AnySec
Private Anycast Acceptance Test: Verify Before Go-Live — AnySec Insights
← InsightsInfrastructure · 8 min read

Private Anycast Acceptance Test: Verify Before Go-Live

A built private Anycast network is not an accepted one. The tests to run per point of presence, per failure and per exit term before production traffic.

By AnySec engineering

The short answer

A private Anycast network is accepted when you have proved, per point of presence and in writing, that it announces only the prefixes it should, carries real traffic to your origin, fails over and withdraws when told to, and can be operated and exited by you. Build completion is the provider's milestone; acceptance is yours.

Who this is for

This is for a head of infrastructure, CTO or network lead who has bought or is about to buy a private Anycast network and has been told it is "ready". It sits between two other decisions. Before the build, who holds the ASN, the IP blocks and the exit terms decides whether the network is yours. After acceptance, a live cutover runbook decides how traffic moves onto it. This article covers the gate in the middle: what you check, and what evidence you keep, before you let production depend on it.

Why "the build is finished" is not acceptance

A provider can truthfully say a network is complete when routers are racked, sessions are up and prefixes are visible somewhere on the internet. None of that tells you whether the network behaves as your business needs. Four gaps are common.

  1. Visibility is not reachability. A prefix seen in one public looking glass may still be unreachable from the networks your users, partners or regulators sit on.
  2. Working is not failing correctly. A site that carries traffic on a good day has not shown it can lose a link, an upstream or a health check without dragging users into a black hole.
  3. Announcing is not withdrawing. The first time you need to pull a site out of the route is a bad moment to learn who has the authority, and the tooling, to do it.
  4. Delivered is not exit-ready. If the credentials, route objects and runbooks live only with the provider, the network is operated for you, not by you.

Acceptance exists to close those four gaps before the first real user depends on the answer.

A network engineer's hand ticks a printed acceptance sheet clipped to a rack door in a dark machine room, with one patch cable tagged in lime green
Acceptance is signed site by site, against a sheet you agreed before the build, not against the provider's completion email.

Write the acceptance criteria before the build

Put the criteria in the contract or statement of work. A criterion that is invented after the build tends to be shaped to what already works. Each one should name the thing tested, the vantage point, the evidence and who witnesses. The matrix below is a starting point for a review meeting; adapt it to your platform and keep it free of thresholds you cannot justify from your own baseline.

The acceptance matrix

AreaWhat to proveEvidence to keepRed flag if missing
Prefix and originOnly the agreed prefixes are announced, from the agreed origin ASN, with matching route objects and authorisation recordsRegistry records, route object exports, announcement snapshots from independent vantage pointsPrefixes visible that you did not list, or records that name someone else
Reachability per siteEach point of presence answers from networks that matter to you: your users' major access networks, partners, payment and data providersProbe results per site, with timestamps and source networksOne looking glass offered as proof for every region
Path to originTraffic accepted at a site reaches the origin over the agreed private path, and the origin sees the client address you expectPacket captures or origin logs showing source address handlingThe origin only reachable through a shared or public path
CatchmentUsers land on the site you intended, and you know which sites a given network prefersTraceroutes and measurement results from several vantage points"It should go to the nearest site" with no measurement
FailoverWhen a site loses a link, an upstream or its health signal, traffic moves to another site and sessions recover as designedTimestamped test log with what was removed, what users saw, how long recovery tookFailover described but never exercised
WithdrawalYou can withdraw a site or prefix on demand, and the route disappears everywhere it shouldA witnessed withdrawal and restore, with who had the authorityOnly the provider can do it, and only by ticket
Health-check behaviourA bad check cannot withdraw every site at once; a site that is up is not pulled by a false negativeThe check logic and a test of its failure modesOne shared check able to take the whole edge offline
Capacity and limitsDocumented limits per site for sessions, bandwidth and state, with the behaviour when they are hitProvider specification plus a controlled test inside agreed limitsLimits unknown until an incident
Monitoring and alertingYour team receives alerts for session loss, route changes and unexpected announcementsAlert test records and the named recipientsAlerts go only to the provider
Access and credentialsYour named staff hold the accounts, MFA and change authority you agreedAccess list reviewed and signedShared logins or provider-only control
Exit readinessYou hold the records needed to leave: prefix authorisation, configuration export, contact paths, handover stepsA dry-run handover list reviewed by youExit terms stated but nothing to execute them with

The two rows that most often get skipped are withdrawal and health-check behaviour. They are also the ones behind the outage pattern described in when your own health checks take a crypto exchange fully offline.

Test per point of presence, not once for the network

Anycast makes a network look like one thing. Operationally it is a set of sites, each with its own upstreams, hardware and local failure modes. Accept them individually.

  • Use the same sheet for every site. If one site passes a stricter test than another, record why, and do not let the weaker standard become the default.
  • Keep an unaccepted site out of production. A site that is announced but not yet accepted will attract traffic you did not plan for.
  • Record the order. The list of accepted sites, with dates and witnesses, is also your first inventory for the operating phase.
  • Re-run a short version after any change. A new upstream, a router replacement or a policy edit can quietly alter catchment.

Failure tests: what to break, and how to stay safe

Failure tests are where acceptance earns its cost, and they should be designed so a mistake does not become an outage.

  1. Agree a window and a stop condition. Name the person who can call off the test and the signal that triggers it.
  2. Start on a site that carries no production traffic. Prove the method there before you touch a live site.
  3. Break one thing at a time. Remove a link, then an upstream session, then a health signal. Record what changed in the routing view and what a test user experienced.
  4. Measure from outside. Use probes from networks you do not control, so that what you observe is what users would observe.
  5. Restore and confirm. Bring the thing back and verify that traffic returns to the intended sites, not only that the alarm cleared.

If the provider needs to authorise or perform any step, put that in writing as part of the test plan. Where a test touches third-party networks, check your upstreams' and hosts' current policies first.

What to keep as the acceptance record

The record is the hand-over. A reviewer, an auditor, a successor engineer or a partner should be able to read it and see what was tested, by whom, against which criteria and with what result. Keep:

  1. The signed criteria that existed before the build.
  2. Per-site results with timestamps in one time zone.
  3. Witness names and the dates they signed.
  4. Any criterion waived, with the reason, the risk owner and a review date.
  5. The open items that remain after go-live, each with an owner.

A waived criterion is not a failure of the process. An unrecorded one is.

Red flags in the acceptance conversation

  • The provider proposes to run and sign off the tests themselves.
  • Acceptance is defined as "sessions established and prefixes visible".
  • A failure test is declined with no alternative site or date offered.
  • The criteria change after the build.
  • Credentials, route objects or runbooks cannot be shown to you before go-live.
  • You are asked to sign for the whole network after a test of one site.

Limitations

This article is not legal, regulatory or network-engineering certification advice. It uses no external statistics or incident case studies, and the matrix has no thresholds because the right figures depend on your traffic, your upstreams and your providers' policies. Acceptance testing shows how a tested network behaved on the day under the conditions you created; it does not prove future availability, does not replace a later authorised DDoS stress test and cannot predict how third-party networks will route changes made after the test. Provider behaviour can differ from documentation. Always confirm the current testing and change policies of your upstream and hosting providers before you break anything on purpose.

Decision and next step

If, for every site, you can show the prefixes announced, the path to origin, one failover that recovered, one withdrawal and restore that you could trigger yourself, and the records needed to exit, accept the site. If you cannot, hold it out of production and ask for the missing evidence. To have the criteria specified before a build, or to review a network you have already been handed, request a network design review with AnySec and we will turn your architecture into an acceptance sheet with named evidence and witnesses.

Sources and review

The method draws on public guidance: RFC 4786 Operation of Anycast Services, which covers deployment, failure and monitoring considerations for Anycast, and MANRS, the routing-security norms for filtering, anti-spoofing and coordination, read 2026-11-22; check the current versions and your own upstream providers' policies. AnySec prepared the acceptance matrix from the public structure of network commissioning practice; no AnySec-original statistics are used. Author: AnySec Engineering. Published 2026-11-22; reviewed by the AnySec security team.


Related reading

To have the acceptance criteria written before the build, request a network design review.

Frequently asked questions

What is an acceptance test for a private Anycast network?
A written set of checks, agreed before the build, that proves each point of presence announces the right prefixes, forwards traffic to your origin, fails over when it should, withdraws cleanly when it must, and leaves you able to operate and exit the network. It is separate from the provider's own commissioning tests: you are verifying the outcome against your requirements, not the provider's checklist.
Who should run the acceptance tests?
Your own engineers, or a party you appoint, should witness or run the tests that matter, using vantage points and logs you control. The provider can execute the changes, but a test the provider runs and reports on alone proves only that the provider believes it works.
Can we accept the network one point of presence at a time?
Yes, and it is usually safer. Accept each point of presence against the same written criteria, keep it out of the production announcement until it passes, and record which ones are live. A staged acceptance also makes the rollback path concrete, because every accepted site has already been withdrawn and restored once.
Should the acceptance test include a DDoS test?
Not as part of the routing acceptance. Confirm the routing, failover and withdrawal behaviour first, then run an authorised DDoS stress test against the accepted network as a separate, scoped exercise with its own rules of engagement and provider notifications.
What if the provider refuses a failure test on a live site?
Ask for it on a site that is not yet carrying production traffic, and ask for the reason in writing. A provider that will not let you withdraw a route, or watch a failover, before you depend on it is telling you something about the exit terms and the operating model.

Rather not learn this in production.

Talk to the engineers behind these write-ups — no sales script, a straight read on where you stand.

Get a fixed quote