RWS Architecture article

ADR-019 Hybrid network rollout: ACI in Location A (AI-only) and EVPN/VXLAN in Location B (target)

The rollout requires fast delivery of an AI cluster in Location A while the multi-tenant, multi-cluster platform is built and validated in Location B. A single network baseline acr

  1. Typeadr
  2. Statusaccepted
  3. Domainnetwork
On this page
  1. Context and Problem Statement
  2. Decision Drivers
  3. Considered Options
  4. Decision Outcome
  5. Positive Consequences
  6. Negative Consequences
  7. Operational Guardrails for the Hybrid Phase
  8. Pros and Cons of the Options
  9. Option 1 - Single EVPN/VXLAN baseline immediately
  10. Option 2 - ACI baseline for both locations
  11. Option 3 - Hybrid approach (chosen)
  12. Links and References

Context and Problem Statement

The rollout requires fast delivery of an AI cluster in Location A while the multi-tenant, multi-cluster platform is built and validated in Location B. A single network baseline across both locations increases short-term delivery risk for the AI lane. We need an explicit transitional model that keeps long-term target-state alignment clear.

Decision Drivers

  • Fast-track the AI cluster deployment in Location A with minimal lead time.
  • Keep the long-term datacenter network target aligned to EVPN/VXLAN in Location B.
  • Prevent uncontrolled dual-operating-model sprawl by hard-scoping the ACI lane.
  • Preserve auditability and cutover governance while both lanes exist.
  • Avoid ambiguity in platform onboarding scope per location.

Considered Options

  1. Single EVPN/VXLAN baseline immediately across both locations.
  2. ACI baseline for both locations in phase 1.
  3. Hybrid approach: ACI in Location A for AI-only, EVPN/VXLAN in Location B for target platform.

Decision Outcome

Chosen option: Option 3 - Hybrid approach (transitional).

Location A uses a single ACI pod only for the AI cluster use case. Location B remains the target-state build lane for EVPN/VXLAN multi-tenant networking and the multi-cluster platform.

This decision supersedes the single-fabric assumption in adr-016-network-topology-2026 for the current rollout phase.

Positive Consequences

  • Reduces short-term delivery risk while the EVPN/VXLAN blueprint is completed for Location B.
  • Provides a bounded acceleration lane for AI without blocking Location B platform build-out.
  • Keeps target-state investment focused on EVPN/VXLAN in the location that will host the multi-cluster baseline.

Negative Consequences

  • Dual-operating models increase complexity and make validation/auditing harder.
  • Creates avoidable migration work while not improving long-term alignment.
  • Requires strict scope controls so non-AI workloads do not expand into the ACI lane.

Operational Guardrails for the Hybrid Phase

  • Location A remains AI-only; non-AI tenant onboarding is blocked by policy and reviewed as an exception.
  • Location B EVPN/VXLAN lane must pass readiness evidence before expanding critical workloads (BGP/EVPN health, MTU validation, and route-leak negative tests).
  • DC-to-backbone connectivity follows the L3 interconnect model in adr-021-dc-evpn-to-backbone-l3vpn-interconnect; L2 cross-site extension is exception-only.
  • Backbone service/transport semantics follow adr-020-backbone-service-vs-transport-l3vpn-sr-mpls to avoid mixed-layer design decisions.

Pros and Cons of the Options

Option 1 - Single EVPN/VXLAN baseline immediately

  • Good: Fastest path to one operating model.
  • Bad: Higher delivery risk if EVPN/VXLAN readiness for the AI timeline is not met.

Option 2 - ACI baseline for both locations

  • Good: Simplifies near-term delivery if ACI capabilities are already available.
  • Bad: Delays target-state EVPN/VXLAN maturity and increases future migration effort.

Option 3 - Hybrid approach (chosen)

  • Good: Reduces short-term delivery risk while the EVPN/VXLAN blueprint is completed for Location B.
  • Bad: Dual-operating models increase complexity and make validation/auditing harder.
  • Bad: Creates avoidable migration work while not improving long-term alignment.