Summary

  • draft-ietf-intarea-reordering-00 distinguishes reordering from resequencing: the first changes arrival order; the second delays available frames to recreate the transmitted order.
  • A link can multiplex many transport contexts. One missing frame can therefore hold packets from unrelated flows even though the link does not know whether those endpoints need in-order delivery.
  • RACK and QUIC use time-based loss inference, while ROHC and ESP define their own bounded reordering behavior. Capability is specific to protocol, implementation, configuration and context.
  • The draft proposes exposing unavoidable beneficial reordering to endpoints and using subnetwork resequencing only when clear evidence shows its benefit exceeds added latency and jitter.
  • Revision 00 is an active Informational Internet-Draft proposing an RFC 3819 update, not an RFC, deployed change or universal performance result.

One hole, several queues of consequence

Imagine a bonded link carrying six live conversations. A frame from the first is lost on the slower member. Five later frames arrive promptly on the faster member. Their sequence numbers prove that a gap exists. The receiver places them in a resequencing buffer and waits.

The buffer is correct about sequence. It has no corresponding view of consequence. The absent frame may belong to a bulk transfer whose sender will recover it. The waiting frames may carry an interactive response, a DNS answer, an acknowledgement or a control message for entirely different transport contexts. Restoring one layer's chronology can consume another layer's deadline.

This is the sharpest point in the new INTAREA working-group draft. A layer-2 link often carries a multiplex of higher-layer contexts, yet resequencing is normally performed without knowing their semantics. The subnetwork can prove that it withheld packet 104 behind missing packet 103. It cannot infer from that record that packet 104 benefited from the wait.

Order is a fact; utility is a counterfactual

The draft uses two terms that operational dashboards often blur. Reordering is the event that makes the received sequence differ from the transmitted sequence. Resequencing is the action that returns it to transmitted order.

That action is not free. It requires buffer space, a loss or expiry rule and a wait budget. It also creates a counterfactual: what would the receiving endpoint have done if the later packets had been released immediately?

A sequence receipt cannot answer that question. To decide whether waiting helped, an operator needs both branches—or at least comparable traces. Record when each frame arrived, which higher-layer context it carried where observable, when the gap was declared lost, when held frames were released, what loss algorithm the endpoint used, whether a replay or compression window constrained acceptance, and what the application saw. “In order” is one field in that record, not the verdict.

The old warning changed underneath the network

RFC 3819 warned subnetwork designers about TCP's sensitivity to reordering and the assumptions of then-standard header compression. The new draft does not say those concerns were imaginary. It says the operational context changed.

Earlier TCP loss detection could treat three duplicate acknowledgements as evidence of loss. Reordering could therefore produce a spurious retransmission and an unnecessary reduction of the congestion window. RFC 8985's RACK instead derives loss from time and acknowledgement evidence. RFC 9002 gives QUIC time and packet thresholds and permits implementations to adapt to observed reordering.

The important word is can. A standards reference is not a live capability receipt. A network still needs to know which endpoint implementation is running, which algorithm is enabled, what thresholds it uses and whether the path's reordering lies inside those bounds. “Modern transport” must not become another unverifiable compatibility label.

The same caution applies to older endpoints. The draft notes that TCP without RACK does not simply cease to work when some packets arrive out of order; its performance may degrade. That is a different claim from saying every old stack tolerates every pattern, and different again from showing that resequencing improves a particular service.

Three links, three different control seams

The draft surveys cellular, Wi-Fi and DOCSIS to show why one universal switch is implausible.

LTE and 5G can reorder traffic because parallel HARQ processes finish at different times, because RLC retransmits after a lower-layer failure, or because Dual Connectivity paths have different delay. Resequencing may occur in RLC or PDCP depending on the system. A radio bearer is not merely “ordered” or “unordered”; mode, retransmission, compression and timer state define the actual contract.

Wi-Fi uses a resequencing buffer per traffic identifier in the described model. A missing frame in a block holds later frames until retransmission fills the gap or the receiver gives up. But sequential packet numbers are also used for replay protection. Releasing late frames safely therefore requires more than shortening a timer: replay logic must distinguish a valid delayed retransmission from an attack. Removing latency without redesigning that evidence boundary can turn a performance experiment into a security regression.

DOCSIS channel bonding creates skew without layer-2 retransmission. Downstream equipment maintains multiple resequencing contexts and bounded storage; upstream segmentation has its own sequence and reassembly requirements. Reassembling the bytes of a frame is necessary. Delaying already complete frames from independent contexts until an earlier sequence resolves is a separate policy decision.

Bounded tolerance is not unlimited tolerance

Header compression and IPsec illustrate why the endpoint argument must remain conditional. RFC 4224 explains how ROHC can operate over reordering channels, and ROHCv2 lists improved reordering tolerance. That does not prove that every deployed compressor uses the newer profile or has the same context window.

ESP's anti-replay service uses a sliding window. A conforming receiver can accept out-of-order packets within its configured range, but a packet that falls behind the window can be rejected even if it is authentic. “ESP handles reordering” omits the value that decides the result.

NAT and fragmentation bring another distinction. A requirement to support out-of-order fragments is not evidence that every installed box does so. The draft itself points to the mismatch between normative expectation and reported implementation behavior. The correct operational response is not to assume either universal fragility or universal tolerance. It is to observe the path and name the uncertain segment.

What the draft actually proposes

Revision 00 does not command every network to release every out-of-order packet. It proposes a decision rule.

Subnetworks should avoid introducing unnecessary reordering. When reordering is an unavoidable consequence of a mechanism that improves capacity or reliability—striping over multiple links or retransmitting damaged frames—the subnetwork should generally expose it to the endpoint. Resequencing should remain only where clear evidence shows that its benefits exceed the latency and jitter it adds.

That is a transfer of interpretive authority. Endpoints own transport and application semantics. The subnetwork knows the cost it imposes; it usually lacks the information needed to value the benefit. A default that spends latency without a capability record reverses the burden of proof.

Status without promotion

The first working-group revision was published on 27 August 2026 and expires on 28 February 2027. It is intended to be Informational and proposes to update Section 15 of RFC 3819 if approved. It requests no IANA action.

No IEEE, 3GPP or CableLabs rule changes merely because this draft exists. No named implementation or operator was tested for this Article. The Wi-Fi latency figures in the draft describe a cited scenario with seven competing stations; they are not universal Wi-Fi values or a measurement performed by BTW. The document may change, be replaced or expire.