TCP latency / RTT (p99 budgets)
Latency is the round-trip time for a request: the customer asks, your servers answer, the answer travels back. Every customer on a Silver, Gold, or Platinum plan buys a promise about that number — specifically about their slowest requests, not their average one. Cross the line and you owe them money back as an SLA credit.
Three things spend latency. Distance: a customer served from a far-away region pays for every mile. Congestion: an oversubscribed switch or a saturated uplink makes everyone behind it wait in line. Contention: a noisy neighbour hammering shared disks drags the whole host down. The painful part is that it's the slowest 1% that breaches — so a service that looks fine on average can still be quietly failing its tail.
Detailed explanation
Silver = 1000ms, Gold = 500ms, Platinum = 250ms p99. Inter-region RTT baseline is 60ms plus ~20ms per id step between source and destination, and the peering kind applies a multiplier on top (transit is worse than a direct peer). p99 is the canonical figure on purpose: averages hide tail-latency contention, and the tail is what the customer feels and what the contract measures.
A sustained tail-latency excursion auto-opens a CustomerObservation incident — the engine surfaces it in #incidents rather than waiting for you to notice the dashboard. Root causes cluster: IOPS contention on shared SSDs (noisy neighbour), uplink saturation under DDoS, oversubscribed CPU during the diurnal peak. Fix the cause; adding capacity only masks the symptom for a tick.