Summary

  • A new GROW working-group draft would enrich BGP Monitoring Protocol reports with measurement-window duration, sample count, extrema, averages and percentiles.
  • The same draft says collection methods and sample populations remain implementation-specific, does not mandate a percentile algorithm and does not export the configuration that produced the values.
  • A valid P95 therefore has local meaning but is not automatically comparable across routers, vendors, releases or collection strategies.
  • The practical complement is a measurement-lineage receipt carried through the management plane: identity, window, sampling rule, exclusions, algorithm, clock quality, implementation and the decision that used the result.
  • That receipt is Daniel Kade’s editorial proposal, not a requirement of BMP, GROW or the IETF.

The calm line that hides the excursion

Imagine a collector receiving the same route-count snapshot every fifteen minutes: 100,000, then 100,000, then 100,000. The chart looks calm. Inside the first interval, however, the router may have fallen to 95,000 routes and climbed to 105,000 before settling at the familiar number. A periodic endpoint can be accurate at every observation and still conceal the movement that matters.

That is the problem addressed by the active GROW working-group draft BMP Statistics Information TLV. Revision 01, dated 11 September 2026, proposes additional distributional information for gauge-type BMP statistics. Instead of sending only a point value, a router could attach a measurement-window duration, a sample count and entries such as minimum, maximum, average, median, P5 and P95. Minimum and maximum carry the times when they were observed.

The change is useful precisely because RFC 7854 designed BMP as a convenient monitoring interface for BGP sessions and route views, not as a routing protocol. Monitoring data is used later: to investigate an event, compare capacity, tune alarms, explain a policy change or decide whether something unusual deserves escalation. The quality of that later decision depends on what the number preserves about its own production.

Revision 01 is still an Internet-Draft. The Datatracker history records the working-group document lineage, while the current status page shows no shepherd, responsible Area Director or telechat date and an IESG state of I-D Exists. The draft header says Standards Track; the status field currently displays no intended RFC status. That discrepancy is a reason to describe the live record, not to manufacture a settled outcome.

The document’s real contribution is twofold. It makes an interval less opaque, and it states the limits of the resulting statistics unusually clearly. Those two moves should be read together. The new fields can improve local observation. They do not turn locally produced aggregates into universal facts.

A window and a count are necessary context

The draft distinguishes the sampling interval from the reporting interval. A router might sample every sixty seconds but transmit a Statistics Report every fifteen minutes. A longer reporting interval reduces message and processing overhead; the faster internal observation can still catch excursions that a final snapshot would miss.

Revision 01 makes the measurement window and sample count explicit. Both must be greater than zero. The window can differ from the report-transmission interval, especially after a session is established or when a report is triggered by an event. If two values for one statistic were calculated over different windows, they must be sent in separate Statistics Information TLVs.

Sample count is more than decoration. Fifteen observations inside a 900-second window tell a different story from two observations inside the same duration. For an average, median or percentile, a low count weakens representativeness. For a reported minimum or maximum, it reduces confidence that the actual extreme between sampling instants was ever observed. The value can be a truthful minimum of the samples and still not be the minimum that the underlying system reached.

The distinction matters because the TLV is optional. A router may omit it entirely, enable it only for some peers, or support it for only a subset of statistics. The first report after session establishment or a peer’s return to Established state may omit the distributional information or calculate it from partial data. Absence can mean unsupported, disabled, selectively configured or not yet populated. It cannot safely be interpreted as proof that the metric did not vary.

The draft also limits the mechanism to gauges. Counters increase monotonically and have different semantics; a sender must not attach this distributional TLV to a counter-type statistic. A collector that receives such a reference should ignore it and may log the event. That rule prevents a named aggregate from escaping the kind of quantity for which it was defined.

The same P95 can have different parents

The central governance problem begins where the packet ends. Revision 01 allows periodic, event-driven, hybrid or other collection strategies. The entry name does not change when the method changes. P95 remains P95, but the population whose 95th percentile is being reported can be assembled in materially different ways.

A periodic sampler observes the metric when its clock fires. A short-lived excursion between ticks may never enter the sample population. An event-driven implementation records observations when routes are added or withdrawn; a turbulent period may therefore contribute many observations while a quiet period contributes few. A hybrid may combine scheduled points and event observations. All three can follow the draft. Their reported percentiles need not answer the same operational question.

Even an identical sample set does not guarantee an identical percentile. The draft explicitly leaves calculation to the implementation. Nearest-rank selection and linear interpolation can return different P95 values for the same ordered observations, especially in a small population. Median, minimum and maximum may look familiar enough to discourage a reader from asking how they were made. Familiar names are not method equivalence.

Windowing is another independent choice. The draft says the period is implementation-dependent. It will often cover the time since the prior report, but it need not. A sliding window, a tumbling interval, an event-bounded segment and a partial post-restart window can all create different populations even when their stated durations match.

The two timestamps do not close this gap. Minimum and maximum include observation times, but snapshot, average, median, P5 and P95 do not. The clock’s precision and accuracy also depend on the implementation. A collector learns when the router says it saw an extreme; it does not receive a full trajectory, a clock-quality statement or the times associated with the other aggregate values.

This is why the draft’s Operational Considerations call the results primarily local to the reporting router. Cross-router comparison requires compatible measurement-window and collection configurations. The document deliberately does not export those configurations or the internal collection methodology. Compatibility must be established through management or configuration channels.

That boundary is not a defect hidden by Daniel Kade’s analysis. It is written into revision 01. The mistake would be to preserve the number while discarding the warning.

A registry names the statistic, not the observation practice

The draft requests new IANA registries for Statistics Information entry types and Statistics Information TLV types. Its initial vocabulary gives stable names to minimum, maximum, snapshot, average, median, P5 and P95. It also asks for a note in the existing BMP Parameters registry distinguishing ordinary statistics from meta-information about another statistic.

That is coordination of meaning and code points. It is not certification of how a router sampled. A registry can say that entry type 7 means P95 and specify whether a timestamp field is present. It cannot, unless the protocol carries additional information, tell a downstream analyst whether the population was periodic, event-driven or hybrid, whether samples were missed, or which percentile convention was used.

RFC 8126 supplies the registration-policy vocabulary behind Standards Action and First Come First Served. Those procedures answer who may allocate and change a protocol parameter. They do not make unseen measurement practices interchangeable. A standardized noun can still refer to observations produced under different local rules.

Nor is every draft request already a live allocation. The current IANA page reflects published assignments and temporary registrations. A proposed registry in an active draft is evidence of intended coordination, not proof that the allocation has been completed or that the document has passed the standards process.

Revision 01 strengthens both the metric and the warning

The comparison with revision 00 is revealing. The earlier text already explained the difference between periodic reporting and internal observation. Revision 01 adds the measurement-window duration and sample count to the wire format, introduces P5 and P95, extends timestamps to the BMP seconds-and-microseconds form, adds explicit cautions about percentile algorithms and creates an Operational Considerations section devoted to local significance and comparability.

The 11 September I-D announcement establishes the publication event. It does not make every new sentence consensus or prove implementation. The GROW charter explains why the subject belongs there: the group covers operational measurement and monitoring of global routing and maintains BMP.

The revision therefore does not merely add more metrics. It creates a better boundary around what those metrics can support. A collector now has enough context to reject some careless comparisons: different windows, very different sample counts, partial post-session data. It still needs an external account of method before treating two results as members of one comparable series.

What the packet can prove

A well-formed Statistics Information TLV can prove a bounded set of facts about the sender’s statement. It identifies the referenced BMP statistic. It states a window duration and count. It names each aggregate entry. For minimum and maximum it attaches claimed observation times. It can be checked for length, allowed flags, duplicate entry types, gauge applicability and, in the AFI/SAFI case, presence of the corresponding ordinary statistic.

It cannot prove that every relevant change was sampled, that the clock was accurate, that two routers used the same population rule, that missing observations were handled alike, that the software implemented its configured algorithm correctly, or that a dashboard did not combine incompatible values. It does not preserve the human decision made from the number.

These are different claims. Valid transport is not complete lineage. A standardized label is not methodological equivalence. A high sample count is not proof of unbiased selection. An observed maximum is not the unknowable continuous maximum. A P95 is not portable merely because two devices encode it under the same entry type.

The measurement-lineage receipt

The practical response is not to burden BMP with every analytical policy. It is to ensure that the configuration channel invoked by the draft leaves a durable receipt that can travel with exported measurements inside the operator’s evidence system.

For each governed series, that receipt should identify the router, peer and referenced statistic; the AFI/SAFI where applicable; the window start, end and duration; the report and session boundary; the sampling trigger or cadence; the number of expected, observed and excluded samples; the treatment of gaps and duplicates; the percentile algorithm and rounding; the clock source, resolution and known uncertainty; the software and configuration version; any downstream aggregation; the accountable owner; and the condition that supersedes the record.

The receipt should not pretend to reproduce every raw observation. Its purpose is to make the comparison rule inspectable. If two routers differ in method, an analyst can keep the series separate, normalize with an explicit caveat or decline the comparison. If a release changes the algorithm, the boundary becomes visible instead of creating a silent break in the chart.

This is Daniel Kade’s proposed governance artifact. It is not normative text from the draft, an IETF requirement or a claim that the protocol should carry private configuration. It applies Heng Lu’s preference for a minimum initial specification and a visible policy mirror: preserve the smallest local record that lets another operator understand what was decided, why the result was admitted and when the method changed.

Heng Lu’s minimum-initial-specification argument matters here because uniform global methodology is neither necessary nor promised. Operators may choose different methods. What must not disappear is the boundary between those methods. The Policy Mirror supplies the institutional discipline: make the local rule and its consequences observable without converting one implementation’s choices into a central mandate.

Sources