... returning HTTP 200 with an empty list, and from the outside it simply looked like they had stopped hiring. Nothing alerted, because technically nothing had failed. I ended up alerting on the shape of the data instead of the statuscode: if a source drops more than half its rows overnight, a human looks at it before anything downstream moves. That threshold is a guess, and it still misses the slow rot, where a site changes one selector ...

... secrets. The failure only exists on a server whose config file is one change behind the code. The fix is three lines: if the secret is empty, refuse before verifying. The part that took thought was the status

... Just turn off prerendering is not a safe move on its own, and I would rather say that out loud than let someone repeat it. The part that bothers me most is not the bug. It is that a statuscode check was never going to catch it. The process returned a fast, clean, completely wrong 200, with another company's name in the title and a canonical pointing at localhost. If a crawler happened to arrive inside that window ... ... monitoring only really start once things settle? And does your health check assert anything about the content of the response, a title, a canonical, an identifier that proves it came from your database, or does it only check the

... reporting what we found, not a turnaround. The check itself is not clever. Fetch your ugliest URLs with Googlebot's user agent using curl, pass the Googlebot user agent string with the -A flag, and read the statuscode back with the write-out flag. Do not start with your homepage. It is the page most likely to be healthy, so it tells you the least. Run it against a paginated listing, a URL with a query string ... ... competing with each other instead of continuing to assume it. 13 posts, 78 pairs, similarity threshold 0.363, highest observed 0.300. The assumption was wrong. If you have ever run a Googlebot user agent fetch against your ugliest URLs, what

... dead rows in our own catalogue rather than for anything we had done. The control gave up something better than a verdict though. The two responses differ on an axis neither the detector nor I was looking at: the statuscode. 200 against 404. Our fetch path calls goto() and throws the Response object away, so every downstream decision was made from the HTML alone, which is the one surface where these two cases genuinely do look alike. The distinguishing ... ... reframes what I had already shipped. I built a contextual guard: if this node succeeded within the last two minutes, an error page is a dead ID rather than a block. It works, but it infers from timing. The