Halo-style ingress assumes GeoDNS picks the closest vpshalo PoP for JP, KR, HK, SG, and US West. Laptops hide behind resolvers that may send EDNS Client Subnet hints. This matrix pairs ECS on versus off with TTL layering and artifact p95 gates for one shared sign-off sheet.

Pair with SVCB, ECH, and TLS p95, split-horizon GeoDNS latency, and SCP and SFTP RTT matrices. Public pages: help, home, pricing.

1) ECS can map the resolver instead of the engineer. 2) One aggressive TTL poisons every cache row during cutover. 3) Artifact p95 spikes often mean wrong-city answers, not slow disks.

Probe mode ECS on ECS off
GeoDNS accuracy Authority sees client prefix when policy allows Answers follow resolver anycast; larger roaming deltas
Drift vs traceroute Lower median gap when VPN hairpins are documented Higher tail mismatch; watch JP and US West public resolvers
Privacy Subnet reaches authority; shorten prefix under GDPR Less telemetry; add synthetic probes per PoP

Treat each mismatch as routing until proven otherwise. Log answers, TTL remaining, and resolver ID beside SSH latency so KR and SG diffs stay in one warehouse.

Resolver drift: Geo answers versus traceroute

GeoDNS keys off the wire source. ECS replaces that with the client subnet option. Without it, the source is the recursive. Corporate VPNs add a third story. Run paired probes per office profile and label outer resolver, inner tunnel, and target vpshalo region.

TTL layering: apex, control plane, artifacts

Keep interactive SSH names on short TTL. Use longer TTL on artifact hosts only after poisoning tests pass. Split apex, health CNAMEs, and HTTPS bindings per the SVCB checklist.

Decision matrix: ECS hints versus static probes

Score with security, networking, and release engineering before changing resolver contracts.

Workload Prefer ECS on Prefer ECS off plus probes
Roaming split-tunnel VPN Yes with redacted logs and short prefix masks Yes when subnet disclosure is blocked; use vantage agents
CI on fixed SG NAT Optional when anycast already matches city Default off if one resolver cluster serves all runners
Artifact backhaul to HK Yes for per-office steering inside one country Yes when answers should follow CDN only; pair with RTT matrix
Tip: Compare ECS on versus off before tuning TCP or BBR or you will chase the wrong continent per MSS and BBR.
# Resolver with ECS hint (replace prefix with your office lab net)
dig @203.0.113.53 ssh.builder.internal A +subnet=198.51.100.0/24 +norecurse +noall +answer

# Same question without ECS emulation (no client subnet option)
dig @203.0.113.53 ssh.builder.internal A +norecurse +noall +answer

# Trace delegation while you iterate TTL decay after a cutover
dig @203.0.113.53 ssh.builder.internal A +subnet=198.51.100.128/27 +trace +noall +answer

Regional artifact p95 gates

Five-minute rolling windows after caches warm. Widen sampling only with a filed exception.

PoP City drift bar One GB artifact p95 TTL residual alarm
JP Ninety-five percent probes match traceroute city Under nine minutes unless object storage is explicit Residual TTL over policy by thirty seconds
KR Same bar; watch resolvers anycast outside peninsula Match JP unless fiber proves faster Same thirty second skew rule
HK Allow SG only when ECS proves intent Eight percent tighter than JP when cables share borders Stricter skew during submarine maintenance seasons
SG Log resolver ASN because ECS variance is higher Default JP ceiling unless throughput proves higher Dual resolver policies differ by at most one TTL step
US West Watch Colorado versus Oregon resolver swaps Slightly looser than HK when Pacific paths win Match JP alarm but allow longer health CNAME chains

Six rollout steps

  • 1 Inventory resolvers; tag ECS capable versus stripping paths per VPN.
  • 2 Save paired dig outputs with and without subnet, join to traceroute cities.
  • 3 Publish TTL bands: short for SSH, medium for health, long only for proven artifact names.
  • 4 Replay one gigabyte fixtures from every region through the production DNS path.
  • 5 Plot ECS delta versus artifact p95 on one dashboard line.
  • 6 Sign with security and networking; attach evidence to help tickets for auditors.

FAQ: ECS and cache collisions

Privacy: Redact subnets before logs leave the zone; use country-level prefix when possible. Caches: Mixed ECS and non-ECS resolvers on one anycast IP can collide—split names or partition resolvers until charts stay clean.

Cite in reviews: Ninety-five percent city match before MSS work. Thirty seconds residual TTL skew triggers rollback. One gigabyte stays the minimum fixture.

Disclaimer: Thresholds support staging sign-off, not contractual SLAs.

Green rows mean you can align checkout with probes via purchase and Tokyo, Seoul, Hong Kong, Singapore, or US West.

Nodes · help · home

Lock the PoP that cleared your ECS and artifact matrix

Start from purchase for the Mac mini M4 monthly bundle aligned to the regions in this checklist. Use help for SSH and console onboarding, home for product positioning, and pricing when finance needs list rates.

Regional carts: Tokyo, Seoul, Hong Kong, Singapore, US West. Continue the Halo ingress series on the blog index.

Open purchase summary Help center Product home Pricing Technology blog