Summary

  • RFC 9744 separates a shared transport resource from the customer circuits it carries. Saving labels and suppressing per-circuit signaling are related, but different, choices.
  • Default FXC does not advertise an individual AC failure. VLAN-signaled FXC retains normalized-VID routes so a remote edge can remove the failed path for the affected traffic.
  • Recovery acceptance should identify the failed AC, its forwarding association and delivered service—not infer every customer's recovery from a working shared tunnel.

The saving appears before the obligation

A design review can count the service labels eliminated by aggregation before anyone has agreed who owns its new mapping table. That imbalance is the most consequential management question in EVPN-VPWS Flexible Cross-Connect, or FXC. The visible resource saving belongs to the network platform; the less visible obligation is to preserve enough circuit identity to repair the right service without disturbing its neighbours.

An attachment circuit, AC, is the customer-side connection: it may be a port, a VLAN on a port or a VLAN group. A provider edge, PE, carries that connection across the provider network. RFC 9744, published in March 2025, extends the EVPN-VPWS baseline so ACs on different interfaces and Ethernet Segments, ESs, can share a service tunnel. Its requirements distinguish label/OAM reduction from reducing BGP signaling as far as possible. It is not a promise that every mode eliminates both kinds of state.

The operational inference is simple: count the state that moves, not just the state that disappears. Configuration, remote attribution and customer-delivery evidence can remain necessary even when the label count falls. The RFC establishes a mechanism; it does not establish a particular vendor's support, a measured saving or an outage history.

One label is not one destination

FXC still has to separate the circuits inside the aggregate. Local VLAN identifiers can overlap on different interfaces. They are translated into normalized VIDs before MPLS encapsulation. At the receiving edge, the service label selects a VID-VRF; a further VID lookup selects the outgoing connection, followed by the required local tag handling. That two-stage association is specified in section 3.

For acceptance purposes, this suggests a chain of evidence: customer AC, local port and VID, normalized VID, service label and remote segment, eligible adjacency, actual delivery. Those are proposed review objects, not newly mandated telemetry fields. A correct label lookup with an incorrect final association is not a recovered customer service. Conversely, a local cross-connect may be delivering correctly without using the shared core path at all. Section 3.3.1 gives local switching precedence and requires unambiguous destination-AC or destination-ES label association in the described case.

A universal “every label means the shared VID-VRF” assumption would miss that exception.

Two ways to buy attribution

Default FXC leaves individual VLANs out of BGP signaling. For single-homed endpoints there is no alternate edge to which a failed connection can be redirected. The remaining cost is potentially carrying traffic across the core to discard it at the failed destination. Application behaviour can limit that waste in some cases; it is not a universal protection guarantee.

Multihoming changes the question from whether an alternative exists to whether the remote edge knows which traffic should use it. The two approaches purchase that knowledge differently. VLAN-signaled FXC advertises each normalized VID on its ES while allowing the multiplexed circuits to share the service label. Default multihoming instead groups ACs on one ES for one destination endpoint. It can require several tunnels between the same pair of PEs. Arbitrary cross-interface pooling is therefore not an entitlement to equally precise bundle-level recovery.

This is not a contest in which signaling is always wasteful and bundling always efficient. A well-scoped bundle can match the service's recovery boundary. Per-AC signaling can justify its cost where customers require independent failure treatment. The wrong comparison rewards the smallest route count without pricing the consequences of losing attribution.

Four failures, four pieces of evidence

The distinctions become concrete in section 5. An individual AC failure in default FXC is not signaled, so a remote PE may keep forwarding toward the failed egress. In VLAN-signaled mode, withdrawal of the affected normalized-VID route changes the outgoing adjacency for that traffic. A port failure has a wider scope: the relevant bundle tunnel or the affected VID routes and ES route are withdrawn. A PE failure is wider again, with withdrawals performed by the route reflector in the described procedure. A service failure may be detected through service OAM such as BFD for VCCV.

It follows that a successful port or PE failover does not prove individual-AC recovery. Nor does a live service OAM session prove delivery through every customer connection. RFC 5885 itself distinguishes pseudowire fault detection from AC/PW status signaling capabilities. The word BFD is not a substitute for identifying the selected capability and the object it observes.

An operator can compare ingress traffic destined for an unavailable AC with egress discards and the interval before it stops to estimate wasted transport. No fixed duration or bandwidth penalty is established here. The useful measurement is specific to the affected service and deployment, not an invented industry average.

Consistency is not a forwarding certificate

Optional mode and VID-normalization indications also need careful treatment. The M field is ignored for forwarding and supports error notification. The V indication can expose a single-tag/double-tag mismatch that prevents tunnel instantiation under section 3.4. An alarm, a permitted tunnel and a correctly delivered circuit are consequently three different observations—not interchangeable green lights.

RFC 9744 remains a Proposed Standard; the official errata search returned no matching entries on 13 September 2026. Neither finding certifies an implementation. The commercial case for FXC should rest on demonstrated mappings, appropriately scoped recovery and real resource measurements. Aggregation has earned its saving when the network can still name—and restore—the service that failed.