Seedlabs

SpillRoute: Topology-Agnostic Path Steering for CDNs

A system for internet networks that finds the fastest data path by analyzing traffic congestion.

A traffic-steering controller for CDNs and multi-cloud networks that learns the best route or origin from the spillover in probe results — measuring one path reveals congestion on links shared with other paths — without needing a map of the network topology.

Decision SciencesAdvanced Bandit Algorithms Research
Hybrid Graph-Based Memory and SchedulingCloud networking and CDN — traffic steering, path selection, and origin selection

Concept

CDNs and multi-cloud routers steer traffic over many candidate paths and origins, guided by active measurements. Probes cost money and add latency, and standard practice wastes them: measuring one path also reveals the state of every path sharing its links, but most controllers either ignore this spillover or depend on a topology model that is stale the moment it is deployed (peering shifts, failures, maintenance). The paper's setting maps directly onto this problem: each candidate path is an action, and the directed observation system is the unknown, time-varying link-sharing structure. EXP3-IX is the first algorithm with near-optimal regret that does not need to know this observation system before selecting actions.

SpillRoute is a route/origin-selection controller built on that result. It (1) maintains candidate paths per point-of-presence-to-origin pair, (2) ingests latency and error measurements and attributes spillover online — when any probe reveals a shared link's state, all dependent paths are updated — and (3) re-selects routes with implicit exploration, which the paper shows is more efficient than explicit exploration, i.e., fewer probes wasted on bad paths for the same measurement quality. It is scoped to seconds-to-minutes route re-selection, matching the paper's round-based model with adversarially determined losses, not per-packet forwarding.

Why now

The paper proves the first algorithm (EXP3-IX) with near-optimal regret for bandit problems with side observations under a directed observation system chosen by the environment, without prior knowledge of that system — precisely the property a network controller needs, since topology is unknown and shifts over time. It further shows implicit exploration is more efficient computationally and information-theoretically than previously studied exploration strategies, which in this setting translates to lower probing overhead for the same learning quality. The evidence is theory-stage (round-based decisions, adversarial losses over T rounds), so SpillRoute targets slow-timescale steering where the round abstraction holds rather than microsecond packet forwarding.

Written by GLM · 11 Oct 2026

AI assessment

Backed by 1 paper83

A high-potential application of a theoretical regret-minimization algorithm to a concrete, high-value problem in CDN traffic steering, though it faces a significant gap between theoretical proof and production-grade network implementation.

Evidence strength
3/5
The idea is a direct application of the EXP3-IX algorithm, but the research is purely theoretical (regret bounds and round-based models) and lacks empirical validation in a real-world network environment.
Market pull
5/5
Major CDNs like Cloudflare and Akamai spend immense resources on probing and traffic steering; reducing probe overhead while improving route selection is a high-value objective.
Novelty & moat
5/5
The approach of topology-agnostic spillover learning is fundamentally different from the existing catalog's telemetry or deterministic routing ideas.
Feasibility
3/5
While the algorithm is defined, implementing it as a controller that interacts with real-world BGP/DNS steering mechanisms requires significant engineering beyond the paper's scope.
Wedge clarity
4/5
The focus on 'seconds-to-minutes route re-selection' provides a clear, manageable entry point that avoids the complexity of per-packet forwarding.
Simplicity / focus
5/5
The product is a single, focused controller with one clear function: optimizing path selection via implicit exploration.

Scored by Gemma 4 · 11 Oct 2026. Rated against a fixed rubric (evidence, market, novelty, feasibility, wedge, simplicity), relative to the closest ideas already in the catalog. A prior estimate to compare ideas before real-world signal arrives.

Persona discussion

AI personas trained on real people's expertise debate this idea as it evolves.

View the discussion →

Act on this idea

Ideas only matter if someone runs with them. Your message goes straight to the founder's inbox — nothing is stored on our servers.

Who benefits

  • Cloudflarecompany

    Argo Smart Routing already selects paths across transit providers and pays for measurement; exploiting probe spillover across shared links cuts measurement cost while adapting to topology changes without a maintained map.

  • Akamaicompany

    Operates one of the largest edge networks with intelligent traffic management; a topology-agnostic learner removes dependence on stale topology models across its delivery network.

  • Fastlycompany

    Its CDN makes real-time routing decisions where wasted probes on bad paths degrade user latency; implicit exploration reduces that waste for the same measurement quality.

  • Netflixcompany

    Open Connect selects origins and paths for video delivery across an appliance-based network where topology knowledge is partial and changes; learning from measurement spillover fits that operating reality.

Research it builds on

  1. Efficient learning by implicit exploration in bandit problems with side observations
    Tomáš Kocák, Gergely Neu, Michal Valko et al. · 2026 · 120 citations
    All ideas from this paper →
Leave feedback
feasibility