A cyan SQLite data page sends one query trace into three illuminated silicon test layers.

Reproducibility instrument / v0.1

Performance evidence that travels with your SQLite release.

Pin the fixture. Repeat the workload. Record query plans and CPU features. Catch a 15% regression before your users catch a SIGILL.

Output
JSON + Markdown
Threshold
> 15% fails CI
Telemetry
None

One record, not a timing screenshot

A release claim becomes a record

The context is part of the measurement.

A fast query and a portable binary are different claims. The lab keeps the workload identical while making the execution environment impossible to miss.

  1. 01

    Declare

    Pin SQLite, fixture bytes, PRAGMAs, repetitions, queries, and at least three CPU profiles in one TOML file.

  2. 02

    Measure

    Run the same binary on hardware, in a container, or under QEMU. Every report keeps its evidence label.

  3. 03

    Gate

    Compare matching contexts. Changed results stop the comparison; median regressions above your threshold stop CI.

Recorded walkthrough

Follow the evidence plane by plane.

This static walkthrough mirrors fields emitted by the real CLI. No workload or report is uploaded.

RUN / candidate 8FB2—A09C / LOCAL
01x86-64-v2Emulator evidence
02x86-64-v3Container evidence
03arm64Hardware evidence
Command
$ sqlite-workload-lab run lab.toml \
  --profile x86-64-v2 --out reports/candidate

 Fixture sha256  7c50…6a2e
 SQLite         3.50.4 / pinned
 Repetitions    12 measured + 2 warmups
! Evidence       emulator / x86_64

Compatibility without ambiguity

Three environments. Three honest labels.

Hardware

Release host

Best evidence for actual latency and throughput on that named machine.

Container

Pinned userspace

Reproduces libraries and build context; it does not change the host CPU.

Emulator

Instruction floor

Exercises an older CPU model and exposes leaked AVX flags before release.

Local by construction

From empty folder to reviewable report.

One Rust binary. A checked-in TOML manifest. No account, token, collector, hosted runner, or automatic tuning.

Read the full workload schema
Quick start
# Install from a clone
$ cargo install --path .

# Create and validate a pinned workload
$ sqlite-workload-lab init lab.toml
$ sqlite-workload-lab check lab.toml

Valid: 1 query across 3 CPU profiles.

Your database never enters this page.

The CLI has no networking or telemetry. The documentation site stores no personal data and uses no analytics, cookies, third-party scripts, or remote fonts.