Benchmarks

Project: benches/Nuventra.NuvexaDB.Benchmarks

dotnet run --project benches/Nuventra.NuvexaDB.Benchmarks -c Release -- --gate

Comparators: NuvexaDB (plain + encrypted), SQLite (Microsoft.Data.Sqlite), LiteDB.

1 crore (10,000,000) write / indexed read / complex query — not in CI. Writes ~/Downloads/nuvexa-1crore.nvx and can take hours and several GB:

dotnet run --project benches/Nuventra.NuvexaDB.Benchmarks -c Release -- --crore
dotnet run --project benches/Nuventra.NuvexaDB.Benchmarks -c Release -- --scale 10000

Reuse the file on a later run (skip insert if customers already has enough rows). --force recreates. --path /tmp/scale.nvx overrides the file.

Indexed equality find + limit is an IXSCAN that stops after the limit (bounded prefix). Numeric ranges still walk the index because G17 keys are not numeric-order-preserving. See changelog.md for local 1,000-row and 1-crore timings.

Frozen v1 SLOs (enforced by --gate):

• Smoke (1000 inserts): fail if NuvexaDB is more than 5× slower than both SQLite and LiteDB
• Batch insert 10k: within 1.5× LiteDB
• Encrypted point get (cache-hot, median of 3): ≤ 30% slower than plaintext, or within +50ms (Windows CI noise floor on sub-20ms totals)
• Point get @ 100k docs (2000 lookups, median of 3 after warmup): within 3× SQLite PK, or within +50ms (macOS CI noise on a ~20ms SQLite baseline)
• Crash recovery: committed insert survives Environment.FailFast (see tests/Nuventra.NuvexaDB.CrashHarness)