Redundancy & Scheduling Logic

Details on the “One Model – Two Nodes” policy, workload distribution, failover systems, and fairness in task assignment.

DeepNode is designed to maximize reliability, fault tolerance, and performance efficiency when running AI models. At the core of this system are two core principles:

• Redundancy — Every model is deployed across multiple nodes to ensure no single point of failure

• Scheduling Logic — Tasks are assigned dynamically based on real-time performance, load, and reputation.

Together, these systems ensure that users receive fast, accurate responses, even during high traffic or partial network disruptions.

---

Redundant Execution

To ensure high availability and resilience:

• Each production model is deployed on at least two Nodes, selected from different ends of the Node reputation spectrum:

  • One top-ranked Node (high reliability, high throughput)

  • One lower-ranked Node (to give new Nodes opportunity for growth and evaluation)

This 1 model → 2 Nodes setup enables:

• Higher uptime and availability

• Redundant fallback if a node fails

• Fair rotation and decentralization

If a Node crashes, times out, or returns invalid output, the system automatically retries on the backup node.

---

Scheduling Logic

When a user request is made, DeepNode routes it to the best-fit Node through a multi-factor scheduling algorithm.

Key factors include:

• Node reputation score

  Prioritizes Nodes with consistent performance, uptime, and low error rates

• Load balancing

  Avoids overloading a single Node to maintain low latency

• Recent failure rate

  Penalizes Nodes with recent downtime or failed executions

• Stake weight (optional override)

  Can be used in certain Domains to give preference to highly staked Nodes

---

Fallback & Retries

If a Node fails to return a valid response within a defined timeframe:

• The system automatically retries on the backup Node

• If both fail, the user receives an error but is not charged

• Node reputations are adjusted accordingly (penalty or reward)

This ensures users aren’t punished for infrastructure issues while incentivizes Miners to maintain reliable deployments.

---

Deployment Rotation

To prevent centralization and allow all Miners to build reputation over time:

• DeepNode rotates deployments periodically across eligible Nodes

• Models are rebalanced between top-tier and underutilized Nodes

• Poor-performing Nodes are gradually phased out or slashed

This keeps the network dynamic, fair, and censorship-resistant, without compromising performance.