Opscraft
Route 53 routing policies · Lesson 3 of 10 · 3 min read

Weighted routing

How weights work

Each record gets a weight from 0 to 255. The probability a record is returned = its weight ÷ the sum of weights of healthy records.

resource "aws_route53_record" "blue" {
  zone_id        = var.zone_id
  name           = "app.example.com"
  type           = "A"
  set_identifier = "blue"
  weighted_routing_policy { weight = 90 }
  alias {
    name                   = aws_lb.blue.dns_name
    zone_id                = aws_lb.blue.zone_id
    evaluate_target_health = true
  }
}
resource "aws_route53_record" "green" {
  zone_id        = var.zone_id
  name           = "app.example.com"
  type           = "A"
  set_identifier = "green"
  weighted_routing_policy { weight = 10 }
  alias {
    name                   = aws_lb.green.dns_name
    zone_id                = aws_lb.green.zone_id
    evaluate_target_health = true
  }
}
  • Weight 0 stops traffic to a record — unless all records are 0, in which case they're returned with equal probability. Useful to keep a standby defined but idle.
  • Unhealthy records are removed from the calculation (their share goes to healthy ones).
Enlarge diagramstep 1old stack 90new 10step 2old stack 50new 50step 3new 100Weight = share of answers; weight 0 stops traffic to a record (unless every record is 0).
Figure 1. Weighted routing for a gradual migration: shift traffic step by step, watching metrics at each stage.
Swipe sideways to see all of it, or tap to open it full screen.

Uses

  • Gradual migration to a new stack/region/account: 90/10 → 50/50 → 0/100, watching metrics at each step.
  • Blue-green at the DNS layer (0/100 flip) — with the caveat that DNS caching makes the switch slow and uneven.
  • A/B testing of whole stacks.
  • Spreading load across multiple regions/ALBs in fixed proportions.

Why weighted DNS is imprecise for canaries

Resolvers cache one answer and serve it to all their users for the TTL. A few large resolvers (ISP resolvers, 8.8.8.8, corporate resolvers) represent huge user populations, so the real traffic split can be far from 95/5 and moves in lumps. Some clients ignore TTL. Rollback also waits on caches. For precise canaries, split at the ALB (weighted target groups), ingress controller, or service mesh; use weighted DNS for coarse, slow shifts between whole stacks or regions.

Enlarge diagramlarge ISP resolvercached the "canary"answer200,000 usersall go to the canary for60 sUse the load balancerinsteadweighted target groups ora mesh split requestsprecisely
Figure 2. Why DNS weights are imprecise for canaries: a big resolver caches one answer and hands it to all its users for the whole TTL.
Swipe sideways to see all of it, or tap to open it full screen.