Data health

Cross-checking against ClickHouse's public github_events dataset proves nothing: it is a copy of GH Archive, so it shares the same upstream and would agree with a broken source just as readily as with a working one.

Sampling GitHub's own public events API is independent, and it disagrees. In a 296-event sample it still returned PullRequestEvent at 29.4% of all events, IssuesEvent at 14.9% and WatchEvent at 8.1%. For the same hour, GH Archive recorded those three at 0.22%, 0.08% and 0.06% — under-reporting them by 112 to 179 times. GitHub is still emitting the events; the archive is not recording them.

The likely mechanism, and this part is inference rather than measurement: the public events endpoint only exposes roughly the 300 most recent events. Push volume grew about 45% across this window, so a poller whose cadence did not keep pace would have the lower-frequency types crowded out between polls.

The loss signature supports that reading. Four unrelated event types fall by near-identical amounts, which is what sampling loss looks like — a per-type policy change at GitHub would have to hit stars, pull requests, issues and forks to within a couple of percent of each other, and would not spare pushes.

Pushes and contributor counts stay trustworthy across the whole window — pushes are the series that rose, and the distinct-account counts derived from them rise with it. Every ranking that has to hold end to end is built on those two.

Star, pull request, issue and fork trends only mean something early in the window. The Rising repos page therefore has a movable “Ranked as of” date: move it into the first half of the window for those metrics, or switch to Pushes or Contributors and leave it where it is. Pull requests are not a safe fallback for stars — they were under-captured just as badly.

Live popularity had to come from somewhere else. That is why Rising stars exists: it ranks on real star counts snapshotted from GitHub's API rather than on WatchEvent, because the event stream can no longer answer the question.

Payloads were truncated as well as dropped, which is why some columns are simply absent rather than always zero: pushes are countable but commits are not, and a pull request carries no title, merged flag or diff size in the archived form.