SuperLink High Availability Mode

SuperLink High Availability Mode is an architecture for running SuperLink without a single SuperLink process acting as the only entry point for a federation. Instead, traffic is routed through a load balancer to multiple SuperLink replicas that share durable state in a PostgreSQL backend.

중요

SuperLink High Availability Mode is available in preview. To discuss production requirements or request preview access, contact the Flower team at enquiries@flower.ai.

This page describes the architecture at a high level. It is not a deployment guide.

Architecture

In a standard Deployment Runtime setup, callers and SuperNodes connect to one long-running SuperLink. In High Availability Mode, callers and SuperNodes connect to a stable load-balanced endpoint. The load balancer distributes requests across multiple SuperLink replicas. Each replica reads and writes shared state through the same PostgreSQL backend, so the replicas observe the same federation, run, node, message, and object state.

flowchart LR CLI["flwr CLI / SuperExec"] Nodes["SuperNodes"] LB["Load balancer"] SL1["SuperLink replica"] SL2["SuperLink replica"] SLN["SuperLink replica"] DB[("PostgreSQL backend")] CLI --> LB Nodes --> LB LB --> SL1 LB --> SL2 LB --> SLN SL1 --> DB SL2 --> DB SLN --> DB

The main components are:

• Load balancer: exposes the stable SuperLink endpoint used by flwr CLI commands, SuperNodes, and SuperExec processes. It routes traffic only to healthy SuperLink replicas.

• SuperLink replicas: run the same SuperLink service role in parallel. Any healthy replica can handle requests for the federation because state is externalized.

• PostgreSQL backend: stores shared SuperLink state durably so replicas can observe the same runs, messages, node registrations, and object metadata.

• SuperNodes and SuperExec: continue to connect to a SuperLink endpoint. They do not need to know which SuperLink replica handles an individual request.

Operational properties

High Availability Mode is designed to reduce the impact of losing an individual SuperLink process. If one SuperLink replica becomes unavailable, the load balancer can route new requests to another healthy replica while the shared PostgreSQL backend continues to provide the federation state.

This changes the deployment architecture of the SuperLink layer, but it does not change the Flower app programming model:

• ServerApp and ClientApp code remains the same.

• SuperNodes connect to the configured SuperLink endpoint, not to individual replicas.

• Operators monitor and scale SuperLink replicas as part of the SuperGrid deployment.

• The load balancer and PostgreSQL backend become critical infrastructure and must be operated with their own availability, backup, and monitoring strategy.

High Availability Mode should be understood as part of a production SuperGrid architecture. It improves resilience at the SuperLink layer, but end-to-end production availability still depends on the surrounding infrastructure, including load balancing, database availability, networking, authentication services, and application-level retry behavior.