Methodology

Everything on the overview page comes from one loop doing one simple thing. This page is the whole of it, so the numbers can be read for exactly what they are.

The probe

Every 60 seconds each endpoint gets a single GET over HTTPS, redirects followed, body read and discarded, with a 10-second ceiling on the whole exchange. The measurement is wall time from the start of the request to the last byte of the response — so it includes DNS, the TCP handshake, the TLS handshake and the server's own thinking time. It is not a ping, and it is not comparable to one.

The three states

• up — answered with an HTTP status below 400, in under 1200 ms.
• slow — answered correctly, but took longer than 1200 ms. Usually the path, not the site.
• down — the request failed outright, timed out, or came back 4xx/5xx.

The numbers on each card

Latest is the most recent sample. Median is the middle value of every successful sample still in memory — median rather than mean, because one 9-second outlier should not move the figure that describes a normal minute. Reachable is the share of samples that succeeded at all.

The chart holds the last 60 samples, roughly an hour. A gap in the line is a failed probe; the short red tick on the baseline marks where it happened.

What this does not measure

One machine in one datacenter in Istanbul watching six endpoints. That is a single vantage point, so it cannot distinguish a site being slow from the route to it being slow, and it cannot tell you anything about how the same site behaves from anywhere else. A real monitoring service probes from dozens of locations and takes a vote. This is a bench, not a service.

History lives in memory and a small file beside the process, capped at two hours. There is no database, no alerting, and nothing is retained beyond that window.