Change Feed with Azure Functions and Processors

What is the change feed?

Amara’s scenario: real-time sensor processing

📡 Amara at SensorFlow ingests 500M IoT events per day into a Cosmos DB container partitioned by /deviceId. She needs to:

• Trigger alerts when a sensor reading exceeds a threshold
• Update a dashboard materialized view in near-real-time
• Feed analytics by streaming changes to a data warehouse

The change feed is the backbone of all three.

Azure Functions Cosmos DB trigger

The simplest way to consume the change feed — Azure Functions handles the infrastructure:

[FunctionName("SensorAlerts")]
public static void Run(
    [CosmosDBTrigger(
        databaseName: "sensorflow",
        containerName: "readings",
        Connection = "CosmosDBConnection",
        LeaseContainerName = "leases",
        CreateLeaseContainerIfNotExists = true
    )] IReadOnlyList<SensorReading> changes,
    ILogger log)
{
    foreach (SensorReading reading in changes)
    {
        if (reading.Temperature > 85.0)
        {
            log.LogWarning($"ALERT: Device {reading.DeviceId} at {reading.Temperature}°C");
            // Send notification, write to alert container, etc.
        }
    }
}

Key points:

• The trigger creates and manages the lease container automatically
• Changes arrive as a batch — process the entire batch, not item by item
• If the function fails, the batch is retried from the last checkpoint
• The function scales automatically based on the number of physical partitions

💡 Exam tip: lease container purpose

The lease container tracks which changes have been processed and by which consumer. It stores one lease document per physical partition of the monitored container. This enables:

• Checkpointing: If your processor crashes, it resumes from the last checkpoint, not from the beginning.
• Parallel processing: Multiple instances claim different leases, distributing work.
• Load balancing: If an instance dies, its leases are redistributed to surviving instances.

The lease container must be in the same Cosmos DB account as the monitored container.

Change Feed Processor library

For more control than Azure Functions, use the Change Feed Processor directly:

Container leaseContainer = cosmosClient
    .GetDatabase("sensorflow")
    .GetContainer("leases");

Container monitoredContainer = cosmosClient
    .GetDatabase("sensorflow")
    .GetContainer("readings");

ChangeFeedProcessor processor = monitoredContainer
    .GetChangeFeedProcessorBuilder<SensorReading>(
        processorName: "dashboardUpdater",
        onChangesDelegate: HandleChangesAsync)
    .WithInstanceName("instance-1")
    .WithLeaseContainer(leaseContainer)
    .WithStartTime(DateTime.MinValue.ToUniversalTime()) // from beginning
    .Build();

await processor.StartAsync();

static async Task HandleChangesAsync(
    ChangeFeedProcessorContext context,
    IReadOnlyCollection<SensorReading> changes,
    CancellationToken cancellationToken)
{
    foreach (SensorReading reading in changes)
    {
        Console.WriteLine($"Partition {context.LeaseToken}: {reading.DeviceId}");
    }
}

Change feed modes

Feature	Latest Version (default)	All Versions and Deletes (preview)
What's captured	Most recent version of inserts/updates	Every version of every change + deletes
Deletes	❌ Not captured (use soft-delete + TTL)	✅ Captured with metadata
Intermediate versions	❌ Only latest if multiple changes before read	✅ Every intermediate version
Retention	Unlimited	Configured by continuous backup retention window
Use case	Event reactions, materialized views	Audit trails, full change history, CDC
Prerequisites	None	Continuous backup enabled, NoSQL API only

💡 Exam tip: soft-delete pattern

In Latest Version mode (default), deletes are NOT in the change feed. The standard workaround is the soft-delete pattern:

1. Instead of deleting, set a deleted: true flag and a TTL value
2. The change feed captures this as an update (the flag change)
3. Your processor sees the deleted: true flag and reacts accordingly
4. The TTL auto-removes the document after the specified period

This is a very common exam question — know the soft-delete workaround.

Starting position

Option	Description
From beginning	Process all historical changes — useful for rebuilding a materialized view
From now	Only process new changes — default for Azure Functions
From specific time	Resume from a point-in-time — disaster recovery scenarios
From continuation token	Resume from exact position — programmatic checkpointing

Parallel processing and scaling

The change feed is partitioned by physical partition. Each physical partition can be processed by exactly one consumer instance at a time.

Container with 6 physical partitions:
  Instance A: processes partitions 1, 2
  Instance B: processes partitions 3, 4
  Instance C: processes partitions 5, 6

If Instance B crashes:
  Instance A: processes partitions 1, 2, 3  ← acquires B's lease
  Instance C: processes partitions 4, 5, 6  ← acquires B's lease

Scaling rule: You can have at most as many consumer instances as physical partitions. More instances than partitions means some sit idle.

🎬 Video walkthrough

🎬 Video coming soon

Change Feed Processing — DP-420 Module 15

Change Feed Processing — DP-420 Module 15

~16 min

Flashcards

Question

What does the lease container do in change feed processing?

Click or press Enter to reveal answer

Answer

It tracks processing state — one lease per physical partition. It enables checkpointing (resume after crash), parallel processing (distribute partitions across instances), and load balancing (redistribute leases if an instance fails). Must be in the same Cosmos DB account.

Click to flip back

💡 Show hint

Question

Does the default change feed mode capture deletes?

Click or press Enter to reveal answer

Answer

No — Latest Version mode only captures inserts and updates. For deletes, use either: (1) the soft-delete pattern (set a deleted flag + TTL), or (2) All Versions and Deletes mode (requires continuous backup and new containers).

Click to flip back

💡 Show hint

Question

What is the maximum number of change feed consumer instances for a container?

Click or press Enter to reveal answer

Answer

Equal to the number of physical partitions. Each physical partition's changes are processed by exactly one instance at a time. Extra instances beyond the partition count sit idle, waiting to acquire a lease.

Click to flip back

💡 Show hint

Question

What happens if a change feed processor instance crashes mid-batch?

Click or press Enter to reveal answer

Answer

The batch is retried from the last checkpoint. The lease is released and acquired by another instance (or the restarted instance). No changes are lost because the change feed is persistent and the lease tracks the last successful position.

Click to flip back

💡 Show hint

Knowledge Check

Knowledge Check

Amara's SensorFlow container has 8 physical partitions. She deploys 12 Azure Functions instances to process the change feed. How many instances will actively process changes?

AAll 12 — each processes a share of every partitionB8 — one per physical partition; 4 instances sit idleC1 — Azure Functions uses a single instance for change feedD6 — Azure Functions caps at 6 concurrent instances

Check Answer

Knowledge Check

Amara needs to detect when sensors are physically removed (deleted from the database). She's using the default change feed mode. What's the best approach?

ARead the change feed for delete eventsBImplement soft-delete: set a 'deleted' flag and TTL, then watch for the flag in the change feedCUse the conflict feed to detect deleted itemsDQuery the container periodically for missing device IDs

Check Answer

Knowledge Check

Amara wants to rebuild a materialized view from scratch by reprocessing all historical changes. How should she configure the change feed processor?

ASet StartTime to DateTime.NowBSet StartTime to DateTime.MinValue.ToUniversalTime() (from beginning)CDelete and recreate the lease containerDUse a pull model query against the container directly

Check Answer

---

Next up: Analytical Workloads — how Synapse Link and Fabric Mirroring bring analytics to your Cosmos DB data without impacting transactional performance.