The eclipse drew a 30% hole in national traffic, and your anomaly detector will not know why

Cloudflare Radar has published a breakdown of HTTP request volume during the total solar eclipse that swept from the North Atlantic across Iceland and northern Spain and Portugal on Wednesday, August 12, with a deep partial eclipse over the rest of Western Europe near local sunset ^[1][22]. The company measured request volume in five-minute buckets and compared it against a same-weekday baseline, and the troughs land under maximum obscuration almost exactly ^[4][9]. That is a clean natural experiment in synchronized demand, and it is also a preview of how badly conventional baselines handle one.

The numbers first. Along the path of totality and in the countries with the deepest partial coverage, Iceland, Ireland, the UK, France, Spain and Portugal, traffic fell in a trough directly beneath peak obscuration ^[10][12]. Cloudflare measured the dip as the average percent change against baseline in the 15-minute window around maximum eclipse, and puts the totality-path decline at roughly 15% to 30% ^[13][14]. Country-level trend lines spanned 9.3% to -46.7%, a spread of 56 percentage points across the affected set ^[15][2]. Where the sun was barely obscured, in Sweden, Denmark, Poland and Switzerland, the effect was all but absent ^[11].

Now the operational part. At five-minute resolution, that 15-minute dip window is three buckets ^[1]. Traffic rebounds typically within minutes of maximum coverage as people return to their screens ^[12]. An alert with a dwell time in that range fires and then self-resolves before anyone has finished reading the page, which is the worst possible outcome: a real, large, correlated deviation that leaves no diagnosable artefact behind.

The shape is worse than the depth. In Cloudflare's heatmap, the rows are countries, one exception being Alaska, ordered by the sequence in which they saw the eclipse ^[7]. Per-region alerting on that same data does not see one event. It sees a fault propagating west to east across a continent, region after region, which is what a routing problem or a bad config rollout looks like to a correlation engine.

Then there is the baseline itself. Cloudflare compared eclipse day against the median of the three previous Wednesdays, matched slot by slot on time of day, and says explicitly that using the median keeps one odd week from skewing the comparison ^[5]. That choice is the tell. Any team baselining on a mean of recent same-weekday windows, or on plain week-over-week, will now carry a -30% trough forward into its own reference series for as long as the lookback runs, and will read next Wednesday's normal evening as a spike.

The obscuration side of the analysis is worth noting because it is the part most operators cannot reproduce. Cloudflare derived coverage from apparent angular sizes of the sun and moon and their separation, taking the geometric overlap of two circles every five minutes, then summed regions for national traffic and averaged regional obscuration to place each country's maximum eclipse time ^[16][17]. That is an external, physically determined covariate. Very few demand shocks come with one.

Cloudflare acknowledges that population density, time of day and cloud cover explain scatter at any given coverage level, while arguing the direction holds and the eclipse was the primary driver ^[18][19]. One caution on the source: the post describes the day as Wednesday, August 12, labels its chart columns August 12, 2025, and cites the 2026 World Cup and the 2024 total eclipse as prior comparable events ^[1][21][20]. The year is internally inconsistent; the mechanism is not.