Dynamic Membership Architecture (DMA) for Distributed Runtimes

In classical distributed systems, membership remains static or is managed by a centralized coordination service (such as ZooKeeper or etcd). When virtual servers spin down or new databases register, the configuration updates across the registry. However, in agentic networks, participant churn is much higher: agents are spawned dynamically to handle a single query, run short-lived operations, or exit due to failure.

PhoenixFlight's Dynamic Membership Architecture handles this volatility at the routing layer. It implements a decentralized control plane where participant state is managed using consistent hashing over virtual nodes. This design separates physical host resource management from logical agent membership.

Virtual Nodes & Consistent Hashing

To distribute tasks evenly and prevent hot-spotting, DMA maps each active agent (physical or container-based) to multiple virtual locations (v-nodes) on a logical 32-bit integer ring. When a task payload (FlightPacket) is routed, its required capabilities are hashed onto the same ring. The task is assigned to the nearest clockwise v-node. This guarantees that adding or removing a node only impacts a local subset of the ring, rather than forcing a full rehash of all workloads.

Stability & Performance Metrics

PhoenixFlight tracks system-wide balance indicators to maintain performance under churn:

• Standard Deviation of Load: Verifies that tasks are evenly distributed among active participants.
• Rebalancing Overhead Ratio: Tracks the ratio of migrated context packets compared to the total active tasks, aiming to keep it minimal during churn events.
• Handoff Latency: Measures the time required to sign, serialize, transmit, and verify task states between nodes.

Frequently Asked Questions

Citation & Archival Records

PhoenixFlight is connected to two Zenodo archival records: the Dynamic Membership Architecture research paper and the PhoenixFlight software artifact.