CONCEPT Cited by 1 source
DNS routing policy at scale¶
Definition¶
DNS routing policy at scale refers to the architectural insight that the choice of DNS routing policy becomes a load-bearing decision at CDN/edge scale — particularly when combined with persistent-connection protocols. At normal scale, routing policies (Weighted, Latency, Multi-Value Answer, Failover) produce similar operational outcomes. At extreme scale (millions of simultaneous connections, CDN edge fleets with many resolving nodes), the difference between returning 1 IP per query (Weighted) vs up to 8 IPs per query (Multi-Value Answer) determines whether traffic distributes or concentrates.
The bitdrift lesson¶
The canonical wiki production instance (Source: sources/2026-07-15-aws-bitdrift-scaled-121-million-grpc-connections-cloudfront):
| Policy | IPs per response | Behavior at 121M devices |
|---|---|---|
| Weighted | 1 | All CloudFront edges resolve to same NLB per TTL → 80% 5xx |
| Multi-Value Answer | Up to 8 | Edges spread connections across all NLBs immediately → 0% 5xx |
The key insight: adding more origins (2 → 6 NLBs) had no observable effect under Weighted routing because the policy still returned only 1 IP per query. The problem was not capacity — it was the routing policy's 1-IP-per-response semantic.
When this matters¶
This concept is specifically relevant when: 1. A CDN or edge fleet resolves your origin domain (many resolving nodes → all converge on the same answer). 2. The protocol uses persistent connections (gRPC, WebSocket) that accumulate rather than redistribute. 3. Traffic surges are sudden (live events, viral content) rather than gradual.
For short-lived HTTP behind a CDN, Weighted routing's single-IP response is usually fine because connections churn before the concentration becomes dangerous. See concepts/persistent-connection-amplification for why persistent protocols change this calculus.
Seen in¶
- sources/2026-07-15-aws-bitdrift-scaled-121-million-grpc-connections-cloudfront — the canonical production incident demonstrating that routing policy selection is architecturally load-bearing at extreme scale.