Summary

  • WorldLink’s APNIC guest post reports more than one million IPoE subscribers enabled for IPv6 and roughly 70% of traffic over IPv6; it gives no measurement window or independent audit.
  • The post attributes an outage loop to DHCPv6 renewal and BNG binding behavior. Its 15,467 Solicit/s is a projection from one CPE, not a network-wide observation.

Three figures, three kinds of evidence

A network can hold a large IPv6 allocation, provision IPv6 to every subscriber and still have a session-recovery problem. Allocation describes address space; enablement describes configuration; a traffic share describes use over a period and denominator; continuity asks whether a household stays connected when a link fails. The public evidence for each is not interchangeable.

Kijush Maharjan, identified by APNIC Blog as WorldLink Communications’ Principal Network Engineer, says the operator began IPv6 testing in May 2016 and moved it into production in July 2019. His 12 October 2026 guest post reports more than one million IPoE subscribers, 100% IPv6 enablement and around 70% of traffic over IPv6. It gives no measurement window, denominator behind “traffic,” method for classifying subscribers as enabled or independent validation. These are operator-reported figures, not a service-level audit.

The post also says WorldLink received an IPv6 /32 from APNIC. A /32 contains 16,777,216 /56-sized blocks as address arithmetic; that capacity says nothing about how many are assigned, active or reachable. Nor does the allocation explain the incident. APNIC is the registry named in the resource account; the described sessions ran on WorldLink’s access and gateway equipment. APNIC Whois associates AS17501 with WorldLink in Kathmandu, but that record does not verify the /32 claim.

When a lease becomes a session boundary

WorldLink says DHCP and DHCPv6 leases serve as session state in its IPoE design where a PPP keepalive is absent. A fibre cut above an optical line terminal may therefore interrupt the customer path without a corresponding failure signal at the broadband network gateway (BNG). When the client’s T1 renewal and T2 rebinding go unanswered, lease expiry can lead customer-premises equipment (CPE) to drop its address and delegated prefix, then begin a new Solicit.

Maharjan says one CPE issued 116 DHCPv6 Solicit messages in a minute. The post multiplies that rate by 8,000 CPEs—10% of an 80,000-subscriber BNG—and estimates about 15,467 messages per second. The calculation is 116 divided by 60, multiplied by 8,000. It is a projection based on one reported device-minute, not an observed aggregate. Device behavior, timing and the number of affected clients would determine whether it resembles a real load.

The author attributes a second loop to “Delete Binding on Renegotiation” (DBOR). A new Solicit encounters an existing binding; the BNG removes and rebuilds session state; the CPE continues using an address or prefix it believes it holds; a later renewal returns NoAddrsAvail or NoPrefixAvail. The post reports IPv6 loss, increased CGNAT use and complaints about slow content. It supplies no packet capture, vendor or firmware version, exact configuration, affected-client count, duration or reproduction package; details were generalized for security and confidentiality.

That is enough to frame an investigation, not to identify a universal protocol defect or vendor nonconformance.

Standards and measurement boundaries

The guest post discusses RFC 8415, but it is no longer the current DHCPv6 baseline. RFC 9915, published in January 2026, obsoletes it. RFC 9915’s customer-edge model generally assumes one set of transactions for address assignment and prefix delegation; it says a client SHOULD use a single transaction for the IA options needed on an interface and directs servers to provide T1 and T2 values that support it. Earlier RFC 7550 explains how separate state machines for IA_NA and IA_PD can add traffic and allow state to diverge; it recommends a single session for common cases while recognizing exceptions.

Maharjan reads RFC 8415 as permitting an independent state machine per identity association. That is his interpretation of the earlier text. Without a CPE/BNG trace, the post does not establish which transaction behavior occurred, whether an implementation followed current guidance or whether DBOR interacted with a device-specific retry sequence. Standards frame the questions; they do not replace the missing trace.

APNIC Labs measures a different object: its published test checks whether end-user browsers can retrieve IPv4 and IPv6 objects from its vantage points, with results reported by economy and origin AS. It does not inspect WorldLink’s internal traffic counters, CPE lease timers or BNG binding database. A reachability test cannot independently verify the roughly 70% traffic figure, just as an operator traffic share cannot show that every subscriber stays reachable during renewal.

Sources