The team reviewed the proposed architecture for HSP and tenant request routing interfaces, debated sequencing (HSP-first vs. tenant-first), and reached broad alignment on the key interfaces required. Four interface categories were identified: Â Â Â Network Mapping, Footprint Planning, Routing Configuration, and Health. Next step is to go one level deeper into data models for each interface.
Â
Action Items
- Team: Go one level deeper on each interface — define specific data primitives exchanged per interface.
- Team: Evaluate whether to produce one document per interface or one document per use case.
- Yoav: Add Delegation Methods section explicitly to the spec document.
- Team: Define tenant use-case flows for footprint planning (derive initial config, validate coverage, carve out subset) to clarify required data structure
Antitrust Statement and Group Discussion
Approach: HSP Routing First vs. Tenant Routing First
Yoav and Sanjay proposed starting with HSP request routing as the baseline, with tenant request routing treated as an additive expansion on top.
- Rationale: HSP routing is simpler, more analogous to today's CDN environment, and delivers a complete, implementable use case end-to-end.
- Alan's pushback: Not fully convinced tenant routing is purely additive — noted there is overlap and that tenant routing could arguably be the more important starting point.
- Resolution: Agreed to proceed with the exercise and revisit the sequencing question once interfaces are better defined.
Network Mapping Interface (Network Operator → HSP)
This interface provides the HSP with information about how network footprints and subnets map to sites within the network.
- Includes: routing permissions and preferences (where routing is allowed preferred), capacity indicators, route health, and special-use cases (private IPs, anycast support, etc.).
- Alan: The interface should reflect a live picture, not a static snapshot — changes in traffic engineering should propagate automatically.
- BGP as a candidate technology: Favored because it avoids building custom protocols — network operators can propagate traffic engineering decisions natively.
- Agreement: This interface is shared regardless of whether routing is HSP-controlled or tenant-controlled. It should be defined generically to allow multiple technology implementations.
Footprint Planning Interface (HSP → Tenant)
This interface enables tenants to understand where they can deploy and plan their deployment strategy accordingly.
Three distinct tenant use cases were identified for this interface:
- Initial deployment derivation: Given a list of sites (with compute, size, and network attributes), derive an initial deployment plan — e.g., "I want Edge-only for low latency" or "I want only large metro sites."
- Coverage validation: Validate that a proposed deployment covers all intended users, with appropriate primary and backup sites.
- Subset carve-out: Select only a subset of users/regions to serve — e.g., "I only need help with Texas users."
Key debates:
- Footprint-to-site vs. site-to-footprint mapping: Both directions were seen as useful. Yoav favored footprint-to-site; Alan noted that a reverse (site-to-footprint) map may be more intuitive for building an initial config. Conclusion: Provide both maps.
- Piotr's example: A footprint like "US, who.net" could map to a list of sites (e.g., Miami 1, Dallas 1, Dallas 2, LA 1, LA 2), allowing tenants to deploy on a cost-effective subset.
- Geolocation for users: Alan suggested geo attributes for users; Yoav cautioned this raises commercial and privacy concerns (geodatabase implications). Site geolocation is fine; user geolocation is not recommended.
- Availability / "when" dimension: Sanjay noted tenants also need real-time capacity information — not just where to deploy, but whether a site is available at the time needed.
- Scope question (unresolved): Should this interface live in the request routing track or the orchestration track? Yoav leaned toward out-of-scope for routing (it's essential for orchestration); Alan argued it must be linked and should remain in scope — ultimately it may need to be a combined interface serving both routing planning and orchestration.
Routing Configuration Interface (Tenant → HSP)
Allows tenants to configure and influence how the HSP routes requests to their deployed instances.
- Tenant defines services (mapped via service name / hostname), the instances associated with each service, and the locations where each service is deployed.
- Alan introduced the concept of "overlay services" — subgroups of instances with their own external endpoints and independent routing rules.
- Tenants need to specify failover and/or load-balancing rules per service.
- Different services can have different deployment profiles — e.g., software downloads at a central location vs. live streaming at all edge sites.
Health Interface (Tenant Instances → HSP)
A previously missing interface that was identified during the meeting as essential.
- Core requirement: HSP must know which tenant instances are currently alive and healthy before routing traffic to them — avoids routing to "black holes."
- Basic capability: Up/down status per instance (analogous to Kubernetes health checks).
- Extended capability: Load indicators — e.g., a tenant instance at 80% capacity can signal the HSP to offload requests to other locations, independent of request rate.
- Push vs. Pull: Alan favored polling instances directly for better scale; Yoav noted that for large numbers of small instances, tenant-push may be more appropriate. Conclusion: Both models should remain in scope; do not constrain the spec to one approach.
Delegation Methods
Alan raised the question of how traffic is handed off from tenant to HSP. This was briefly discussed but flagged for explicit documentation.
- DNS-based delegation: Assumed to be the primary method (~99.9% of cases).
- Other options: HTTP redirect, URL injection to cache.
- Action: Delegation methods must be explicitly defined in a coherent manner within the spec.