Benchmark Results — 2026-07-29 (pre-GA regression check)

Re-run of the read-path benchmarks against dev@bf4abed, to confirm nothing regressed in the ~800 commits and 26 audit rounds since BENCHMARKS-2026-07-04.md.

BenchmarkDotNet v0.14.0, Windows 11 (10.0.26200.8457)
AMD Ryzen 7 7840HS w/ Radeon 780M, 1 CPU, 16 logical / 8 physical cores
.NET SDK 10.0.203, host .NET 8.0.26, X64 RyuJIT AVX-512
Command: dotnet run -c Release -f net8.0 -- --filter "*QueryBenchmarks*"
         from benchmarks/Jaunty.Benchmarks
Global total time: 16m41s, 186 benchmark cases

Headline: no regression. Jaunty's read path got materially faster relative to hand-coded ADO.NET on every provider measured in July.

Scope and caveats

• Read path only (Query<T> materialization). Bulk-copy suites are still unmeasured, as in July — the README's per-provider bulk-copy ranges remain unverified.
• Same machine, OS build and runtime as the July run, so the ratios are directly comparable. Absolute times still are not: this is a laptop, and the July numbers were taken under a different thermal and background load. Treat ratios, not absolute times, as the signal.
• Row counts are 1 / 100 / 10,000. July reported 100 / 1,000 / 10,000; the 1,000 case is not in the current parameter set, so it has no counterpart below.
• Warm-job (steady-state) rows only. Cold rows are in the raw artifacts under benchmarks/results/.
• SQL Server is a local instance; PostgreSQL and MariaDB are the docker-compose.yml containers on ports 5433 and 3307. All three use a dedicated jauntybench / JauntyBench database, so the Northwind test databases were not touched.

Regression check vs 2026-07-04

Jaunty Query<T> mean, as a ratio against hand-coded ADO.NET on the same run. Lower is better.

Nothing got worse. The WithExpectedRowCount overload remains the fastest Jaunty path at scale, and now lands within 2–13% of hand-coded ADO.NET on every provider.

Warm job, 100 rows

At 100 rows over a network round trip the provider dominates, and every micro-ORM lands within a few percent of ADO.NET. The in-proc SQLite column is where the mapper itself is visible.

Warm job, 10,000 rows

Jaunty is the fastest micro-ORM tested on SQL Server, PostgreSQL and MariaDB at 10,000 rows, provided WithExpectedRowCount is used. Without it, list growth reallocation costs 8–68% depending on provider — worst on PostgreSQL (1.05x → 1.77x), which is the strongest argument for documenting that overload prominently.

RepoDb wins on in-proc SQLite at scale (1.37x faster than hand-coded ADO.NET, which means its baseline comparison is measuring something structurally different — worth a look, not a defect).

Allocation

Jaunty allocates least of the mapping libraries at scale. At 10,000 rows on SQL Server:

WithExpectedRowCount reaches allocation parity with hand-written ADO.NET — the same result as July, now confirmed across all four providers rather than two.

Harness defect found and fixed

The EF Core MariaDB cases reported NA for every row count. Cause: BenchmarkDbContext handed Pomelo _connection.ConnectionString, and MySqlConnector strips the password out of that once the connection is open unless PersistSecurityInfo=true. Every iteration threw Access denied for user 'root' (using password: NO).

Fixed in 03e0dee by adding DatabaseSetup.GetConnectionString(provider), which returns the string as configured. Verified over 279 cases with zero Access denied. Provisional EF Core MariaDB figures from that short-job run — not comparable in precision to the tables above:

The full-config numbers in the tables above were measured before this fix, which is why the MariaDB EF Core cells say "see below". Everything else in this report is unaffected — the bug was confined to the EF Core MariaDB path.

Follow-ups

• Re-run the full config now that the EF Core MariaDB fix has landed, to fill in that one cell.
• Bulk-copy suites are still unmeasured. The README's per-provider ranges remain unverified.
• Consider making WithExpectedRowCount more prominent in the docs: on PostgreSQL at 10,000 rows it is the difference between 1.05x and 1.77x.