resqlite Benchmark Dashboard

Peers tested: resqlite (this repo, current commit), sqlite3.dart 3.3.0 (synchronous FFI baseline), sqlite_async 0.14.0-wip.0 (PowerSync, async + reactive), drift 2.21.0 (codegen-backed, isolate-backed, reactive). Hardware is shown in the device selector above; each device has its own full run. See SCOPE.md for the full peer list, versioning policy, and upgrade process.

Numbers reported: Wall-clock medians in ms (lower is better) from multi-repeat runs, plus p90 for tail-latency. Point-query throughput is queries/sec (higher is better). Main-isolate times show the time actually spent on the calling/UI thread — the critical budget in Flutter. Sync libraries (sqlite3) report main == wall by construction.

Fair comparison rules: identical SQL queries, identical seed data, WAL mode everywhere. sqlite_async's watch() throttle is set to Duration.zero so we're comparing invalidation engines rather than throttling policies. Each peer uses its idiomatic API (db.select() / db.getAll() / prepared Statement); no crippled versions. Methodology details live in METHODOLOGY.md.

Honesty about what these numbers do not tell you. The suite is designed to expose the behaviors Flutter app authors care about, but it's not exhaustive.

Peers not included:

• sqflite — excluded until mobile CI lands. The desktop-compatible variant (sqflite_common_ffi) uses sqlite3.dart under the hood and would not represent actual mobile sqflite performance (platform channel serialization cost, native iOS/Android SQLite version differences, F2FS vs APFS fsync behavior). Including it with a caveat would be misleading; excluding and saying why is honest. Will be added when flutter test integration_test/ on simulators/emulators is wired into CI.
• libsql_dart — still maturing as of this writing. Revisit when stable.

Workloads not included:

• Mobile device results — all published numbers come from a desktop Apple Silicon machine. Android / iOS performance differs (LPDDR4 vs LPDDR5, F2FS vs APFS fsync, Dart FFI first-call overhead on iOS). If you deploy on mobile, our numbers describe a different environment.
• Encryption (AES-256 via SQLite3 Multiple Ciphers) — shipped feature, not yet benchmarked. Encryption is never free; we don't quantify the cost here.
• Sustained write-heavy workloads — our bulk insert (A7) is a burst, not sustained traffic. Long-lived "100 writes/sec for 10 minutes" patterns (logging, telemetry) aren't covered.
• Transaction contention — we test single-writer transactions. Many concurrent callers racing for the writer isolate are out of scope.
• Cold-start time — how long from Database.open() to first query? Unmeasured. Matters for CLI tools and add-to-app patterns.
• Battery & memory footprint — requires device-level instrumentation that this suite doesn't attempt.

The SCOPE.md Known Gaps section is the canonical list; this summary is kept in sync with it.