Automated canary analysis, live

Every deployment gate below is real. Move the sliders to change what the candidate build does, and the same statistical judge that runs in the pipeline decides — right here, in your browser — whether it reaches production.

Mann–Whitney U Cliff's delta Weighted judgment Red/black promotion Progressive delivery

The judge

A canary is only trustworthy if it catches real regressions and refuses to cry wolf. Try the presets: a clean build promotes, a 4% error rate fails on a critical metric alone, and a 3% latency drift is caught as significant but deliberately tolerated.

Baseline holds at 0.2%. This metric is marked critical.
Baseline holds at 18 ms.
How large a difference must be before it counts. At 0, any statistically significant change fails — which is how canaries get switched off.
More samples detect smaller differences. Watch p-values fall as you raise this.
MetricResult Canary Baseline p-value Cliff's δ Weight

Latency distribution

Baseline Canary
Right-skewed, which is why the judge uses a rank-based test rather than comparing means. A long tail moves an average around long before it means anything.

What the pipeline does with that verdict

The judgment above is not advisory. It decides whether traffic moves, and a rejection halts promotion before the next environment is touched.

Why it is built this way

Significance is not the same as importance

With enough samples, a 0.3 ms latency increase is statistically significant and operationally meaningless. A gate that fails on significance alone gets disabled within a month. Every metric must clear both a p-value and an effect-size threshold.

One critical metric outranks the score

The weighted score can be dragged up by healthy metrics. An error-rate regression fails the canary on its own, regardless of how good everything else looks.

Missing data is not good news

A metric that reported nothing is excluded from the score and disclosed — never counted as a pass. Treating silence as health is how bad builds reach production.

Rollback is a pointer swap

Red/black keeps the previous version running and healthy while the canary is judged. If the canary fails, nothing needs to be redeployed under pressure — traffic simply never moved.