Performance Benchmarks
Real numbers from production-validated parsing infrastructure
1.38M+
ops/sec sustained
End-to-end parse throughput under continuous load
8M+
tokens/sec
Zero-copy tokenizer throughput
20K+
concurrent ops
Thread-safe with zero race conditions
Parse Benchmarks
| Benchmark | Query Type | Apple M4 | Baseline (CI) |
|---|---|---|---|
| Simple SELECT | SELECT id, name FROM users | ~470 ns/op | 700 ns/op |
| Complex Query | JOIN + WHERE + ORDER BY + LIMIT | ~2000 ns/op | 3000 ns/op |
| Window Function | ROW_NUMBER() OVER (PARTITION BY ...) | ~860 ns/op | 1300 ns/op |
| CTE | WITH clause query | ~800 ns/op | 1200 ns/op |
| INSERT | Simple INSERT statement | ~670 ns/op | 1000 ns/op |
Methodology
- All benchmarks use
ParseFromModelTokens— the production code path with zero-copy tokenization. - Apple M4 numbers are measured locally; CI baselines include tolerance (60-65%) for shared runner variability.
- Memory efficiency achieved via
sync.Poolobject pooling for AST nodes, tokenizers, and statements. - Concurrency tested with 20,000+ goroutines under race detection — zero races found.
- Data sourced from
performance_baselines.json(v1.10.4, updated 2026-03-13).