Every status page forgets.

A count is how often a provider announced something, not how often it broke. A company that publishes every degradation will out-count one that publishes nothing, so this ranks disclosure as much as reliability. Never read a low number as a good one. Time with a record open is not downtime: an incident record usually covers one component or one region while everything else keeps serving, and it stays open until the provider closes it. Where the resolved figure carries a second number, the difference is rows kept from before this read the providers’ own JSON, which have no status on record either way.

Announced nothing in the window: Atlassian. Read that as published, not as proven.

Every figure here is built from two timestamps the provider published. Nothing is measured independently, and nothing here is a reliability rating.

Titles are the provider’s own wording, linking to their own write-up. Length is the gap between the start and the resolution they published, on the 768 of 850 incidents in this window where they published both — it is not a measure of how long anything was broken.

Counted from one Hacker News hiring thread a month — 386 posts, skewed hard toward American startups, and evidence about that population and no wider one. A post counts once however often it names a thing. Terms marked "floor" collide with ordinary English and are matched only in unmistakably technical context, so their number is a minimum rather than a count.

Asked for more often than last month

Asked for less often than last month