Networking

Description of traffic flow for simple(no Istio) Single cluster traffic flow.

Vehicle data processing

Vehicle data processing

Traffic flow:

1. Vehicle sends measurements to Vehicle API service (provided by Idneo – data preprocessing +Redis DB)
2. transmission is received by antenna
3. Transmission enters data(IPv4/IPv6) network
4. Data is forwarded to the closest Kubernetes Edge node using (i) Anycast IP address, shared by all Edge Nodes, and (ii) nodePort defined in NodePort service
5. Edge Node receives the data, checks the IP table in kube-proxy, and forwards it to local Pod A as the service is configured with externalTrafficPolicy: Local
   • If there is no local Pod for Vehicle API service, the packet is dropped
6. Vehicle API service forwards the data to the ML prediction service(Kserve inference)
7. Data is forwarded to Pod B as Kserve inference service is configured with internalTrafficPolicy: Local
   • If there is no local Pod for Kserve inference service, the packet is dropped
8. or ML prediction service pod periodically queries local Redis DB in Vehicle API service – as Vehicle API service is also configured with internalTrafficPolicy: Local
   • If there is no local Pod for Vehicle API service, the queries/requests are dropped
9. Kserve inference service pushes the data into InfluxDB
10. Grafana periodically queries data from InfluxDB

Energy consumption monitoring

Energy consumption monitoring

Traffic flow:

1. Kepler exporter uses eBPF to probe CPU, GPU, RAM performance counters and exposes them over time via HTTP
2. Prometheus periodically queries Kepler exporter for a new data
3. Grafana periodically queries Prometheus for anew data

Vehicle data and energy consumption monitoring

Vehicle data and energy consumption monitoring

Traffic flow:

1. User enters Grafana FQDN in the web browser and is forwarded to the Ingress controller
2. Ingress controller translates the FQDN to the appropriate service and forwards it
3. Grafana GUI comes up in the Users web browser

ML model development/update

ML model development/update

Traffic flow:

1. User enters MLflow FQDN in the web browser and is forwarded to the Ingress controller
2. Ingress controller translates the FQDN to the appropriate service and forwards it to Mlflow GUI where user can check current models
3. User updates the model locally or in cloud and uploads it to MLflow API
4. Mlflow API forwards the model to the backend database - Minio

ML model deployment

ML model deployment

Traffic flow:

1. Admin user instructs Kserve to deploy a model
2. Kserve queries the model from Minio storage
3. Kserve updates/”serves” the model in all Kserver inference Pods