Step 13 - Vision inspection Application

Goal

Deploy the vision-inspection Application with three microservices on three Edgelet nodes. Wafer thumbnails flow from site-a through ONNX inference on edge-c to an inspection store on ops-b.

Scenario context

Acme adds fab line fab-01 beside the existing plant floor on site-a. A frame-generator replays curated wafer map images every two seconds. An object-detector on edge-c runs the bundled wafer-defect.onnx model and publishes defect class plus confidence. A snapshot-service on ops-b keeps the latest inspection in memory for HTTP gallery access.

Ideal topology:

Microservice	Edgelet node	Role
frame-generator	site-a	Publish vision.wafers.fab-01
object-detector	edge-c	Subscribe wafers, publish vision.inspections.fab-01
snapshot-service	ops-b	Subscribe inspections, serve HTTP on port 8080

All three bind to account rule vision-inspection-local (step 12) and their user rules from step 11.

YAML walkthrough

File: deploy/steps/13-application-vision-inspection.yaml

Application header

vision-inspection Application

---
apiVersion: datasance.com/v3
kind: Application
metadata:
name: vision-inspection
spec:
natsConfig:
  natsAccess: true
  natsRule: vision-inspection-local

frame-generator

Image: ghcr.io/datasance/pot-edge-patterns/frame-generator:1.1.0

Variable	Value	Purpose
PUBLISH_INTERVAL	2s	Wafer publish cadence
FRAMES_MODE	wafer	v1.1.0 wafer dataset (not forklift v1.0.0)
LINE_ID	fab-01	Subject suffix vision.wafers.fab-01
FRAMES_DATASET_DIR	/assets/wafers	Bundled wafer thumbnails

Override agent.name if your production node is not site-a.

object-detector

Image: ghcr.io/datasance/pot-edge-patterns/object-detector:1.1.0 (amd64 and arm64 only; ONNX runtime)

Variable	Value	Purpose
LINE_ID	fab-01	Match frame-generator subject
MODEL_PATH	/models/wafer-defect.onnx	Wafer defect ONNX model
CLASSES_PATH	/config/wafer-classes.json	Defect class labels

Deploy on edge-c when available. ONNX inference is the edge AI story in this tutorial: inference runs near the fab line GPU/CPU on edge-c, not on the central ops node.

snapshot-service

Image: ghcr.io/datasance/pot-edge-patterns/snapshot-service:1.1.0

External port 8083 maps to HTTP 8080 inside the microservice. Step 14 creates the Router Service bridge. Leave SNAPSHOT_URL on the dashboard as a placeholder until step 15.

Single-node or two-node fallback

If you lack edge-c or ops-b, override all agent.name values in the YAML:

• Single node: set frame-generator, object-detector, and snapshot-service to your sole Edgelet node (for example site-a).
• Two nodes: move object-detector to site-a; keep snapshot-service on ops-b or colocate it with the generator.

Comments in the pot-edge-patterns manifest describe these fallbacks.

Why this design

Data path	Mechanism
Wafer thumbnails	NATS vision.wafers.fab-01 (in-account)
Defect inference	ONNX on edge-c, publish vision.inspections.fab-01
Gallery HTTP	snapshot-service NATS consumer + Router bridge (steps 14–15)

Multi-agent placement keeps heavy ONNX off site-a plant floor workloads and off ops-b presentation services. NATS subjects still work because all microservices share vision-inspection-local credentials, not because of cross-account import.

Deploy

Edit agent names for your ECN if needed, then:

potctl deploy -f deploy/steps/13-application-vision-inspection.yaml

Verify

potctl get microservices
potctl describe microservice frame-generator
potctl describe microservice object-detector
potctl describe microservice snapshot-service

Confirm each microservice shows:

• natsAccess: true with the correct user rule (frame-publisher, vision-detector, or snapshot-consumer)
• Expected agent.name for your topology

Log checks (after ~30 s):

• frame-generator: published waferId values on vision.wafers.fab-01
• object-detector: defect class and confidence on vision.inspections.fab-01
• snapshot-service: received inspection JSON (no NATS publish expected)

Common mistakes

• Missing edge-c. object-detector fails or never schedules if the agent is not provisioned. Use the single-node fallback or provision edge-c first.
• Wrong LINE_ID mismatch. frame-generator and object-detector must use the same LINE_ID or subjects will not match.
• Using v1.0.0 images or subjects. Redeploy with :1.1.0 and vision.wafers.> / vision.inspections.>.
• ONNX on unsupported arch. object-detector omits riscv64 and armv7 images. Use amd64 or arm64 Edgelet nodes for inference.
• Skipping steps 11–12. Application deploy fails or NATS auth violations appear without user and account rules.

Next step

Step 14: Snapshot Service: create the Router bridge so operators reach the inspection gallery.

See anything wrong with the document? Help us improve it!

Edit this page on GitHub!