Summary

  • Revision 18 of the NETCONF adaptive-subscription draft lets a publisher choose a YANG-Push reporting period by evaluating mutually exclusive XPath conditions at a configured or server-selected cadence.
  • A connected impacted receiver must get adaptive-period-update; the notification cannot be dropped or filtered, yet it also cannot be stored in replay buffers.
  • Replayed telemetry can therefore lack the control-plane event that explains why its cadence changed. That is a recovery boundary, not evidence that the mechanism failed.
  • Condition evaluation, period switching and the next scheduled content update occur on different clocks. A faster period does not prove that a transient was evaluated or that every update reached the client.
  • Revision 18 is an Experimental Internet-Draft in the RFC Editor queue, not an RFC or a deployment result. Its contributor-supplied prototype is not independent interoperability evidence.

Three clocks behind one faster stream

The simple story is that a threshold is crossed and telemetry accelerates. The protocol story is more exact.

The subscriber supplies an eval-expression for a particular data-node instance and a reporting period. It can also supply eval-interval, which controls how often the server evaluates the condition. If that interval is absent, the server chooses the minimum cadence it can support. The draft says that choice is implementation-specific and may change with available CPU or memory.

When the expression evaluates true, the publisher switches to the associated period. Updates still occur on boundaries calculated from the period and anchor-time. The condition may therefore become true at one instant, be evaluated later, cause a switch at another instant and produce the next content record on a later scheduled boundary.

Those four moments should not be collapsed into “the alarm happened at 10:02.” The protocol can identify the switch time through period-update-time; it does not make continuous observation out of discrete evaluation.

XPath is policy, not a sensor

The trigger expression does not alter the content selection filter and does not itself generate a telemetry record. It chooses which configured period applies. That distinction protects a useful minimum: one expression governs cadence while another filter governs what the publisher sends.

It also exposes a precision boundary. XPath 1.0 represents numbers as IEEE 754 double precision. Some int64, uint64 and decimal64 values cannot be represented exactly. A complex comparison can be expensive, unsupported or behave differently from an operator's mental model. Revision 18 recommends a small node set, matching data types and constant numeric thresholds.

Multiple conditions in one adaptive configuration must be mutually exclusive. A server can reject a conflict through multi-xpath-criteria-conflict, reject an interval it cannot sustain, or reject unsupported syntax. Acceptance proves only that this publisher admitted the policy. It does not prove the threshold is operationally wise or that future load will preserve the same evaluation cadence.

Mandatory live delivery, deliberately absent history

The draft adds adaptive-period-update as a subscription state-change notification. It identifies the subscription and new period, and may include the switch time, the data that satisfied the condition and the content filter. For an online receiver, that is a valuable control receipt.

The delivery rules are unusually revealing. The notification cannot be dropped or filtered out. It goes only to impacted receivers. It also cannot be stored in replay buffers.

These are not contradictory requirements. Subscription state is live control context, not ordinary replayable subscribed content. But a recovery consequence follows. If a collector disconnects before the switch and returns after it, later records can arrive at the faster cadence without the replay buffer supplying the intervening switch notice. A replacement collector can observe the present period without possessing the earlier reason.

An operator who needs a durable explanation must preserve it elsewhere: in an append-only subscription ledger, collector journal or independent audit stream. That is BTW's evidence recommendation, not a hidden requirement in revision 18.

A dense stream can still have a hole

A shorter reporting period improves temporal resolution only under stated conditions. The server must actually evaluate the expression, schedule the period, produce records and deliver them. The receiver must absorb and process the increased volume. A reconnecting collector must know whether sequence gaps reflect transport loss, suspension, replay limits or a cadence change it never saw.

The draft acknowledges that burden. Its experiment asks for results on client buffers, processing overhead, sudden bursts, system resources, scalability and telemetry savings. It also says that a high volume of period-update notifications in a short window is a primary sign of oscillation or unstable expressions.

That makes the control notification an operational signal, but not a diagnosis. Rapid switching may reflect a noisy metric, a threshold without hysteresis, resource-driven evaluation behavior or a genuinely volatile system. The receipt needs enough context to distinguish those cases before automation treats them alike.

Status without inflation

The Datatracker places revision 18 in the RFC Editor queue for Experimental publication. The archived draft is dated 22 April 2026 and still carries Internet-Draft status. Queue entry is real process evidence; it is not an RFC number, deployment certificate or finding that the experiment succeeded.

The implementation-status section lists one contributor-supplied prototype for campus Wi-Fi telemetry. The text itself says such listings are not IETF endorsements and were not independently verified. A prototype proves that someone reported running code. It does not establish general XPath coverage, receiver recovery, cross-vendor interoperability or production stability.

The right claim is narrower and more useful: the draft supplies a concrete experimental contract for adaptive cadence, including explicit error identities and a visible switch message. The remaining evidence must come from implementations and operations.