A regression opens only after degradation across two consecutive 24-hour windows with at least 75 valid samples in each
A regression alert is deliberately hard to open. It needs a sustained degradation confirmed by enough samples on two separate days, which is why a healthy site usually shows nothing at all — and why the rare alert is worth reading properly instead of dismissing.
1. Read the regression header#
Scroll to the regression panel. Its header reads either No sustained field regressions detected or {n} sustained field regression(s), and the sub-line states the bar an alert must clear: Alerts require 75 valid samples in each window and degradation across two consecutive 24-hour windows. When alerts have been acknowledged, a count of {n} acknowledged appears beside the header. Resolved alerts drop off this list; acknowledged ones stay visible until they resolve.
2. Read an alert row#
Each row names the metric and the percentile in short form — LCP p75, INP p75, CLS p75, or TTFB p75 — followed by its status and the configuration v{n} the regression is attributed to, with the current value set large on the right. The line beneath gives the comparison directly: the Baseline value, the percentage increase, and the correlation label. Correlation prefers a change on your side — Recent configuration change, Recent origin change, or, when neither is found inside the window, No matching configuration or origin event, which is itself useful because it points away from a release.
3. Understand the floors#
An alert opens only when both recent 24-hour windows degrade against the preceding baseline, and each degradation must clear both a relative and an absolute floor. The floors are per metric: LCP at 20% and 250 ms, INP at 20% and 50 ms, CLS at 25% and 20 milli-points, and TTFB at 20% and 100 ms. A site with thin traffic will often not reach 75 samples a day and will therefore never raise an alert — that is the floor doing its job, not a broken report.
Expected result#
You can tell whether a field regression is open, which metric and percentile it concerns, how far it has moved from baseline, and which configuration version it is attributed to — before you decide whether it matters.
This panel is a report, not an action. Nothing here changes traffic, rolls back, or restarts anything. A regression that does matter is what opens the guarded workflow on Post-promotion validation, and any traffic response there stays a manual decision made by a person.