5.50x faster |
8b. Relaxed Durability (synchronous = OFF) |
rows/s |
4,545,455 | 657,895 | ๐ 6.91x faster |
| 9. DROP TABLE Catalog Cleanup | ms |
0.004 | 0.916 | ๐ 229.05x faster |
Durability Distinctions: Comparing buffered group-commit writes against SQLite's default per-commit fsync() (which yields ~1,064 tx/s due to NVMe disk sync latency) is not apples-to-apples. Under matching WAL durability (PRAGMA synchronous = NORMAL), SQLite reaches 94,437 tx/s, while UltSQL achieves 519,481 tx/s. When disk synchronization is relaxed entirely (synchronous = OFF), SQLite achieves 657,895 rows/s on disk and 1.32M rows/s in memory, while UltSQL's batch group-commit reaches 4.55M rows/s.
Out-of-Process Crash Safety Proof: UltSQL is rigorously verified with an out-of-process crash harness (test/out_of_process_crash_test.dart) executing kill -9 / taskkill /F abruptly mid-batch. In both single-crash and multi-cycle recovery tests, 100% of acknowledged committed transactions survive, the Write-Ahead Log replays cleanly, and relational SQL correctness is verified against standardized SQL logic conformance suites.
# Clone and run the native benchmark script (requires only Python 3 + bundled sqlite3)
git clone https://github.com/ompatel3158/ULTSQL.git
cd ULTSQL
python tool/benchmarks/benchmark_native_sqlite_vs_ultsql.py
Empirical statistical medians across 3 timed runs per workload.
| Workload Benchmark | Unit | โก ULTSQL (Pure Dart) | SQLite (Native C/FFI) | Winner / Speedup |
|---|
Clone the repository and run the open-source benchmark harness directly via Dart:
# 1. Clone ULTSQL
git clone https://github.com/ompatel3158/ULTSQL.git
cd ULTSQL
# 2. Run automated comparative benchmark harness (relational, nosql, kv, vector)
dart run tool/benchmarks/run_all.dart --rows 50000 --runs 3 --big 200000
# Results will be generated and saved to web_site/data/benchmarks.json