The Architect's Guide to the AIoT - Part 2

Asheesh Goja
Asheesh Goja

Wednesday, April 6th, 2022

"So once you do know what the question actually is, you'll know what the answer means.” - HG2G


In part 1 of this series, we explored the AIoT problem space, the emergent behaviors, and the architecturally significant challenges. We learned how to address them using AIoT patterns and comprehensive reference architecture. In this post, I will show you how to apply the principles and patterns of this reference architecture to build a real-world AIoT application that can run on resource constrained edge devices.

Reference Implementation

While the reference architecture formalizes the recurring scenarios and repeatable best practices into abstract AIoT patterns, the reference implementation offers concrete archetypes that can be used as foundational building blocks for any AIoT application.
In this implementation, I have attempted to maximize the use of open source projects, however, in certain areas, none existed, so I wrote my own. I have coded these unopinionated modules with a deliberate openness to both extension and modification.

reference implementation

The reference implementation can be used as a collection of individual reusable libraries and templates, or as a unified application framework.

The reference implementation is organized into two sections:

  1. Reference Infrastructure : This is the infrastructure aspect of the reference implementation and is built using the technology mappings described below.
  2. Reference Application : This is the application aspect of the reference implementation that shows you how to build a "Real-world" AIoT solution on the reference infrastructure. The reference application is discussed in detail in the next part of this series.

The Reference Infrastructure

Technology Mappings - MLOps and Platform Services

The core platform and MLOps services of the reference infrastructure uses various CNCF projects from the Kubernetes ecosystem such as K3S, Argo, Longhorn, and Strimzi along with custom-coded modules in Go and Python. Here is the complete list of the mappings.

TierLayerReference ArchitectureReference Implementation
PlatformPlatform ServicesLightweight Pub/Sub Broker
Protocol Bridge
Event Streaming Broker
Model OTA Service
Model Registry
Device Registry
Training DataStore
Container Registry
Container Orchestration Engine
Container Workflow Engine
Edge Native Storage
Embedded Go MQTT Broker
MQTT-Kafka Protocol Bridge
Model OTA Server
Model Registry μService
Device Registry μService
Training Datastore μService
Docker Registry Service
Argo Workflows
PlatformMLOpsMLOps CD
Control and Data Events
Training Pipelines
Ingest Pipelines
Argo CD
Argo Dashboard
Control and Data Topics
Argo Workflows Training Pipeline
Data Ingest μService
Argo Demo DAG

Technology Mappings - Application services

The AIoT application services, which are covered in detail in the next post, are primarily comprised of custom coded modules in C++, Python, and Go.

TierLayerReference ArchitectureReference Implementation
Compressed ML Model
Context Specific Inferencing
Streaming Data
Orchestration Agent
Motor Condition Alerts
Quantized Model
TF Lite PyCoral Logistic Regression Module
K3S Agent
ThingsPerceptionProtocol Gateway
Sensor Data Acquisition
Pre Processing Filter
Actuator Control
Closed Loop Inferencing
Sensor module
FFT DSP Module
TF Lite Model Download
Servo Controller Module
TFLM Module
Aggregation Module

Infrastructure Hardware Specifications

Each infrastructure tier of this implementation uses a particular type of hardware and AI acceleration to ensure the resource availability, scalability, security, and durability guarantees of the tier are met. Each tier can independently scale and fail, enabling services on each tier to be deployed, managed, and secured independently. The hardware and OS specifications for each tier are listed here

Infrastructure TierDeviceAI AcceleratorComputeMemoryOS/Kernel
GPU - 128-core NVIDIA Maxwell™CPU – Quad-core ARM®
A57 @ 1.43 GHz
2 GB 64-bit LPDDR4Ubuntu
18.04.6 LTS
PlatformRaspberry Pi 4NoneQuad Cortex-A72
@ 1.5GHz
4GB LPDDR4Debian
GNU/Linux 10 (buster)
InferenceCoral Dev BoardGPU - Vivante GC7000Lite
TPU - Edge TPU
VPU - 4Kp60 HEVC/H.265
Quad Cortex-A53
@ 1.5 GHz
1 GB LPDDR4Mendel
GNU/Linux 5 (Eagle)
NoneMCU - Dual Core
Xtensa® 32-bit LX6 @ 40Mhz
448 KB ROM
NoneMCU - Dual Core
Xtensa® 32-bit LX6 @ 40Mhz
448 KB ROM

I will now show you how to configure each tier and prepare it to host an AIoT application.

Infrastructure Configuration

Configuring the Things Tier

The concrete implementation of this tier runs on an ESP32 SoC. The next post gets into the details of the hardware setup.

Configuring the Inference Tier

The concrete implementation of this tier runs on a cluster of three Coral Dev Boards and an ESP32 SoC. This tier hosts the following services:

The cluster of TPU Dev boards are ARM devices running Mendel Linux. These devices host the TFLite PyCoral modules.


We will first install the latest Linux Mendel OS on the Dev Boards by following these steps:
(Note: These steps are specific to the macOS)

  1. Install ADB tools on your laptop or PC
bash brew install android-platform-tools
  1. Install the CP210x USB to UART Bridge VCP Drivers
  2. Use a USB-micro-B cable and connect to the serial console port of the Dev Board
  3. Use serial terminal at 115200 baud to connect to the device
screen /dev/tty.SLAB_USBtoUART 115200
  1. Flash the latest firmware on the Coral Dev Board by following these instructions.
  2. Change the hostname of each of the Coral Dev Boards to agentnode-coral-tpu1, agentnode-coral-tpu2 and agentnode-coral-tpu3.

Configuring the Platform Tier

The concrete implementation of this tier runs a cluster of two raspberry pi devices and a NVIDIA Jetson Nano device.

  • The Jetson Nano device hosts the MLOps services that:

    • Runs extract, train, drift detection, and quantization tasks
    • Executes Argo DAGs that declaratively express the training workflow pipeline
  • The Raspberry pi cluster hosts platform services that:

    • Provides a browser based Argo MLOps dashboard
    • Runs data ingest jobs that subscribe to sensor data topics from the Kafka broker
    • Provides a private docker registry server
    • Hosts a K3S server
    • Hosts Argo workflows server
    • Provides a MQTT-Kafka protocol bridge
    • Hosts an embeded MQTT broker service
    • Provides ML model download OTA service
    • Hosts model registry, device registry and training datastore services
    • Hosts Longhorn services

Here are the steps to configure this tier.

Raspberry Pi configuration

  1. Download and flash the device with the “Debian Buster with Raspberry Pi” 64-bit ARM image.
  2. SSH into the device and confirm the OS is 64bit ARM by running
dpkg --print-architecture
  1. Update the OS using
sudo apt-get update
sudo apt-get upgrade
  1. Add the following lines to /boot/cmdline.txt (This is required for K3S and containerd to work correctly)
add cgroup_enable=cpuset cgroup_enable=memory cgroup_memory=1
  1. Change the hostname of each of the Raspberry Pis agentnode-raspi1 and agentnode-raspi2.
  2. Reboot the device.

NVIDIA Jetson Nano configuration

  1. SSH into the device and remove docker using the following commands
dpkg -l | grep -i docker
sudo apt-get purge -y docker-engine docker docker-ce docker-ce-cli
sudo apt-get autoremove -y --purge docker-engine docker docker-ce
sudo rm -rf /var/lib/docker /etc/docker
sudo rm /etc/apparmor.d/docker
sudo groupdel docker
sudo rm -rf /var/run/docker.sock
sudo rm -rf ~/.docker
  1. Change the hostname of each of the Jetson Nano device agentnode-nvidia-jetson.
  2. Reboot the device

At this point, the edge devices have all the prerequisite firmware and OS configurations needed to install and run the platform services. We will now install and configure various platform services for MLOps, communication, and container orchestration.

Container Orchestration Engine - K3S setup

In this reference infrastructure, K3S is set up in a single-server node configuration with an embedded SQLite database and requires two separate steps.


Step 1 - Server Node

The first step is to install and run the K3S server on the platform tier (a Raspberry Pi4 device or an equivalent VM). Here are the steps:

  • Open Firewall ports 6443, 32199, 1883, and 5000 if using a cloud VM.
  • Make sure the hostname resolves to the IP Address.
  • Get the IP Address of the device or the external IP address of the VM (if using a VM)
  • Install and run the server control node
#replace the <IP Address> with the IP Address of the device or VM
curl -sfL | INSTALL_K3S_EXEC="--write-kubeconfig ~/.kube/config --write-kubeconfig-mode 666 --tls-san <IP Address> --node-external-ip=<IP Address>" sh -
  • Confirm proper setup by using crictl
crictl info
  • Get the token to authorize the agent nodes
cat /var/lib/rancher/k3s/server/token

Step 2 - Agent Nodes

The agent nodes get installed on all the tiers except the things tier. Install the K3S agent on the Jetson Nano and Coral TPU Dev Kits, and then confirm proper setup using crictl

#replace the <IP Address> with the IP Address of the K3S server node
#replace the <TOKEN> with the token from the server node
curl -sfL | K3S_URL=https://<IP Address>:6443 K3S_TOKEN=<TOKEN> sh -
crictl info

With each successful agent node setup, you should be able to see the entire cluster by running this command on the K3S server node

kubectl get nodes -o wide -w

This is what I see on my cluster cluster-view

Edge Native Storage - Longhorn

Install longhorn by following these steps:

  1. On the platform tier (a Raspberry Pi4 device or an equivalent VM) install longhorn by following these instructions

  2. Create a new namespace architectsguide2aiot and label the raspberrypi device 1

kubectl create ns architectsguide2aiot
kubectl label nodes agentnode-raspi1 controlnode=active
  1. Add a node selector in the longhorn.yaml file to run the following longhorn CRDs only on devices labeled controlnode=active
apiVersion: v1
kind: ConfigMap
  name: longhorn-default-setting
  namespace: longhorn-system
  default-setting.yaml: |-
    system-managed-components-node-selector:"controlnode: active"
# add this for each of the the following CRDs
# DaemonSet/longhorn-manager
# Service/longhorn-ui
# Deploymentlonghorn-driver-deployer

  controlnode: active
  1. Install the ingress controller by following these instructions
kind: Ingress
  name: longhorn-ingress
  namespace: longhorn-system
    # type of authentication basic
    # prevent the controller from redirecting (308) to HTTPS "false"
    # name of the secret that contains the user/password definitions basic-auth
    # message to display with an appropriate context why the authentication is required "Authentication Required "
    - http:
          - pathType: Prefix
            path: "/"
                name: longhorn-frontend
                  number: 80
  1. Open the longhorn dashboard and navigate to settings->general. Set the configuration to following settings and save.
- Replica Node Level Soft Anti-Affinity : true
- Replica Zone Level Soft Anti-Affinity : true
- System Managed Components Node Selector : controlnode: active
  1. Label the raspberry pi device 2
kubectl label nodes agentnode-raspi2 controlnode=active
  1. Wait till all the CSI drivers and plugins are deployed and running on the raspberry pi device 2
NAME                                       READY   STATUS    RESTARTS      AGE    IP            NODE                        NOMINATED NODE   READINESS GATES
longhorn-csi-plugin-rw5qv                  2/2     Running   4 (18h ago)   10d    agentnode-raspi2            <none>           <none>
longhorn-manager-dtbp5                     1/1     Running   2 (18h ago)   10d    agentnode-raspi2            <none>           <none>
instance-manager-e-f74eeb54                1/1     Running   0             172m    agentnode-raspi2            <none>           <none>
engine-image-ei-4dbdb778-jbw5g             1/1     Running   2 (18h ago)   10d    agentnode-raspi2            <none>           <none>
instance-manager-r-9f692f5b                1/1     Running   0             171m    agentnode-raspi2            <none>           <none>
  1. On the dashboard confirm that you see two active nodes longhorn_nodes

  2. Open the volumes panel and then create a new volume with the following settings

Name      : artifacts-registry-volm
Size: 1 Gi 
Replicas: 1
Frontend  : Block Device
  1. Attach this volume to the agentnode-raspi2 device. Try attaching and detaching a few times. For some reason, it takes a few retries before the volume attaches.

  2. Using the dashboard create a PV and PVC in the namespace architectsguide2aiot and name it artifacts-registry-volm

kubectl get pv,pvc -n architectsguide2aiot

NAME                                      CAPACITY   ACCESS MODES   RECLAIM POLICY   STATUS   CLAIM                                         STORAGECLASS      REASON   AGE
persistentvolume/artifacts-registry-volm   1Gi        RWO            Retain           Bound    architectsguide2aiot/artifacts-registry-volm   longhorn-static            12d

NAME                                           STATUS   VOLUME                   CAPACITY   ACCESS MODES   STORAGECLASS      AGE
persistentvolumeclaim/artifacts-registry-volm   Bound    artifacts-registry-volm   1Gi        RWO            longhorn-static   12d

Container Registry Service - Private Docker Registry setup

Here are the steps to install and configure a private docker registry on the platform tier:

  1. Install docker a Raspberry Pi4 device or an equivalent VM
sudo apt-get update
	sudo apt-get remove docker docker-engine
	sudo apt install
	sudo systemctl start docker
	sudo systemctl enable docker
  1. Now start the docker distribution service on the device or VM. This is the local docker registry. The -d flag will run it in a detached mode.
docker run -d -p 5000:5000 --restart=always --name registry registry:2
  1. Edit /etc/docker/daemon.yaml to add an insecure registry entry
	"insecure-registries": ["localhost:5000"]

Note: I highly recommended that you use a secure registry using a proper CA and signed certs by following these instructions. But for this reference infrastructure, I am taking a shortcut and configuring an insecure registry.

  1. Restart the docker service
systemctl restart docker.service

K3S - Mirror Endpoints

  1. Configure a mirror endpoint in the K3S server node by editing the /etc/rancher/k3s/registries.yaml
#replace the <IP Address> with the IP Address of the node hosting the docker registry service
  docker.<IP Address>
      - "http://docker.<IP Address>"
  1. On each agent, node edit the containerd config file to add the private container registry mirror by following these steps:
  • Go to the folder /var/lib/rancher/k3s/agent/etc/containerd
  • In this folder make a copy of the config.toml file and name it config.toml.tmpl
  • Add this section to config.toml.tmpl file
  • Replace the <IP Address> with the IP Address of the node hosting the docker registry service
#replace the <IP Address> with the IP Address of the node hosting the docker registry service
    endpoint = [""]
  [plugins.cri.registry.mirrors."docker.<IP Address>"]
    endpoint = ["http://docker.<IP Address>"]
  1. Restart the k3s-agent service and verify the proper configuration of the k3s-agent service using crictl
systemctl restart k3s-agent.service
crictl info

Docker buildx

We also need to set up docker buildx which is used to build the ARM64 compatible inference modules images. On the device hosting the docker registry, initialize and setup docker buildx

docker buildx
docker buildx create --name mybuilder

Container Workflow Engine - Argo workflows setup

Argo workflow is used in this reference infrastructure to run parallel ML jobs expressed as DAGs.

argo-setup Here are the installation and configuration steps:

  1. Deploy the Argo workflow CRDs
kubectl create ns architectsguide2aiot
kubectl apply -n architectsguide2aiot -f
  1. Switch the workflow executor to the Kubernetes API. A workflow executor is a process that conforms to a specific interface that allows Argo to perform certain actions like monitoring pod logs, collecting artifacts, managing container lifecycles, etc
kubectl patch configmap/workflow-controller-configmap \
-n architectsguide2aiot \
--type merge \
-p '{"data":{"containerRuntimeExecutor":"k8sapi"}}'
  1. Port forward to open the argo console in a browser
kubectl -n architectsguide2aiot port-forward svc/argo-server 2746:2746
  1. Get the auth token
kubectl -n architectsguide2aiot exec argo-server-<pod name> -- argo auth token
  1. Open the Argo console in your browser and use the auth token from the previous step

Event Streaming Broker - Kafka Operator Strimzi

Strimzi provides the images and operators to run and manage Kafka on a Kubernetes cluster. We will now install and configure Strimzi on one of the Raspberry Pi devices.


This deployment includes the following components

  • Kafka - cluster of broker nodes
  • Kafka Connect - cluster for external data connections
  • Kafka MirrorMaker - cluster to mirror the Kafka cluster in a secondary cluster
  • Kafka Bridge - make HTTP-based requests to the Kafka cluster
  • ZooKeeper - cluster of replicated ZooKeeper instances


This deployment also includes the following Strimzi Operators:

  • Cluster Operator
  • Entity Operator
  • Topic Operator
  • User Operator

Here are the Installation steps:

  1. Create a namespace for strimzi deployment
kubectl create ns architectsguide2aiot
  1. Apply the Strimzi install file and then provision the Kafka Cluster
kubectl create -f '' -n architectsguide2aiot
kubectl apply -f '' -n architectsguide2aiot
  1. Modify the kafka-persistent-single.yaml to start the node port external listeners
kind: Kafka
  name: architectsguide2aiot-aiotops-cluster
    version: 2.8.0
    replicas: 1
      - name: plain
        port: 9092
        type: internal
        tls: false
      - name: tls
        port: 9093
        type: internal
        tls: true
      - name: external
        port: 9094
        type: nodeport
        tls: false
            nodePort: 32199
            - broker: 0
              nodePort: 32000
            - broker: 1
              nodePort: 32001
            - broker: 2
              nodePort: 32002
      offsets.topic.replication.factor: 1
      transaction.state.log.replication.factor: 1
      transaction.state.log.min.isr: 1
  1. Modify the tolerations and affinities to limit scheduling of pods to specific nodes
      - key: "dedicated"
        operator: "Equal"
        value: "Kafka"
        effect: "NoSchedule"
            - matchExpressions:
                - key: dedicated
                  operator: In
                    - Kafka
  1. Apply the modified configuration and wait for all the services to start
kubectl apply -f '' -n architectsguide2aiot
kubectl wait kafka/my-cluster --for=condition=Ready --timeout=300s -n architectsguide2aiot

Lightweight Pub/Sub Broker - Embedded MQTT broker setup

See this section from the next post.

Protocol bridge - MQTT-Kafka bridge setup

See this section from the next post.

AI Acceleration - Taints and Labels

The devices with AI accelerators such as GPUs or TPUs need to be labeled so as the ensure placement of ML workloads on the proper AI accelerated device.

kubectl label nodes agentnode-coral-tpu1 tpuAccelerator=true
kubectl label nodes agentnode-coral-tpu2 tpuAccelerator=true
kubectl label nodes agentnode-coral-tpu3 tpuAccelerator=true
kubectl label nodes agentnode-nvidia-jetson gpuAccelerator=true

In order to prevent strimzi from scheduling workloads on the devices in the inference tier use the following taints:

kubectl taint nodes agentnode-coral-tpu1 dedicated=Kafka:NoSchedule
kubectl taint nodes agentnode-coral-tpu2 dedicated=Kafka:NoSchedule
kubectl taint nodes agentnode-coral-tpu3 dedicated=Kafka:NoSchedule


In this post, we followed a detailed step-by-step guide for establishing a reference infrastructure on edge devices by installing and configuring various CNCF projects such as Argo, K3S, Strimzi, Longhorn, and various custom services.

In the concluding part of the series, we will see how to build, deploy and manage a “real world” AIoT reference application using TensorFlow Lite and TFLM and deploy it on this infrastructure.