ADR-020 Backbone service-vs-transport model: L3VPN over SR-MPLS

Context and Problem Statement

Network reviews still mix service policy and transport behavior into one layer. That causes route-policy ambiguity, weak ownership boundaries, and inconsistent change review.

The backbone runs tenant service semantics on L3VPN and transport semantics on SR-MPLS. We need one explicit model that keeps those concerns separate.

Decision Drivers

• Clear separation of service policy from transport path behavior.
• Consistent tenant VRF design at DC and backbone boundaries.
• Reviewable ownership and failover evidence per layer.

Considered Options

1. Treat L3VPN and SR-MPLS as one combined architecture layer.
2. Separate service and transport explicitly: L3VPN service over SR-MPLS transport.
3. Collapse backbone and datacenter into one EVPN-only model.

Decision Outcome

Decision remains draft. Working direction is Option 2.

Service semantics are modeled as L3VPN (RFC 4364) and transport semantics as SR-MPLS. Architecture and interconnect notes must keep this separation explicit.

Next Decision Trigger

• Decision owner: Network Architecture.
• Promote to accepted when adr-021-dc-evpn-to-backbone-l3vpn-interconnect is accepted and evidence confirms service-policy and transport-behavior separation in interconnect change gates.
• Required evidence sources: border-interconnect gate artifacts in topic-network-nxos-evpn-implementation and acceptance evidence for WP-01 DC3.0 Network and Supporting Services.
• Re-open this ADR if backbone transport strategy changes from SR-MPLS.

Pros and Cons of the Options

Option 1 – Combined service+transport layer

• Good: Simpler short-term language.
• Bad: Hides design boundaries and increases policy errors.

Option 2 – Explicit service/transport separation (working direction)

• Good: Makes policy, transport behavior, and evidence expectations explicit.
• Bad: Requires stricter documentation discipline.

Option 3 – One EVPN-only model end-to-end

• Good: One protocol family may simplify long-term operations.
• Bad: Does not match current backbone reality and adds migration risk.