What this covers

Polling wastes compute and adds latency. Azure Event Grid flips that around: when something changes (a blob lands, an inference completes), Event Grid pushes a small notification to whoever subscribed. This page introduces the Event Grid building blocks, the topic types, push vs pull delivery, and the event-driven patterns that show up in AI pipelines. It also draws the line between Event Grid and Service Bus, a favourite exam comparison.

Exam weight: Connect to and consume Azure services (20–25%)

Skills measured

"Implement event-driven workflows by using Azure Event Grid, including filters, custom events, and retries"

Key concepts

The five components

Component	What it is
Event	A small record that says something happened (who, what, when) plus a compact `data` payload. Max 1 MB per event; it references the resource rather than carrying it
Event source	The emitter: an Azure service (Blob Storage, Key Vault, Container Registry) or your code
Topic	The endpoint sources publish to
Event subscription	Which events from a topic to take (filters) and where to send them, plus retry and dead-letter settings
Event handler	The destination: Azure Functions, webhook, Service Bus queue or topic, Storage queue, Event Hubs

Topic types

Type	Who publishes	Created by	Notes
System topic	An Azure service only	Automatically when you subscribe to that resource's events (or explicitly)	You can't post to it; schema defined by the service, e.g. `Microsoft.Storage.BlobCreated`
Custom topic	Your application	You, before publishing	You choose the input schema (CloudEvents recommended)
Namespace topic	Your application or MQTT clients	Inside an Event Grid namespace	Supports pull delivery (and push to Event Hubs/webhooks); CloudEvents only

Domains (a management container for thousands of related custom topics, one per tenant) also exist for multitenant publishers.

Push vs pull delivery

Push: Event Grid POSTs to the handler and retries on failure. Used by system, custom and namespace topics.
Pull: the consumer connects to a namespace topic subscription and receives events when ready, then acknowledges, releases or rejects them by lock token. Good when the consumer can't expose a public endpoint or needs to control its own rate.

Delivery guarantees

At-least-once delivery, no ordering guarantee, retries with exponential backoff for up to 24 hours by default. Handlers must be idempotent (dedupe on event id). Details are on Configure delivery and retry policies.

Event-driven patterns for AI

Reactive data processing: subscribe to Microsoft.Storage.BlobCreated to start chunking and embedding as soon as a document is uploaded. No polling loop.
Pipeline stage coordination: each stage publishes a custom event (com.contoso.ai.EmbeddingsRefreshed) that triggers the next. Stages scale and deploy independently.
Model lifecycle: ModelRetrained or ModelPromoted events tell serving endpoints to refresh caches and run smoke tests.

How it works

A content moderation platform shows both topic types: the upload raises a system event, the classifier publishes a custom event, and each downstream service has its own subscription with its own filter. A failing subscriber doesn't block the others.

Hands-on

Azure portal

Open a storage account, select Events > + Event Subscription. The portal creates the system topic for you.
Choose Event Schema: Cloud Event Schema v1.0, filter to Blob Created, and pick Azure Function as the endpoint type.
Search Event Grid Topics > + Create to create a custom topic; on Advanced, set the input schema to Cloud Event Schema v1.0.

Azure CLI

Azure CLI / Bash

# Event Grid topic (custom)
az eventgrid topic create \
  --name <topic-name> \
  --resource-group <resource-group> \
  --location eastus \
  --input-schema cloudeventschemav1_0

# System-topic subscription: new blobs -> Azure Function
STORAGE_ID=$(az storage account show -n <storage-account> -g <resource-group> --query id -o tsv)
az eventgrid event-subscription create \
  --name blob-to-embed \
  --source-resource-id $STORAGE_ID \
  --endpoint-type azurefunction \
  --endpoint <function-resource-id> \
  --included-event-types Microsoft.Storage.BlobCreated \
  --subject-begins-with /blobServices/default/containers/documents/

Python

An Azure Function that handles the push delivery (Python v2 model). The Event Grid trigger handles the endpoint validation handshake for you.

Python

import azure.functions as func
import logging

app = func.FunctionApp()

@app.event_grid_trigger(arg_name="event")
def on_blob_created(event: func.EventGridEvent):
    data = event.get_json()
    logging.info("Event %s type=%s subject=%s url=%s",
                 event.id, event.event_type, event.subject, data.get("url"))
    # Download the blob by URL, chunk, embed, upsert into the vector store

Exam traps

Requirement	Choose
React to a state change in near real time, fan out to many handlers	Event Grid
Ordered, transactional processing of commands; consumer pulls at its own pace	Service Bus
High-volume telemetry stream with replay	Event Hubs
React to Azure resource events (blob created, secret near expiry)	Event Grid system topic
Publish your own app events	Custom topic (or namespace topic)
Consumer behind a firewall, no public endpoint	Namespace topic with pull delivery

Quick review

Event, source, topic, event subscription, handler: know each role.
System topics are Azure-emitted and read-only for publishers; custom topics are yours.
Namespace topics add pull delivery and MQTT.
Delivery is at-least-once and unordered; dedupe on id.
Max event size is 1 MB; send references, not payloads.
Microsoft.Storage.BlobCreated + Functions is the classic ingest trigger.

Practice

Practice: Event Grid questions

Understand Azure Event Grid concepts and event-driven patterns for AI solutions

What this covers

Key concepts

The five components

Topic types

Push vs pull delivery

Delivery guarantees

Event-driven patterns for AI

How it works

Hands-on

Azure portal

Azure CLI

Python

Exam traps

Quick review

Practice

Official references