The arithmetic
If a failure begins one second after a passing check, it stays invisible until the next one. Your worst-case detection time equals your interval, and your average is roughly half of it.
A five-minute interval means an average of two and a half minutes of undetected outage on every incident, before anyone can begin responding. Over a year of incidents, that adds up to real money.
What faster costs you
Halving the interval doubles the request volume against your own infrastructure. Multiply by monitor count and probe regions and the number grows faster than people expect.
Ten monitors checked every thirty seconds from five regions is a hundred requests per minute, continuously, forever. If those endpoints are expensive to serve, you have added meaningful baseline load — and you have added it in the place that hurts most, since it does not stop when you are already struggling.
The reason to keep health endpoints lightweight is not elegance. It is that a cheap endpoint can be checked ten times more often for the same cost, and interval is what determines detection time.
Tier your endpoints
A single global interval is almost always wrong in both directions — too slow for the endpoints that matter, too fast for the ones that do not.
- Tier 1 — revenue pathCheckout, payment callbacks, login. 30 seconds or faster. The cost of a minute of downtime dwarfs the cost of checking.
- Tier 2 — core servicesPrimary APIs and anything other systems depend on. 60 seconds is a good default.
- Tier 3 — supportingMarketing site, docs, blog. 5 minutes. Important, not transactional.
- Tier 4 — internalAdmin tools, batch dashboards. 15 to 30 minutes. Nobody needs paging at 3am for these.
Confirmation adds a little back
Cross-region confirmation costs a few extra seconds before an alert fires, because the system verifies from other locations before declaring an outage.
That trade is almost always worth taking. A handful of seconds of additional detection time in exchange for eliminating most false pages is one of the better bargains in operations.
Do not forget the human side
Detecting an outage in ten seconds accomplishes nothing if the alert goes to a channel nobody watches overnight, or if acknowledgement typically takes twenty minutes.
Match your interval investment to your response capability. If your realistic response time is fifteen minutes, moving from sixty-second to ten-second checks changes very little. Fix the slower half of the equation first.