Zone-local traffic and the cross-zone bill

Spreading across three availability zones is billed twice: in standing idle capacity, and at a cent per gigabyte for every crossing. This guide reconstructs, from Slack, DoorDash, Grab, Grafana Labs, Uber, Polar Signals and AWS's own design documents, the five-layer stack every zone-affinity implementation assembles, which layer the cost-driven versions leave out, and what the issue trackers show going wrong. A reader finishes able to decide whether to make a given hop zone-local, what to instrument so the optimisation cannot silently switch itself off, and what the recovery path costs.

What you get out of it

• Zone-local routing was published first as a reliability pattern, not a cost feature: AWS's Builders' Library derives it from survival odds, (2/3)^N for N regional hops against a constant 2/3 for zonal ones, and mentions the money once at the end.
• Every published implementation assembles the same five layers, and the cost-driven ones stop at four. The missing layer is the drain, the ability to move all traffic out of a zone in minutes, which Slack budgets at five minutes with 1% granularity and Kubernetes Topology Aware Routing does not have at all.
• The optimisation disables itself silently. Scaling nine replicas to eleven turns [3,3,3] into [4,4,3] and removes every hint, and there is no default alert on the optimisation being off, only on the outcome.
• Same-zone routing distributes by endpoint count, not by load. Envoy's maintainers state it plainly in their own 2026 proposal: it 'has no visibility into actual backend load'. Grab needed extra brokers in the zone their consumers happened to cluster in.
• The headline cloud outages in this corpus were regional, not zonal, so multi-AZ bought nothing against them. Uber measured the event it was provisioning against at 0.23% of the year and cut standing capacity from 2x to 1.3x on the strength of it.

Scope

Why this, now. Three things converged inside fifteen months: Kubernetes renamed and re-scoped its zone-routing option, Kafka is choosing between three competing redesigns whose stated motivation is this one line item, and AWS lost a single availability zone in us-east-1 to a thermal event on 7 May 2026.

What it does not cover. Cross-region traffic and multi-region disaster recovery, internet egress and CDN economics, quorum and replica placement for consensus latency, data residency constraints, and on-premises rack topology.