Best practices for large-scale IoT device deployments

Scaling an IoT solution to millions of devices can be challenging. Large-scale solutions often need to be designed in accordance with service and subscription limits. When customers use Azure IoT Device Provisioning Service, they use it in combination with other Azure IoT platform services and components, such as IoT Hub and Azure IoT device SDKs. This article describes best practices, patterns, and sample code you can incorporate in your design to take advantage of these services and allow your deployments to scale out. By following these patterns and practices starting from the design phase of the project, you can maximize the performance of your IoT devices.

Provision new devices

First-time provisioning is the process of onboarding a device for the first time as a part of an IoT solution. When working with large-scale deployments, it's important to schedule the provisioning process to avoid overload situations caused by all the devices attempting to connect at the same time.

Use a staggered provisioning schedule

For deployment of devices in the scale of millions, registering all the devices at once may result in the DPS instance being overwhelmed due to throttling (HTTP response code 429, Too Many Requests) and a failure to register your devices. To prevent such throttling, use a staggered registration schedule for the devices. Configure your device registration batch sizes in accordance with DPS quotas and limits. For instance, if the registration rate is 200 devices per minute, the batch size for onboarding would be 200 devices per batch.

Retry operations

If transient faults occur due to a service being busy, retry logic enables devices to successfully connect to the IoT cloud. However, a large number of retries could further degrade a busy service that's running close to or at its capacity. As with any Azure service, you should implement an intelligent retry mechanism with exponential backoff. More information on different retry patterns can be found in the retry design pattern and transient fault handling.

Rather than immediately retrying a deployment when throttled, wait until the time specified in the retry-after header. If there's no retry header available from the service, this algorithm can help achieve a smoother device onboarding experience:

min_retry_delay_msec = 1000
max_retry_delay_msec = (1.0 / <load>) * <T> * 1000
max_random_jitter_msec = max_retry_delay_msec

With this logic, devices delay reconnecting for a random amount of time, between min_retry_delay_msec and max_retry_delay_msec. The maximum retry delay is calculated with the following variables:

  • <load> is a configurable factor with values > 0, which indicates that the load will perform at an average of load time multiplied by the number of connections per second
  • <T> is the absolute minimum time to cold boot the devices (calculated as T = N / cps where N is the total number of devices and cps is the service limit for number of connections per second).

For more information on the timing of retry operations, see Retry timing.

Reprovision devices

Reprovisioning is the process where a device needs to be provisioned to an IoT Hub after having been successfully connected previously. There can be many reasons that result in a need for a device to reconnect to an IoT Hub, such as:

  • A device could reboot due to power outage, loss in network connectivity, geo-relocation, firmware updates, factory reset, or certificate key rotation.
  • The IoT Hub instance could be unavailable due to an unplanned IoT Hub outage.

You shouldn't need to go through the provisioning process every time a device reboots. Most devices that are reprovisioned end up connected to the same IoT hub. Instead, a device should attempt to connect to its IoT hub directly using the information that was cached from a previous successful connection.

Devices that can store a connection string

Devices that have the ability to store their connection string after initial provisioning should do so and attempt to reconnect directly to IoT Hub after reboot. This pattern reduces the latency in successfully connecting to the appropriate IoT Hub. There are two possible cases here:

  • The IoT Hub to connect upon device reboot is the same as the previously connected IoT Hub.

    The connection string retrieved from the cache should work fine and the device can reconnect to the same endpoint. No need for a fresh start for the provisioning process.

  • The IoT Hub to connect upon device reboot is different from the previously connected IoT Hub.

    The connection string stored in memory is inaccurate. Attempting to connect to the same endpoint won't be successful and so the retry mechanism for the IoT Hub connection is triggered. Once the threshold for the IoT Hub connection failure is reached, the retry mechanism automatically triggers a fresh start to the provisioning process.

Devices that can't store a connection string

Some devices don't have a large enough footprint or memory to accommodate caching of the connection string from a past successful IoT Hub connection. These devices need to reprovision through DPS after rebooting. Use the DPS registration API to re-register. Keep in mind that the number of re-registrations per minute is limited based on the DPS device registration limit.

Reprovisioning sample

The code examples in this section show a class for reading to and writing from the device cache, followed by code that attempts to reconnect a device to the IoT Hub if a connection string is found and reprovision through DPS if it isn't.

using Newtonsoft.Json;
using System;
using System.Collections.Generic;
using System.IO;
using System.Linq;
using System.Text;

namespace ProvisioningCache
  public class ProvisioningDetailsFileStorage : IProvisioningDetailCache
    private string dataDirectory = null;

    public ProvisioningDetailsFileStorage()
      dataDirectory = Environment.GetEnvironmentVariable("ProvisioningDetailsDataDirectory");

    public ProvisioningResponse GetProvisioningDetailResponseFromCache(string registrationId)
          var provisioningResponseFile = File.ReadAllText(Path.Combine(dataDirectory, registrationId));

          ProvisioningResponse response = JsonConvert.DeserializeObject<ProvisioningResponse>(provisioningResponseFile);

          return response;
      catch (Exception ex)
        return null;

    public void SetProvisioningDetailResponse(string registrationId, ProvisioningResponse provisioningDetails)
      var provisioningDetailsJson = JsonConvert.SerializeObject(provisioningDetails);

      File.WriteAllText(Path.Combine(dataDirectory, registrationId), provisioningDetailsJson);

You could use code similar to the following to determine how to proceed with reconnecting a device after determining whether there's connection info in the cache:

IProvisioningDetailCache provisioningDetailCache = new ProvisioningDetailsFileStorage();

var provisioningDetails = provisioningDetailCache.GetProvisioningDetailResponseFromCache(registrationId);

// If no info is available in cache, go through DPS for provisioning
if(provisioningDetails == null)
  logger.LogInformation($"Initializing the device provisioning client...");
  using var transport = new ProvisioningTransportHandlerAmqp();
  ProvisioningDeviceClient provClient = ProvisioningDeviceClient.Create(dpsEndpoint, dpsScopeId, security, transport);
  logger.LogInformation($"Initialized for registration Id {security.GetRegistrationID()}.");
  logger.LogInformation("Registering with the device provisioning service... ");

  // This method will attempt to retry in case of a transient fault
  DeviceRegistrationResult result = await registerDevice(provClient);
  provisioningDetails = new ProvisioningResponse() { iotHubHostName = result.AssignedHub, deviceId = result.DeviceId };
  provisioningDetailCache.SetProvisioningDetailResponse(registrationId, provisioningDetails);

// If there was IoT Hub info from previous provisioning in the cache, try connecting to the IoT Hub directly
// If trying to connect to the IoT Hub returns status 429, make sure to retry operation honoring
//   the retry-after header
// If trying to connect to the IoT Hub returns a 500-series server error, have an exponential backoff with
//   at least 5 seconds of wait-time
// For all response codes 429 and 5xx, reprovision through DPS
// Ideally, you should also support a method to manually trigger provisioning on demand
if (provisioningDetails != null)
  logger.LogInformation($"Device {provisioningDetails.deviceId} registered to {provisioningDetails.iotHubHostName}.");
  logger.LogInformation("Creating TPM authentication for IoT Hub...");
  IAuthenticationMethod auth = new DeviceAuthenticationWithTpm(provisioningDetails.deviceId, security);
  logger.LogInformation($"Testing the provisioned device with IoT Hub...");
  DeviceClient iotClient = DeviceClient.Create(provisioningDetails.iotHubHostName, auth, TransportType.Amqp);
  logger.LogInformation($"Registering the Method Call back for Reprovisioning...");
  await iotClient.SetMethodHandlerAsync("Reprovision",reprovisionDirectMethodCallback, iotClient);

  // Now you should start a thread into this method and do your business while the DeviceClient is still connected
  await startBackgroundWork(iotClient);
  logger.LogInformation("Wait until closed...");

  // Wait until the app unloads or is cancelled
  var cts = new CancellationTokenSource();
  AssemblyLoadContext.Default.Unloading += (ctx) => cts.Cancel();
  Console.CancelKeyPress += (sender, cpe) => cts.Cancel();

  await WhenCancelled(cts.Token);
  await iotClient.CloseAsync();

IoT Hub connectivity considerations

Any single IoT hub is limited to 1 million devices plus modules. If you plan to have more than a million devices, cap the number of devices to 1 million per hub and add hubs as needed when increasing the scale of your deployment. For more information, see IoT Hub quotas. If you have plans for more than a million devices and you need to support them in a specific region (such as in an EU region for data residency requirements), you can contact us to ensure that the region you're deploying to has the capacity to support your current and future scale.

When connecting to IoT Hub via DPS, devices should use the following logic in response to error codes when connecting:

  • When receiving any of the 500-series of server error responses, retry the connection using either cached credentials or the results of a Device Registration Status Lookup API call.
  • When receiving 401, Unauthorized or 403, Forbidden or 404, Not Found, perform a full re-registration by calling the DPS registration API.

At any time, devices should be capable of responding to a user-initiated reprovisioning command.

If devices get disconnected from IoT Hub, devices should try to reconnect directly to the same IoT Hub for 15-30 minutes before attempting to go back to DPS.

Other IoT Hub scenarios when using DPS:

  • IoT Hub failover: Devices should continue to work as connection information shouldn't change and logic is in place to retry the connection once the hub is available again.
  • Change of IoT Hub: Assigning devices to a different IoT Hub should be done by using a custom allocation policy.
  • Retry IoT Hub connection: You shouldn't use an aggressive retry strategy. Instead, allow a gap of at least a minute before a retry.
  • IoT Hub partitions: If your device strategy leans heavily on telemetry, the number of device-to-cloud partitions should be increased.

Monitor devices

An important part of the overall deployment is monitoring the solution end-to-end to make sure that the system is performing appropriately. There are several ways to monitor the health of a service for large-scale deployment of IoT devices. The following patterns have proven effective in monitoring the service:

  • Create an application to query each enrollment group on a DPS instance, get the total devices registered to that group, and then aggregate the numbers from across various enrollment groups. This number provides an exact count of the devices that are currently registered via DPS and can be used to monitor the state of the service.
  • Monitor device registrations over a specific period. For instance, monitor registration rates for a DPS instance over the prior five days. Note that this approach only provides an approximate figure and is also capped to a time period.

Next steps