The Big Picture
A network of local brokers that trade short price signals can place AI pipeline stages across edge and cloud nearly as well as a central controller, improving mean latency and staying resilient while honoring domain data rules.
ON THIS PAGE
Key Findings
Local brokers run a simple buy/sell-style market using marginal-cost prices and tiny subscription summaries to decide whether a pipeline stage runs at the edge or in the cloud. Across a 4-domain testbed and three realistic pipeline shapes, the market-based federation matched or beat centralised baselines on end-to-end latency, kept work flowing under failures, and enforced per-domain data-sovereignty constraints with no measurable runtime cost. The design relies on well-understood economics results to explain when decentralised pricing can reach near-optimal placements, and it preserves privacy by exchanging only compact embeddings and prices rather than raw requests. Agent Service Mesh Pattern
Data Highlights
1Market beats a single-process centralised oracle by 2–4% mean latency across all 9 pipeline/load combos, winning every seed run (45/45).
2Performance stays within ±1.5% of a fair process-count sharded centralised oracle across all 9 cells — near-optimal decentralised allocation.
3System survives broker kills, network partitions, and worker failures with completion rates ≥98.7% across 75 fault-injection cells; applying sovereignty rules showed zero measurable runtime overhead across 60 governance runs.
What This Means
Platform engineers and site reliability teams who place multi-stage AI pipelines across edge and cloud can use price-based federation to avoid a central coordinator while keeping performance close to optimal. Telecom/cloud operators and technical leaders worried about enforcing data residency or cross-vendor policies get a way to enforce those rules without extra runtime cost or centralised control. A2A Protocol Pattern
Avoid common pitfallsLearn what failures to watch for
Key Figures
Fig 1: Figure 1 . Conceptual overview of Neural Pub/Sub. A pipeline (gray) is dispatched stage-by-stage (orange) onto workers across administrative domains ; each domain runs one broker closing a MAPE-K loop (expanded for Domain A). Brokers coordinate peer-to-peer over clearing-price signals (teal bus), without a central coordinator; local-only stages cannot cross a domain boundary ( ⊘ \oslash ). Schematic of the mechanism ( Section 4 ); not the evaluation testbed ( Section 5 ).
Fig 3: (a) Near-optimality: market vs single-process oracle, all cells.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreConsiderations
Results come from a 4-VM, 4-domain, 48-worker testbed with configurable stage delays rather than full ML inference, so absolute numbers may shift at production scale. The implemented pricing algorithm approximates the theoretically optimal auction, so a measurable efficiency gap may appear at larger scale or with different workloads. The design assumes cooperative brokers and reduces but does not eliminate privacy risks from embedding inversion; adversarial multi-vendor settings need added accountability and commitment mechanisms. Guardrails Pattern
Deep Dive
Neural Pub/Sub replaces a single global scheduler with one broker per administrative domain. Each broker monitors local workers, publishes compact subscription summaries (centroid embeddings and radii) and a clearing price, then places incoming pipeline stages locally or on a peer domain using a marginal-cost trade rule. That local Monitor-Analyse-Plan-Execute loop closes with a shared, bounded-staleness knowledge plane and supports per-domain sovereignty predicates as simple placement bounds. The execution showed a federated market matching or beating a centralised oracle, with strong resilience under stress and minimal overhead for sovereignty enforcement. The approach leverages planning-like structures in its loop; see Tree of Thoughts Pattern for related planning concepts and Blackboard Pattern for the shared knowledge plane concept. Evaluation on a 4-domain edge-to-cloud testbed used three 8-stage pipeline templates and a 1005-run campaign. The federated market matched or beat a centralised single-process oracle (2–4% mean latency improvement, 45/45 seed wins) and stayed within ±1.5% of a sharded centralised oracle with equal process count. The market preserved high completion rates under stress (≥98.7% under failures) and added zero measurable overhead when enforcing domain sovereignty. Limitations include the small-scale testbed, simulated stage processing times, and the assumption of cooperative brokers; future work targets larger heterogeneous domains, adversarial accounting, and running real inference workloads at scale. Blackboard Pattern
Need expert guidance?We can help implement this
Credibility Assessment:
Includes Sasu Tarkoma (h-index ~48) — top researcher signal despite arXiv venue, strong credibility.