Kate Goldenring presented Akri, an open source Kubernetes resource interface that connects IoT devices to Kubernetes clusters. Akri discovers devices, represents them as native Kubernetes resources, and schedules workloads based on connected devices. It simplifies scenarios like video streaming from IP cameras to Kubernetes by automatically deploying pods to access devices. Akri uses a controller and agents and supports common protocols like ONVIF. It was demonstrated connecting IP cameras to a K3s cluster and deploying video servers and a firmware upgrade job from Kubernetes.

1. Discovering and
Managing IoT
Devices from
Kubernetes with
Akri
Kate Goldenring
docs.akri.sh

2. Who am I?
• Software Engineer II @ Microsoft
• Maintainer of Akri
• Open-source enthusiast, contributor to
Rust and Kubernetes projects (Krustlet,
kube-rs)
@kate-goldenring @KateGoldenring
docs.akri.sh

3. Agenda
1. Motivation
2. What is Akri?
3. Scenarios simplified with Akri
4. Akri Architecture
5. Demo
6. Roadmap
docs.akri.sh

4. Motivation
• Environment of nodes/servers in the cloud is homogenous
• Edge nodes are surrounded by a variety of IoT devices
• These IoT devices are often too small to run Kubernetes
• How can these devices be dynamically leveraged and managed by Kubernetes
workloads?
• Goal: The open source, standard way to connect devices to a cluster and
manage devices from a cluster
docs.akri.sh

5. Akri: A Kubernetes
Resource Interface
for the edge
Discovers IoT devices. Handles dynamic appearance
and disappearance of devices
Connects IoT devices to a Kubernetes cluster by
representing them as a native Kubernetes resources
Schedules workloads based on what devices are
connected to the cluster
Built with Rust to optimize for the edge and low
footprint clusters (K3s, MicroK8s, etc.)
Open-source CNCF Sandbox Project
docs.akri.sh

6. What IoT
devices?
• IoT devices do not need to be modified to
work with Akri
• Akri does not run anything on the IoT devices
rather communicates via common IoT
protocols
• udev (local devices such as USB cameras,
GPUs, microphones, etc),
• OPC UA (industrial machinery)
• ONVIF (IP cameras)
• Discovery interface is extensible*. Simple
gRPC interface for adding support for new
protocols.
* https://docs.akri.sh/development/handler-development

7. Akri for device use and management
Akri’s Controller automatically deploys “brokers” to discovered devices.
Brokers are either:
Long-running Kubernetes Pods – Ideal for device use Single-task Kubernetes Jobs to discovered devices –
Ideal for device management (introduced in Akri v0.8.4)
docs.akri.sh

8. IoT device use scenario with K8s
Without Akri
Node
Node
K8s Cluster
F = Frame Server Pod
I = Inferencing Pod
A = Alerting Pod
GPU GPU
F
F
F
F
F
F
I I
A
per camera
X FPS
GPU
I A
FFF
Steps 1
2
3
4
5
6
7
8
9

9. IoT device use scenario
Simplified with Akri
Node
Node
K8s Cluster
F = Frame Server Pod
I = Inferencing Pod
A = Alerting Pod
GPU GPU
F
F
F
F
F
F
I I
A
per camera
X FPS
AKRI
per camera
X FPS
Steps 1
2
3

10. IoT device management scenario
Simplified with Akri
Node
Node
v1 v1 v1 v1 v1 v1
K8s Cluster
F = Frame Server Pod
I = Inferencing Pod
A = Alerting Pod
U = Upgrade Job
GPU GPU
F
F
F
F
F
F
I I
A
AKRI
per camera
X FPS
U
U
U
U
U
U
v2 v2 v2 v2 v2 v2

11. Akri Architecture
Edge Cluster
Control Plane
Kubernetes
Scheduler
Akri
Controller
API Server
Node
Kubelet
Akri
Agent
<protocol>
kind: Configuration
metadata:
..name: akri-<protocol>
spec:
..discoveryHandler:
name: <protocol>
..brokerSpec:
containers:
- name: custom-broker
image: "ghcr.io/
etcd
Configuration
CRD
Instance
CRD
<protocol>
Configuration
<protocol>
Instance
Leaf
Device
Leaf
Device
Leaf Device
Broker
Broker
custom-
broker
<protocol>
Discovery
Handler

12. Installing with Helm
helm install akri akri-helm-charts/akri
--set kubernetesDistro=k3s
--set onvif.discovery.enabled=true
--set onvif.configuration.enabled=true
--set onvif.configuration.name='akri-onvif-fw-update'
--set onvif.configuration.brokerJob.image.repository='ghcr.io/../onvif-upgrade-broker'
--set onvif.configuration.capacity='2'
Akri Controller and Agent
Discovery Handler
Configuration

13. 2 Edge Nodes, 2 IP Cameras, Akri installed via Helm
Node
Node
IP Cam B
v1.0
Akri
Demo
Configuration CRD
Instance CRD
ONVIF Frame Server
Configuration
1 Install Akri with ONVIF
frame server Configuration
to use the cameras
2 Akri Agent discovers cameras
and creates Instance CRs
3 The Akri Controller
deploys the frame server
Pods and creates Services
IP Cam A
v1.0
Cam A Instance
Cam B Instance
Frame
Server
Pod
Frame
Server
Pod
Frame
Server
Pod
Frame
Server
Pod
K3s Cluster
Streaming
Application
4 Deploy streaming
application
5 Install ONVIF firmware
upgrade Configuration
to manage the cameras
ONVIF Firmware
Upgrade Configuration
Firmware
Upgrade
Job
Firmware
Upgrade
Job
6 The Akri Controller
deploys a firmware
upgrade Job to each
camera
v2.0 v2.0
Steps to reproduce this demo: https://hackmd.io/@akri/HJZORp5l9

14. Future of Device Management for Akri
• What Akri currently supports:
• Deployment of Kubernetes Jobs to each discovered device
• Deploying multiple Configurations to use the same devices in different ways
• What Akri does not currently support:
• Managed roll outs
• Displaying device management information in the Akri Instance
docs.akri.sh

15. Resources
Learn more
docs.akri.sh
Akri overview at KubeCon EU
May 2021
https://youtu.be/mcKNistZkrY
Join the community via
Slack and monthly Zooms
K8s #akri Slack
Try it out: End to end demo
discovering mock USB cameras
https://docs.akri.sh/demos
This IP camera demo: https://hackmd.io/@akri/HJZORp5l9