Archive status

Why this page exists

Every vendor in this market publishes latency figures. None publishes a per-source uptime you can check. That’s not an oversight — scrapers break weekly when a platform changes something, and a number that admits it is uncomfortable to publish.

The uncomfortable version is the useful one. If you’re going to depend on trend data, the question isn’t how fast the API responds. It’s which sources actually collected today.

What the per-source states mean

• healthy — the last run brought items in, and recent runs mostly did too.
• intermittent — collecting, but failing more than one run in five. This state exists because a source spent a whole night losing roughly half its windows and showed up as fine everywhere: it kept succeeding often enough that no “consecutive failures” rule ever tripped.
• not collecting — three runs in a row without a single item, or nothing successful for over three hours.
• never covered — this source has never produced a row. Nothing broke; there is nothing to recover.

Switched off

These sources are not running at all. They are not broken and they are not late — they were turned off, and the reason is below. None of them appears in the health table above, because a red label next to a deliberate choice is an accusation, not a status.

• Bluesky — Switched off for collection. Bluesky's search API returns 403 ("Request forbidden by administrative rules") to our datacenter IP on the very first call, not after a burst. Through a residential proxy it succeeds 8 times out of 15, at random — which would produce a partial window every hour, and a partial window is not comparable to a full one. Measured 2026-09-09. The bluesky/profile API endpoint is unaffected and still live.
• Facebook — Not collected, by choice.
• Instagram — Switched off. Collection here depends on a third-party provider we have not turned on.
• LinkedIn — Not collected, by choice.
• Mastodon — Not collected. Post search needs an authenticated token, so it is not free and account-less like our other sources.
• Packagist (PHP) — Not collected. Only lifetime download totals are exposed, and a number that barely moves is not a trend.
• Pinterest — Switched off. Collection here depends on a third-party provider we have not turned on.
• Reddit — Switched off. Reddit's free API tier is described as non-commercial, and this product charges. We stopped rather than read under terms we cannot meet.
• Spotify — Not collected. Spotify's developer terms forbid selling or passing on data obtained from their API, including derived and aggregated data.
• Stack Overflow — Not collected. Question volume on these subjects is too low to form a time series.
• TikTok — Switched off. Collection here depends on a third-party provider we have not turned on.
• Twitch — Not collected. Twitch's developer agreement forbids commercially distributing data from their API, including derived insight.
• X / Twitter — Switched off. Collection here depends on a third-party provider we have not turned on.

Removed on purpose

These source-and-niche pairs are not collected, and that is a decision — not a failure and not a gap. Nothing here is waiting to be recovered.

• lemmy × ai-agents — Lemmy search is full text over a general news aggregator. Measured on 2026-09-09: the term “ai agent” returned 30 posts and 7 of them were US politics, unrelated to AI agents. Requiring the term to appear in the title would leave 2 of those 30. It does not serve this niche, and removing it is the honest answer.

Rows already archived from these pairs are kept. Each one is true — the post exists and has those votes; what was wrong was publishing it under that niche. Deleting would destroy real history and would not stop it happening again.

What the per-niche states mean

• collecting — captured within the last 90 minutes.
• delayed — 1.5 to 3 hours. Usually the scheduler: GitHub’s cron is best-effort and slips under load.
• stale — nothing recorded for over 3 hours. Either the collector broke, or the source had nothing to say. This page can’t tell those apart, and says so rather than guessing.
• no data — nothing recorded yet, or the source was reset.

Three states, not two

A window can be complete, a partial sample, or not collected, and the middle one is the dangerous one. Missing data announces itself. A partial sample looks exactly like a good window: real rows, real scores, a full-looking ranking — drawn from a narrower set of searches than the windows around it.

It happens when a source starts refusing part of a run. One of ours returns errors in bursts after the first few requests, with no header saying so, and the collector treated a refusal the same as “nothing matched”. The early searches landed, the later ones vanished, and the window was written as if whole.

That matters most for anything comparing two windows: a rise measured between a full window and a partial one is not a rise, it is the difference between two pools. Those windows are listed above rather than quietly deleted, because the rows are genuine — they just can’t carry the weight of a comparison.

A source can be up and still be wrong

This page reports whether data arrived, not whether it was right. Those are different failures, and the second is harder.

One of these sources spent two hours collecting stories about an e-ink bike computer because a single overly broad search term dominated the results. Every counter looked healthy the whole time. Those windows were deleted rather than kept, because a contaminated baseline poisons every comparison made against it afterwards.

A green row means the pipe is open. It does not mean the water is clean.

An empty source used to look exactly like a broken one

Until 7 September 2026 this page recorded only success. A source that failed wrote no row, and a source that ran and found nothing also wrote no row — so “we didn’t collect” and “we collected and there was nothing” were the same absence. Three sources sat dead for roughly fifteen hours behind that ambiguity, and the archive recorded it as quiet.

Every attempt is now recorded, whether or not it produced anything, so the two are distinguishable going forward. The per-niche table above still can’t tell them apart on its own — Indie SaaS is collected mostly from Hacker News, which barely discusses it, so that niche goes quiet legitimately. The per-source table is where the difference shows.

This does not work backwards. For anything before that date there is no record of attempts, only of results, and no amount of querying recovers what was never written down. Intervals we can positively identify as uncollected are listed under Known gaps; the rest of the early archive carries an ambiguity we can describe but not remove.