Summary

  • RFC 9947 defines an experimental SRH AltMark TLV for a controlled SRv6 domain. It gives a packet a bounded measurement vocabulary; it does not make an Internet standard, authenticate a measurement, or certify a deployed service.
  • An L or D flag, FlowMonID, timestamp or sequence number becomes useful only after the operator reconciles domain admission, SID-function support, marking periods, identifier collisions, collection points and the separately observed service outcome.
  • The practical question is not “did a marked packet appear?” but “which administrator configured which measurement scope, which participating endpoints processed it, and which independent evidence supports the result now being claimed?”

The experiment moved a field, not the burden of proof

Alternate Marking separates consecutive packet batches so an operator can derive loss, delay and jitter observations from live traffic. RFC 9341 describes that general method. RFC 9343 supplies an IPv6 encoding using option headers. RFC 9947 proposes another location: an experimental Type-Length-Value item in the SRH.

That relocation can matter operationally. The RFC observes that SR nodes are expected to parse and process the SRH quickly, while some implementations may not handle Destination Options, may not handle multiple extension headers, or may process an option header on a slow path. But the RFC treats those statements as an experimental question, not as a result. It asks participants to test survivability, device processing and forwarding-plane effect against the existing alternative.

This distinction protects an operator from a familiar measurement error. A design rationale is not a benchmark. A packet format is not a capability inventory. A published RFC is not a declaration that a particular router, silicon path, policy, service chain or customer route accepts the construct. Running-Code Primacy begins here: the evidence must come from the chosen path and the implementation actually carrying it, not from a document whose job is to define an experiment.

A controlled domain is a boundary, not a global passport

RFC 9947 says that the Alternate-Marking experiment MUST be confined to a controlled SR domain. It calls for a coordinated value from the experimental SRH-TLV range, and says packets with the AltMark TLV must be kept from entering or leaving the domain. The RFC also records no IANA action. Those facts have a simple implication: a locally coordinated experiment can be precise within its own administrative boundary without becoming a globally assigned protocol fact.

The boundary is security-relevant as well as administrative. A domain operator selects the nodes, controls admission and should prevent injected or escaping marked packets from contaminating the observation. But “trusted domain” is not the same as “trustworthy conclusion.” A controlled perimeter cannot prove that a collector selected the intended period, that an identifier remained unambiguous, that every planned endpoint had the required SID function, or that a route that was valid at 09:00 remained the customer’s route at 09:15.

The proper first record is therefore not a dashboard percentile. It is a measurement charter: domain owner; experimental TLV value; ingress and egress boundary; SR policy revision; intended segment list; participating SID functions; flow-selection rule; period; collector identity; and explicit statement of what leaves the experiment unproved. That small ledger gives later readers a way to distinguish a repeatable observation from a colorful packet capture.

The marker identifies a candidate, not necessarily a unique flow

The base TLV carries a 20-bit FlowMonID plus the loss (L) and delay (D) flags. The identifier helps implementations find a monitored flow because the IPv6 Flow Label cannot be used for that purpose. It is not a principal, a customer identity or a universal primary key.

RFC 9947 explicitly warns that a domain entry node choosing the 20-bit value, even pseudorandomly, can collide with another flow. The optional enhanced fields can extend the identifier and carry further metadata. That reduces ambiguity when the domain actually provisions and interprets those fields correctly. It does not eliminate the need to retain the matched flow attributes, policy revision and observation interval alongside a collector result.

This is where a good report refuses false precision. A line reading “FlowMonID 41718 lost 0.02%” has no standing by itself. It might mean a stable, uniquely mapped flow in one measurement period. It might mean two candidate flows collided. It might mean the selector changed at ingress. It might mean different collectors associated the same number with different rule versions. The number needs an accompanying mapping record before it can carry a claim.

Segment endpoints decide whether the field is read

SRH placement also narrows who processes the marker. With a Hop-by-Hop option, every hop on the path processes the option. With a Destination Option before the SRH, or with the AltMark SRH TLV, the relevant processing follows the segment endpoints indicated by the evolving destination address. RFC 9947 says an intermediate or egress SR node that cannot process the TLV simply ignores it under the SRH rules. A packet directed to a SID without the Alternate-Marking function need not even have the TLV examined.

This is a valuable operational choice: a domain can collect from supporting endpoints rather than demand a universal rollout. It also prevents an unwarranted sentence in every incident report: “the path was measured.” What was measured may instead be the subset of segment endpoints whose configured SID functions processed the TLV. The uninstrumented hops between them, a changed segment list, an ignored field or a nonparticipating function remain outside that assertion.

The optional M flag makes the boundary explicit: segment-by-segment and end-to-end measurement are different declared modes. Neither mode magically removes the need to show its participating points. An end-to-end label is a requested scope inside the domain, not proof of end-user experience beyond the domain. A segment result is a local observation, not proof that an application traversing several service domains met its objective.

Delay metadata needs its own chain of custody

Enhanced Alternate Marking can carry a timestamp; intermediate nodes can compare it with local time to derive one-way delay. It can also carry a sequence number to help identify reordering and loss, and controls for setting up reverse-direction monitoring. These are richer instruments, not self-executing measurements.

For a one-way delay claim, an operator should retain the timestamp source and epoch convention, the clocks used at each asserted observation point, the treatment of skew and wrap, the point at which the packet was captured, and the calculation that transformed packet fields into a reported value. This is an evidence recommendation, not a new RFC requirement. Its purpose is modest: prevent a local wall-clock difference from being represented as a path property.

The same applies to sequence analysis. A sequence field can show that a collector saw an order. It cannot distinguish every possible cause of an apparent gap: sampling, capture loss, a rule revision, a fragment treatment issue, a collector outage or actual packet loss. The enhanced F flag tells a processor to count a fragmented original packet only once; it does not prove that all fragments arrived or that the receiving application reconstructed content. Counted packets and completed application work remain different facts.

An experiment asks a question; it does not answer it in advance

The strongest lines in RFC 9947 are its questions. Does the SRH TLV survive better or worse than an extension-header alternative? Does it improve or hinder device processing? What does it do to the forwarding plane across architectures? How do extended fields compare with other on-path telemetry approaches? The document anticipates results being reported for later consideration. It does not supply them.

That is the right operational posture. Instrument a controlled trial as a comparison, not as a ceremony. Freeze the policy and software versions. Send comparable candidate traffic through the RFC 9343 and RFC 9947 forms where that is safe. Record treatment at each intended endpoint, collector loss, CPU or forwarding-path counters, identifier collisions, unexpected ignore behavior and the service canary. Then publish both the null result and the favorable result. “No measurable advantage on this path” is useful evidence. “The TLV exists” is not.

The evidence ladder

A defensible operational statement has at least five separate rungs:

  1. Declared experiment: the controlled domain, selected experimental value, flow selector, SR policy and participating SID functions were configured.
  2. Packet observation: the ingress actually added the intended TLV and specified endpoints actually observed or processed it during the recorded interval.
  3. Measurement inference: identifiers, periods, sequence handling and clocks were reconciled well enough to state a bounded loss, delay or jitter observation.
  4. Service correlation: an independent transaction, synthetic probe or application canary saw the corresponding service behavior on the relevant route.
  5. Decision: the accountable operator decided what to change, accept, escalate or roll back—and recorded why.

Skipping a rung changes the sentence that is permitted. A configuration supports “enabled.” A packet capture supports “observed here.” A reconciled collector supports “measured within the stated scope.” Only independent service evidence supports “the service was affected” or “the service met the target.” A protocol marker never selects the remediation by itself.

Sources