DDIENGINEER.COM / FICTIONAL NETWORKS. REAL TROUBLESHOOTING.NO PRODUCTION SYSTEMS WERE REBOOTED.
⌕ PACKET DETECTIVE
DNS + BGP / ADVANCED

Anycast DNS troubleshooting: BGP service-health withdrawal

One office gets DNS timeouts while another resolves normally through the same Anycast address. Both sites still advertise it. The routers would like to be excluded from the incident.

What you’ll practise

Trace the supplied evidence, challenge alternative explanations, and choose a safe next action. Use the topology and private notebook to document your reasoning.

Lab scope: Two internal Anycast resolver instances announce 10.200.0.53/32 using BGP. Clients at each office prefer the nearer site. No external prefix propagation is permitted.

Validation: Original synthetic operational lab; not a reproduced vendor defect

DDIENGINEER.COM / INTERACTIVE NETWORK INVESTIGATIONS

DHCP says yes.
DNS says yesterday.

A laptop moved networks. Its name didn’t. Find out who’s telling the truth — and what you should check next.

ADVANCED · DNS + DHCP · ~8 MIN

Realistic DNS, DHCP and IPAM mysteries.
No lab setup. No production access. Just your judgement.

01 Inspect the clues02 Build your theory03 Make the call
FEATURED / PD–009DNS + DHCP
THE LAPTOP WITH YESTERDAY’S ADDRESS

New lease.
Old address.
Same confused user.

09:30 / DHCP ACK10.62.4.27 ✓↓ six minutes and three cache flushes later09:36 / DNS ANSWER10.61.4.27 ?

“Have you tried flushing it again?”
The fourth time is not a strategy.

FIELD NOTES / READ BEFORE OR AFTER THE LAB

Troubleshooting guide

Recognise the symptoms

One office gets DNS timeouts while another resolves normally through the same Anycast address. Both sites still advertise it. The routers would like to be excluded from the incident.

Checks to compare

  1. Identify the serving paths. Compare what this check establishes with the other devices in the path.
  2. Compare service and route health. Compare what this check establishes with the other devices in the path.
  3. Test controlled withdrawal. Compare what this check establishes with the other devices in the path.

Scope matters. Two internal Anycast resolver instances announce 10.200.0.53/32 using BGP. Clients at each office prefer the nearer site. No external prefix propagation is permitted.

Reveal the diagnosis, common mistakes and next action

What the evidence supports

Per-node queries establish a failed DNS service at E while BGP still sends clients there. Controlled withdrawal restores the affected path without changing zone data.

Common mistakes to avoid in this scenario

  • The shared address is unreachable: Remove both sites’ BGP advertisements because DNS fails in East.
  • East clients have a stale DNS cache: Flush client caches; leave the service advertisement unchanged.

A proportionate next action

Withdraw only the failed instance’s service route using the approved procedure; add service-aware checks and recovery hysteresis.

Limits of the conclusion

Anycast is not inherently self-healing. This case supplies path and node evidence; route changes require capacity and convergence validation.

Read the technical reference →

Browse all investigations →

Browse troubleshooting case guides →The missing networkThe gap in the alibiThe record that refused to existThe name with two addressesThe office with no vacanciesThe relay to nowhereThe duplicate that wasn’tThe suspiciously quiet subnetThe laptop with yesterday’s addressThe security upgrade that broke renewalsThe container that couldn’t find anythingThe lease that rewrote historyThe intermittent DHCP lotteryThe edit that never reached DNSThe policy inherited from upstairsThe updater with the wrong badgeThe fixed address with no nameThe duplicate in another network viewThe update at the wrong front doorThe lease with no permission slipThe reservation the partner forgotThe immortal DNS recordThe pipeline that retried yesterdayThe route that outlived its resolverThe adjacency with no routeThe domain controller nobody could findThe backup with a missing keyThe green dashboard with the wrong answerThe two masters of one recordThe zone transfer with no guest listThe inventory that stopped at page oneThe allocation raceThe redundant pair on one plugThe answer too large for the doorThe IPv6 address with nowhere to go