success6g-edge

Work in progress architecture

This repository is for the edge pod implementation, monitoring and analysis in SUCCESS6G project. Implementation is inspired by digits-recognizer-kubeflow.

Step by step procedure to run the implementation can be found in procedure.md

data and notebooks directories include analysis code used for intial edge model deployment and testing.

[PROD] Microk8s deployment

Description of the components:

• Grafana - dashboards
• Ingress - expose services to operator
• Prometheus - gather (i) pod metrics, (ii) measurements and predictions
• MinIO - store models and training/testing data
• Jupyter notebook (Kubeflow extension) - develop new models
• Kubeflow CI/CD pipelines
• Kserve - serve inference models to predefined pods
• Kepler - gather energy consumption data
• Redis - API for transfer of OBU measurements to Kubernetes

TODO

Edge pod #1

• test and prepare Redis helm chart
• deploy it only on specific nodes

Edge pod #2

• prepare CI/CD platform:
  • Kubeflow with Jupyternotebook extension, Kserve and MinIO for model/trining data storage
• expose measurements and predictions over http
• scrape data from edge by Prometheus, i.e. using service like kepler_exporter
  • TBD success6g service monitor
• deploy Kserve model only on specific nodes, i.e. solved github issue

Prometheus-stack

• finish Thanos configuration
• how to do tls-credential, hosts and externalUrl in NearbyComputing orchestration
  • ingress_prometheus_stack
  • prometheus_stack

Grafana

• make v2x dashboard

Open issues/questions

• organizational
  • Prometheus/Grafana configured by NBC or CTTC?
  • Redis db helm chart/docker provided by Idneo?
• technical
  • ensure comunication flows from OBU through edge Redis pod #1 -> Kserve pod #2
  • transfer of measuremetns (only a fraction?) from Prometheus to MinIO
  • federated elarning/training in Kubernetes? KubeFATE?

Additional

• impose security wherever possible
  • communication between Vehicle and Redis
  • exposing measurements and predictions via https
  • user/pass for Grafana

[DEV]KubeEdge deployment

• same as Microk8s except with KubeEdge and Kubeflow/Kserver is swapped for Sedna
• implement multimodel pods e.g. by ModelMesh, or alpha feature of Kserve