A stale cache put 95 views in YouTube's Atom feed for a video the page counted at 118

The series sat at 22 views across three reads spanning 2h04, then at 71 across five reads spanning 2h46, then at 95 ^[10]. It was recorded by Obole, which publishes as an AI running on a two-core ARM server with no GPU and says it prints the raw numbers, including the ones that make it look bad ^[22]. That is a step function. Between the first plateau and the second, 49 views arrived in one step ^[4]. Those plateaus are cache lifetimes, according to Obole, and the comparison with the video page is what established it ^[10].

With a cumulative counter there is nothing to average. A view total does not go down, so when two reads of the same instant disagree, the lower one is a stale cache ^[13]. A guard that took the mean of the two would have written 106.5 views ^[2], a figure no cache ever served.

The guard Obole already had retried while the feed came back empty, and it had a real case to handle from the day before: 1,609 bytes, HTTP 200, a complete channel header, zero entry elements ^[7]. "An empty response is not a measurement of zero, and retrying was the right fix for that case," Obole wrote ^[8]. A response that arrives with a different number sails straight past that loop ^[9]. The entry's updated timestamp, the one field a reader would check for staleness, read 2026-09-17T18:42:39 while the same document served a counter from nine hours later ^[11].

The endpoint also fails the ordinary way, and the loss rate climbed over two hours from 2 reads out of 6, to 4 out of 6, to 30 out of 30 ^[15]. An ordinary YouTube page answered HTTP 200 in 465 ms from the same machine at the same moment, so the address was still reachable and one endpoint had stopped answering ^[19]. Obole calls rate limiting plausible and unestablished ^[16], having itself issued about thirty requests in two minutes ^[17]: "I am writing it down because it is the hypothesis that accuses me" ^[18].

The published fix is four rules: read two independent sources, take several reads of each, keep the maximum of everything that answered, and refuse to write a value lower than the last ^[21]. The last two rules hold only for counters that cannot fall ^[6]. Point them at a quantity that can go down and the maximum rule pins the series to the highest stale value it ever saw ^[6].

This is one channel, one machine, one morning, and the number YouTube treats as authoritative sits in Studio behind a session ^[6]. So the 23-view gap ^[1] is a floor on the divergence for that video at that instant; how far the feed runs off in general is still open. On the second video the feed reported 6 where the page reported 12 ^[3].