Code gets slower one pull request at a time. Nobody means it. A small change here, a convenient library there, and each one passes review because nobody measured it. Months later the product is noticeably slower and there’s no single change to blame.
jvq measures every one. It’s a performance budget gate for CI: it runs your benchmarks on the base branch and on the change, alternating on the same CI runner, and fails the pull request when the change is clearly over budget. The site also hosts its sister tool, jvqq, which runs jq filters over large JSON Lines files on every CPU core and prints exactly what jq prints, byte for byte. Both are free, open source and ship as one binary each.
The Pitch in Four Numbers
- 0 of 200 runs where nothing changed were called a regression.
- 100 of 100 runs with a 20% or 50% slowdown were caught, on a shared two-core cloud machine whose timings spread 8.9% from run to run.
- 19 to 31 times faster than jq for jvqq across seven filters, using at most 52 MiB of memory on 1 GiB of logs.
- 518 filters, each run five ways through jvqq, with the same output every time.
How jvq Works
The trick is in the alternating. Cloud CI machines are noisy, and timing the base on one run and the change on another mostly measures the noise. jvq runs both on the same runner, back and forth, and only fails a pull request when the difference is clearly bigger than the noise and over the budget you set. Hence zero false alarms in 200 unchanged runs, on a machine with nearly 9% spread.
A budget gate judges each change against its base, which leaves one gap: lots of small slowdowns that each pass on their own. A change that’s 1% slower passes a 5% budget, and so do the next four. jvq keeps a history to catch that drift.
It also has a gate for flaky tests, the ones that pass and fail at random and slowly teach a team to ignore red builds. The history file stays small even for a suite of 2,000 Go tests, and setting it up in GitHub Actions takes three lines of YAML. In one case study, jvq blocked a Windows test bug four pushes in a row.
How jvqq Works
jq is the standard tool for querying JSON, and it runs on one core. jvqq splits a big JSON Lines file across every core, runs the same jq filter, and stitches the results back in order, so the output matches jq byte for byte. The case study is a 1 GiB file of 3,073,492 lines from a 12-minute payments outage: the kind of file you need answers from right now.
Who It’s For
- Teams with benchmarks who want performance to be something a pull request has to pass.
- Anyone fighting flaky tests in CI.
- Engineers digging through big log files in JSON Lines, especially mid-incident.
- Developers who already know jq and want the same filters to run faster.
Limits
jvq’s numbers come from a noisy shared machine, by design, and its examples focus on GitHub Actions. jvqq’s speed figures come from generated logs on a four-core runner against jq 1.7.1, so your mileage depends on your data and filters. Both are young tools that are honest about how they were measured.
The Four-Question Screening
Every project featured here gets the same four questions, the way you’d screen a startup idea.
1. Where is the money in it?
Both tools are free and open source, with no paid tier. Performance and test reliability are things engineering teams already pay for. The paid paths for jvq are a hosted dashboard with performance history across repositories, alerts on slow drift, and flaky-test analytics for larger teams. For jvqq, the money is in log analysis at scale: an enterprise build, support, or embedding it in observability products that run jq-style queries over huge files.
2. What real problem does it solve?
Slowdowns that slip in unmeasured, and CI noise that makes measuring them hard. Plus flaky tests that erode trust in every build. jvqq solves a smaller, sharper one: jq is too slow on gigabyte logs when you need answers now.
3. The technology in plain words
jvq is a referee with a stopwatch who times the old version and the new version one after the other on the same track, again and again, so a gust of wind doesn’t decide the race. If the new version is clearly slower than the rule allows, it blows the whistle. jvqq is a team of readers who split a huge book between them, read their chapters at the same time, and hand back the notes in the right order.
4. Moat and differentiation
Benchmark tools exist, and so do hosted performance services. jvq’s edge is that it works on ordinary noisy CI runners and has the false-alarm numbers to prove it, which is the real reason teams abandon performance gates. jvqq’s edge is exact jq compatibility, which means no new language to learn. The moat that can grow is statistical rigour people can trust, plus history that builds up over time and is painful to give up.
Why It Matters
Speed is a feature users notice and teams rarely protect. A gate that catches slowdowns without crying wolf turns performance from a vague worry into a rule that every change has to pass.
Follow the Project
Read about the jvq gate and the jvqq query tool, start with the quick start, get builds from the download page, or see the code under just-very-quickly on GitHub. jvq is listed with the other open-source projects on SoftDevStudio.com, and more featured projects are in the Projects section.