The edgeHub is a store and forward broker: routes between modules keep the local pipeline running with no cloud in sight, messages destined upstream queue on disk (size the storage for your worst realistic outage), and the child devices pattern lets the edge box act as gateway for downstream sensors that authenticate through it. Design the split deliberately: the edge runs the always needed loop, ingest from the equipment protocols, filter, aggregate, detect, act locally within the latency budget, while the cloud gets the diet version, aggregates, exceptions, and samples, which is routinely a hundredfold bandwidth reduction, and the Stream Analytics on Edge option from September literally deploys the same windowed SQL to the device. The gateway patterns cover brownfield reality, the OPC UA and Modbus translation modules speaking to equipment older than the engineers maintaining it, protocol translation at the edge so the cloud side sees one coherent scheme.

Edge AI and Operations

The inspection and detection scenarios run models at the edge: containers carrying ONNX runtime or Triton with GPU or NPU acceleration where the hardware offers it, models trained in the Azure ML factory from this week, packaged as modules, and rolled out through the same manifests, with the drift monitoring loop closing over sampled data shipped back for evaluation. That train in cloud, infer at edge, monitor from cloud cycle is the reference MLOps extension for physical operations. The operational notes that keep fleets healthy: modules restart on failure but watch the twin reported health and edgeAgent metrics centrally through the monitoring stack, updates roll wave by wave (a bad module deployment to ten thousand sites is the DR post’s lessons in miniature, canary sites first), certificates on edge devices rotate on the same lifecycle discipline as everything else, and physical security belongs in the threat model because the device is in someone else’s building. Defender for IoT extends the behavioral watch to the edge and the OT networks behind it. The estate now computes where it must and centralizes where it can, which is the honest definition of edge architecture.

Cheers
Osama

Offline First and the Data Diet

The edgeHub is a store and forward broker: routes between modules keep the local pipeline running with no cloud in sight, messages destined upstream queue on disk (size the storage for your worst realistic outage), and the child devices pattern lets the edge box act as gateway for downstream sensors that authenticate through it. Design the split deliberately: the edge runs the always needed loop, ingest from the equipment protocols, filter, aggregate, detect, act locally within the latency budget, while the cloud gets the diet version, aggregates, exceptions, and samples, which is routinely a hundredfold bandwidth reduction, and the Stream Analytics on Edge option from September literally deploys the same windowed SQL to the device. The gateway patterns cover brownfield reality, the OPC UA and Modbus translation modules speaking to equipment older than the engineers maintaining it, protocol translation at the edge so the cloud side sees one coherent scheme.

Edge AI and Operations

The inspection and detection scenarios run models at the edge: containers carrying ONNX runtime or Triton with GPU or NPU acceleration where the hardware offers it, models trained in the Azure ML factory from this week, packaged as modules, and rolled out through the same manifests, with the drift monitoring loop closing over sampled data shipped back for evaluation. That train in cloud, infer at edge, monitor from cloud cycle is the reference MLOps extension for physical operations. The operational notes that keep fleets healthy: modules restart on failure but watch the twin reported health and edgeAgent metrics centrally through the monitoring stack, updates roll wave by wave (a bad module deployment to ten thousand sites is the DR post’s lessons in miniature, canary sites first), certificates on edge devices rotate on the same lifecycle discipline as everything else, and physical security belongs in the threat model because the device is in someone else’s building. Defender for IoT extends the behavioral watch to the edge and the OT networks behind it. The estate now computes where it must and centralizes where it can, which is the honest definition of edge architecture.

Cheers
Osama

The edgeHub is a store and forward broker: routes between modules keep the local pipeline running with no cloud in sight, messages destined upstream queue on disk (size the storage for your worst realistic outage), and the child devices pattern lets the edge box act as gateway for downstream sensors that authenticate through it. Design the split deliberately: the edge runs the always needed loop, ingest from the equipment protocols, filter, aggregate, detect, act locally within the latency budget, while the cloud gets the diet version, aggregates, exceptions, and samples, which is routinely a hundredfold bandwidth reduction, and the Stream Analytics on Edge option from September literally deploys the same windowed SQL to the device. The gateway patterns cover brownfield reality, the OPC UA and Modbus translation modules speaking to equipment older than the engineers maintaining it, protocol translation at the edge so the cloud side sees one coherent scheme.

Edge AI and Operations

The inspection and detection scenarios run models at the edge: containers carrying ONNX runtime or Triton with GPU or NPU acceleration where the hardware offers it, models trained in the Azure ML factory from this week, packaged as modules, and rolled out through the same manifests, with the drift monitoring loop closing over sampled data shipped back for evaluation. That train in cloud, infer at edge, monitor from cloud cycle is the reference MLOps extension for physical operations. The operational notes that keep fleets healthy: modules restart on failure but watch the twin reported health and edgeAgent metrics centrally through the monitoring stack, updates roll wave by wave (a bad module deployment to ten thousand sites is the DR post’s lessons in miniature, canary sites first), certificates on edge devices rotate on the same lifecycle discipline as everything else, and physical security belongs in the threat model because the device is in someone else’s building. Defender for IoT extends the behavioral watch to the edge and the OT networks behind it. The estate now computes where it must and centralizes where it can, which is the honest definition of edge architecture.

Cheers
Osama

Offline First and the Data Diet

The edgeHub is a store and forward broker: routes between modules keep the local pipeline running with no cloud in sight, messages destined upstream queue on disk (size the storage for your worst realistic outage), and the child devices pattern lets the edge box act as gateway for downstream sensors that authenticate through it. Design the split deliberately: the edge runs the always needed loop, ingest from the equipment protocols, filter, aggregate, detect, act locally within the latency budget, while the cloud gets the diet version, aggregates, exceptions, and samples, which is routinely a hundredfold bandwidth reduction, and the Stream Analytics on Edge option from September literally deploys the same windowed SQL to the device. The gateway patterns cover brownfield reality, the OPC UA and Modbus translation modules speaking to equipment older than the engineers maintaining it, protocol translation at the edge so the cloud side sees one coherent scheme.

Edge AI and Operations

The inspection and detection scenarios run models at the edge: containers carrying ONNX runtime or Triton with GPU or NPU acceleration where the hardware offers it, models trained in the Azure ML factory from this week, packaged as modules, and rolled out through the same manifests, with the drift monitoring loop closing over sampled data shipped back for evaluation. That train in cloud, infer at edge, monitor from cloud cycle is the reference MLOps extension for physical operations. The operational notes that keep fleets healthy: modules restart on failure but watch the twin reported health and edgeAgent metrics centrally through the monitoring stack, updates roll wave by wave (a bad module deployment to ten thousand sites is the DR post’s lessons in miniature, canary sites first), certificates on edge devices rotate on the same lifecycle discipline as everything else, and physical security belongs in the threat model because the device is in someone else’s building. Defender for IoT extends the behavioral watch to the edge and the OT networks behind it. The estate now computes where it must and centralizes where it can, which is the honest definition of edge architecture.

Cheers
Osama

Offline First and the Data Diet

The edgeHub is a store and forward broker: routes between modules keep the local pipeline running with no cloud in sight, messages destined upstream queue on disk (size the storage for your worst realistic outage), and the child devices pattern lets the edge box act as gateway for downstream sensors that authenticate through it. Design the split deliberately: the edge runs the always needed loop, ingest from the equipment protocols, filter, aggregate, detect, act locally within the latency budget, while the cloud gets the diet version, aggregates, exceptions, and samples, which is routinely a hundredfold bandwidth reduction, and the Stream Analytics on Edge option from September literally deploys the same windowed SQL to the device. The gateway patterns cover brownfield reality, the OPC UA and Modbus translation modules speaking to equipment older than the engineers maintaining it, protocol translation at the edge so the cloud side sees one coherent scheme.

Edge AI and Operations

The inspection and detection scenarios run models at the edge: containers carrying ONNX runtime or Triton with GPU or NPU acceleration where the hardware offers it, models trained in the Azure ML factory from this week, packaged as modules, and rolled out through the same manifests, with the drift monitoring loop closing over sampled data shipped back for evaluation. That train in cloud, infer at edge, monitor from cloud cycle is the reference MLOps extension for physical operations. The operational notes that keep fleets healthy: modules restart on failure but watch the twin reported health and edgeAgent metrics centrally through the monitoring stack, updates roll wave by wave (a bad module deployment to ten thousand sites is the DR post’s lessons in miniature, canary sites first), certificates on edge devices rotate on the same lifecycle discipline as everything else, and physical security belongs in the threat model because the device is in someone else’s building. Defender for IoT extends the behavioral watch to the edge and the OT networks behind it. The estate now computes where it must and centralizes where it can, which is the honest definition of edge architecture.

Cheers
Osama

Offline First and the Data Diet

The edgeHub is a store and forward broker: routes between modules keep the local pipeline running with no cloud in sight, messages destined upstream queue on disk (size the storage for your worst realistic outage), and the child devices pattern lets the edge box act as gateway for downstream sensors that authenticate through it. Design the split deliberately: the edge runs the always needed loop, ingest from the equipment protocols, filter, aggregate, detect, act locally within the latency budget, while the cloud gets the diet version, aggregates, exceptions, and samples, which is routinely a hundredfold bandwidth reduction, and the Stream Analytics on Edge option from September literally deploys the same windowed SQL to the device. The gateway patterns cover brownfield reality, the OPC UA and Modbus translation modules speaking to equipment older than the engineers maintaining it, protocol translation at the edge so the cloud side sees one coherent scheme.

Edge AI and Operations

The inspection and detection scenarios run models at the edge: containers carrying ONNX runtime or Triton with GPU or NPU acceleration where the hardware offers it, models trained in the Azure ML factory from this week, packaged as modules, and rolled out through the same manifests, with the drift monitoring loop closing over sampled data shipped back for evaluation. That train in cloud, infer at edge, monitor from cloud cycle is the reference MLOps extension for physical operations. The operational notes that keep fleets healthy: modules restart on failure but watch the twin reported health and edgeAgent metrics centrally through the monitoring stack, updates roll wave by wave (a bad module deployment to ten thousand sites is the DR post’s lessons in miniature, canary sites first), certificates on edge devices rotate on the same lifecycle discipline as everything else, and physical security belongs in the threat model because the device is in someone else’s building. Defender for IoT extends the behavioral watch to the edge and the OT networks behind it. The estate now computes where it must and centralizes where it can, which is the honest definition of edge architecture.

Cheers
Osama

The edgeHub is a store and forward broker: routes between modules keep the local pipeline running with no cloud in sight, messages destined upstream queue on disk (size the storage for your worst realistic outage), and the child devices pattern lets the edge box act as gateway for downstream sensors that authenticate through it. Design the split deliberately: the edge runs the always needed loop, ingest from the equipment protocols, filter, aggregate, detect, act locally within the latency budget, while the cloud gets the diet version, aggregates, exceptions, and samples, which is routinely a hundredfold bandwidth reduction, and the Stream Analytics on Edge option from September literally deploys the same windowed SQL to the device. The gateway patterns cover brownfield reality, the OPC UA and Modbus translation modules speaking to equipment older than the engineers maintaining it, protocol translation at the edge so the cloud side sees one coherent scheme.

Edge AI and Operations

The inspection and detection scenarios run models at the edge: containers carrying ONNX runtime or Triton with GPU or NPU acceleration where the hardware offers it, models trained in the Azure ML factory from this week, packaged as modules, and rolled out through the same manifests, with the drift monitoring loop closing over sampled data shipped back for evaluation. That train in cloud, infer at edge, monitor from cloud cycle is the reference MLOps extension for physical operations. The operational notes that keep fleets healthy: modules restart on failure but watch the twin reported health and edgeAgent metrics centrally through the monitoring stack, updates roll wave by wave (a bad module deployment to ten thousand sites is the DR post’s lessons in miniature, canary sites first), certificates on edge devices rotate on the same lifecycle discipline as everything else, and physical security belongs in the threat model because the device is in someone else’s building. Defender for IoT extends the behavioral watch to the edge and the OT networks behind it. The estate now computes where it must and centralizes where it can, which is the honest definition of edge architecture.

Cheers
Osama

Offline First and the Data Diet

The edgeHub is a store and forward broker: routes between modules keep the local pipeline running with no cloud in sight, messages destined upstream queue on disk (size the storage for your worst realistic outage), and the child devices pattern lets the edge box act as gateway for downstream sensors that authenticate through it. Design the split deliberately: the edge runs the always needed loop, ingest from the equipment protocols, filter, aggregate, detect, act locally within the latency budget, while the cloud gets the diet version, aggregates, exceptions, and samples, which is routinely a hundredfold bandwidth reduction, and the Stream Analytics on Edge option from September literally deploys the same windowed SQL to the device. The gateway patterns cover brownfield reality, the OPC UA and Modbus translation modules speaking to equipment older than the engineers maintaining it, protocol translation at the edge so the cloud side sees one coherent scheme.

Edge AI and Operations

The inspection and detection scenarios run models at the edge: containers carrying ONNX runtime or Triton with GPU or NPU acceleration where the hardware offers it, models trained in the Azure ML factory from this week, packaged as modules, and rolled out through the same manifests, with the drift monitoring loop closing over sampled data shipped back for evaluation. That train in cloud, infer at edge, monitor from cloud cycle is the reference MLOps extension for physical operations. The operational notes that keep fleets healthy: modules restart on failure but watch the twin reported health and edgeAgent metrics centrally through the monitoring stack, updates roll wave by wave (a bad module deployment to ten thousand sites is the DR post’s lessons in miniature, canary sites first), certificates on edge devices rotate on the same lifecycle discipline as everything else, and physical security belongs in the threat model because the device is in someone else’s building. Defender for IoT extends the behavioral watch to the edge and the OT networks behind it. The estate now computes where it must and centralizes where it can, which is the honest definition of edge architecture.

Cheers
Osama

Offline First and the Data Diet

The edgeHub is a store and forward broker: routes between modules keep the local pipeline running with no cloud in sight, messages destined upstream queue on disk (size the storage for your worst realistic outage), and the child devices pattern lets the edge box act as gateway for downstream sensors that authenticate through it. Design the split deliberately: the edge runs the always needed loop, ingest from the equipment protocols, filter, aggregate, detect, act locally within the latency budget, while the cloud gets the diet version, aggregates, exceptions, and samples, which is routinely a hundredfold bandwidth reduction, and the Stream Analytics on Edge option from September literally deploys the same windowed SQL to the device. The gateway patterns cover brownfield reality, the OPC UA and Modbus translation modules speaking to equipment older than the engineers maintaining it, protocol translation at the edge so the cloud side sees one coherent scheme.

Edge AI and Operations

The inspection and detection scenarios run models at the edge: containers carrying ONNX runtime or Triton with GPU or NPU acceleration where the hardware offers it, models trained in the Azure ML factory from this week, packaged as modules, and rolled out through the same manifests, with the drift monitoring loop closing over sampled data shipped back for evaluation. That train in cloud, infer at edge, monitor from cloud cycle is the reference MLOps extension for physical operations. The operational notes that keep fleets healthy: modules restart on failure but watch the twin reported health and edgeAgent metrics centrally through the monitoring stack, updates roll wave by wave (a bad module deployment to ten thousand sites is the DR post’s lessons in miniature, canary sites first), certificates on edge devices rotate on the same lifecycle discipline as everything else, and physical security belongs in the threat model because the device is in someone else’s building. Defender for IoT extends the behavioral watch to the edge and the OT networks behind it. The estate now computes where it must and centralizes where it can, which is the honest definition of edge architecture.

Cheers
Osama

Yesterday’s hub assumes telemetry travels to the cloud for processing. Physics and economics disagree often enough: the vision model inspecting parts needs milliseconds, the factory’s bandwidth cannot carry raw sensor streams, the site loses connectivity weekly, and some data may not leave the premises at all. IoT Edge moves the compute to the device, and it does it with the container discipline this series already speaks.

Modules, Manifests, and the Runtime

An edge device runs the IoT Edge runtime: the edgeAgent supervising workloads and the edgeHub brokering messages, both managing modules, which are simply containers from your ACR. The deployment manifest declares which modules run, their create options, routes between them, and desired properties, and it deploys through the twin machinery from yesterday: at scale deployments target twin tag defined groups (all devices where tags.site equals plant7), layered deployments compose the base stack plus site specific additions, and the fleet converges declaratively with metrics on who has not. The composition is exactly the October GitOps mental model: manifests in Git, CI building module images through the registry patterns, deployments as reviewed changes, and the fleet as a set of convergence targets rather than machines someone configures.

The inspection and detection scenarios run models at the edge: containers carrying ONNX runtime or Triton with GPU or NPU acceleration where the hardware offers it, models trained in the Azure ML factory from this week, packaged as modules, and rolled out through the same manifests, with the drift monitoring loop closing over sampled data shipped back for evaluation. That train in cloud, infer at edge, monitor from cloud cycle is the reference MLOps extension for physical operations. The operational notes that keep fleets healthy: modules restart on failure but watch the twin reported health and edgeAgent metrics centrally through the monitoring stack, updates roll wave by wave (a bad module deployment to ten thousand sites is the DR post’s lessons in miniature, canary sites first), certificates on edge devices rotate on the same lifecycle discipline as everything else, and physical security belongs in the threat model because the device is in someone else’s building. Defender for IoT extends the behavioral watch to the edge and the OT networks behind it. The estate now computes where it must and centralizes where it can, which is the honest definition of edge architecture.

Cheers
Osama

The edgeHub is a store and forward broker: routes between modules keep the local pipeline running with no cloud in sight, messages destined upstream queue on disk (size the storage for your worst realistic outage), and the child devices pattern lets the edge box act as gateway for downstream sensors that authenticate through it. Design the split deliberately: the edge runs the always needed loop, ingest from the equipment protocols, filter, aggregate, detect, act locally within the latency budget, while the cloud gets the diet version, aggregates, exceptions, and samples, which is routinely a hundredfold bandwidth reduction, and the Stream Analytics on Edge option from September literally deploys the same windowed SQL to the device. The gateway patterns cover brownfield reality, the OPC UA and Modbus translation modules speaking to equipment older than the engineers maintaining it, protocol translation at the edge so the cloud side sees one coherent scheme.

Edge AI and Operations

The inspection and detection scenarios run models at the edge: containers carrying ONNX runtime or Triton with GPU or NPU acceleration where the hardware offers it, models trained in the Azure ML factory from this week, packaged as modules, and rolled out through the same manifests, with the drift monitoring loop closing over sampled data shipped back for evaluation. That train in cloud, infer at edge, monitor from cloud cycle is the reference MLOps extension for physical operations. The operational notes that keep fleets healthy: modules restart on failure but watch the twin reported health and edgeAgent metrics centrally through the monitoring stack, updates roll wave by wave (a bad module deployment to ten thousand sites is the DR post’s lessons in miniature, canary sites first), certificates on edge devices rotate on the same lifecycle discipline as everything else, and physical security belongs in the threat model because the device is in someone else’s building. Defender for IoT extends the behavioral watch to the edge and the OT networks behind it. The estate now computes where it must and centralizes where it can, which is the honest definition of edge architecture.

Cheers
Osama

Offline First and the Data Diet

The edgeHub is a store and forward broker: routes between modules keep the local pipeline running with no cloud in sight, messages destined upstream queue on disk (size the storage for your worst realistic outage), and the child devices pattern lets the edge box act as gateway for downstream sensors that authenticate through it. Design the split deliberately: the edge runs the always needed loop, ingest from the equipment protocols, filter, aggregate, detect, act locally within the latency budget, while the cloud gets the diet version, aggregates, exceptions, and samples, which is routinely a hundredfold bandwidth reduction, and the Stream Analytics on Edge option from September literally deploys the same windowed SQL to the device. The gateway patterns cover brownfield reality, the OPC UA and Modbus translation modules speaking to equipment older than the engineers maintaining it, protocol translation at the edge so the cloud side sees one coherent scheme.

Edge AI and Operations

The inspection and detection scenarios run models at the edge: containers carrying ONNX runtime or Triton with GPU or NPU acceleration where the hardware offers it, models trained in the Azure ML factory from this week, packaged as modules, and rolled out through the same manifests, with the drift monitoring loop closing over sampled data shipped back for evaluation. That train in cloud, infer at edge, monitor from cloud cycle is the reference MLOps extension for physical operations. The operational notes that keep fleets healthy: modules restart on failure but watch the twin reported health and edgeAgent metrics centrally through the monitoring stack, updates roll wave by wave (a bad module deployment to ten thousand sites is the DR post’s lessons in miniature, canary sites first), certificates on edge devices rotate on the same lifecycle discipline as everything else, and physical security belongs in the threat model because the device is in someone else’s building. Defender for IoT extends the behavioral watch to the edge and the OT networks behind it. The estate now computes where it must and centralizes where it can, which is the honest definition of edge architecture.

Cheers
Osama

Offline First and the Data Diet

The edgeHub is a store and forward broker: routes between modules keep the local pipeline running with no cloud in sight, messages destined upstream queue on disk (size the storage for your worst realistic outage), and the child devices pattern lets the edge box act as gateway for downstream sensors that authenticate through it. Design the split deliberately: the edge runs the always needed loop, ingest from the equipment protocols, filter, aggregate, detect, act locally within the latency budget, while the cloud gets the diet version, aggregates, exceptions, and samples, which is routinely a hundredfold bandwidth reduction, and the Stream Analytics on Edge option from September literally deploys the same windowed SQL to the device. The gateway patterns cover brownfield reality, the OPC UA and Modbus translation modules speaking to equipment older than the engineers maintaining it, protocol translation at the edge so the cloud side sees one coherent scheme.

Edge AI and Operations

The inspection and detection scenarios run models at the edge: containers carrying ONNX runtime or Triton with GPU or NPU acceleration where the hardware offers it, models trained in the Azure ML factory from this week, packaged as modules, and rolled out through the same manifests, with the drift monitoring loop closing over sampled data shipped back for evaluation. That train in cloud, infer at edge, monitor from cloud cycle is the reference MLOps extension for physical operations. The operational notes that keep fleets healthy: modules restart on failure but watch the twin reported health and edgeAgent metrics centrally through the monitoring stack, updates roll wave by wave (a bad module deployment to ten thousand sites is the DR post’s lessons in miniature, canary sites first), certificates on edge devices rotate on the same lifecycle discipline as everything else, and physical security belongs in the threat model because the device is in someone else’s building. Defender for IoT extends the behavioral watch to the edge and the OT networks behind it. The estate now computes where it must and centralizes where it can, which is the honest definition of edge architecture.

Cheers
Osama