Benchmarks

Criterion benchmarks live in the siderite-bench crate (not published). Its README, crates/siderite-bench/README.md, lists what each benchmark measures and the latest snapshot with the machine they were taken on.

cargo bench -p siderite-bench -- --warm-up-time 1 --measurement-time 2

Run a single group by name, for example cargo bench -p siderite-bench -- routing. The groups are routing, extract, and orm. Criterion writes HTML reports to target/criterion.

Numbers are machine-specific. Compare runs only on the same machine, and treat the README’s table as a snapshot, not a guarantee.

Group	Measures
routing	dispatch through siderite’s RouterService versus a raw axum Router
extract	validating Json<T> versus raw axum::Json; Dump versus serde_json::to_vec
orm	QuerySet → QueryPlan, SQL compilation (SQLite and PostgreSQL), fetch of 100 SQLite rows

HTTP benchmarks versus FastAPI

End-to-end throughput measured with ApacheBench (ab -l) on the same machine, one server at a time (DB rounds) or side by side on different ports (plain-HTTP round). Siderite runs a --release build; FastAPI runs under Uvicorn with a single worker, no reload, and warning log level. Automated benchmark harnesses live in benchmarks/.

Machine: Darwin arm64 (Apple M3 Pro), rustc 1.99.0, Python 3.13.14, FastAPI 0.142.1, Uvicorn 0.54.0. Databases run in local containers: PostgreSQL 17, MySQL 8.4, MongoDB 8 (single-node replica set).

Plain HTTP (examples/hello_world)

Direct streaming response serialization without intermediate JSON tree construction:

Test	Siderite (req/s)	FastAPI (req/s)	Speedup	Latency p99 (S / F)
GET /	36,701	18,441	2.08x	3.5 ms / 3.6 ms
GET /hello/{name}	37,884	12,692	2.98x	3.3 ms / 5.8 ms
POST /echo (JSON)	24,084	13,510	2.00x	3.2 ms / 3.3 ms

Database-backed Todo API

A minimal Todo API (id, title, done) with GET /todos (latest 20), GET /todos/{id}, and POST /todos → 201, implemented once with the siderite ORM (SqliteBackend, PgBackend, MySqlBackend, MongoBackend, matching pool of 10) and once with FastAPI (sqlite3, asyncpg, aiomysql, motor, matching pool of 10). Tables are truncated and reseeded with 100 rows before each phase.

DB	Test	Siderite (req/s)	FastAPI (req/s)	Speedup	Latency p99 (S / F)	Notes
SQLite (default)	list 20	18,108	1,709	9.10x	2.3 ms / 28.1 ms	Rollback journal, pool 10
SQLite (default)	get one	23,592	5,598	4.37x	2.7 ms / 11.4 ms	Single row lookup
SQLite (default)	insert	2,715	1,286	2.00x	5.4 ms / 150.3 ms	In-process WriteGate eliminates busy-sleep
SQLite (group commit)	insert	8,589	1,457	5.59x	3.7 ms / 120.1 ms	--siderite-group-commit shares commit sync
SQLite (WAL)	list 20	17,406	1,705	10.62x	2.8 ms / 24.4 ms	--sqlite-wal
SQLite (WAL)	get one	20,818	6,052	3.68x	3.2 ms / 8.2 ms
SQLite (WAL)	insert	14,508	4,414	3.31x	1.9 ms / 27.7 ms
PostgreSQL 17	list 20	9,583	7,606	1.12x	6.1 ms / 9.0 ms	Pool 10
PostgreSQL 17	get one	10,723	8,422	1.01x	3.5 ms / 8.0 ms
PostgreSQL 17	insert	8,546	5,957	1.42x	1.7 ms / 2.7 ms
MySQL 8.4	list 20	10,473	5,089	2.04x	3.7 ms / 15.6 ms	Pool 10
MySQL 8.4	get one	9,886	7,386	1.39x	3.7 ms / 4.7 ms
MySQL 8.4	insert	3,371	3,433	1.02x	6.5 ms / 5.9 ms
MongoDB 8	list 20	10,729	3,825	2.92x	5.5 ms / 10.2 ms	Replica set
MongoDB 8	get one	13,049	4,116	3.05x	3.5 ms / 11.5 ms
MongoDB 8	insert	2,970	2,580	1.20x	6.8 ms / 7.2 ms

Optimization Highlights

• SQLite write gate: Siderite’s WriteGate serializes write requests across pool connections in-process via a FIFO queue. This avoids SQLite file lock races where losing connections sleep in SQLite’s busy handler (exponential backoff up to 100 ms), dropping p99 tail latency from ~150 ms to 5.4 ms.
• SQLite group commit: The opt-in SqliteBackend::group_commit batches concurrent autocommit writes into a single transaction and commit sync, driving insert throughput to 8,589 req/s (5.59x speedup).
• SQLite WAL mode: In write-ahead log mode, readers and writers operate concurrently without blocking, reaching 14,508 req/s (3.31x speedup) at 1.9 ms p99.

See also

• Development
• Architecture