This article is contributed. See the original author and article here.
We continue to expand the Azure Marketplace ecosystem. For this volume, 115 new offers successfully met the onboarding criteria and went live. See details of the new offers below: | ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
This article is contributed. See the original author and article here.
This public preview of Microsoft Azure Active Directory (Azure AD) custom security attributes and user attributes in ABAC (Attribute Based Access Control) conditions builds on the previous public preview of ABAC conditions for Azure Storage. Azure AD custom security attributes (custom attributes, here after) are key-value pairs that can be defined in Azure AD and assigned to Azure AD objects, such as users, service principals (Enterprise Applications) and Azure managed identities. Using custom attributes, you can add business-specific information, such as the user’s cost center or the business unit that owns an enterprise application, and allow specific users to manage those attributes. User attributes can be used in ABAC conditions in Azure Role Assignments to achieve even more fine-grained access control than resource attributes alone. Azure AD custom security attributes require Azure AD Premium licenses.
We created the custom attributes feature based on the feedback we received for managing attributes in Azure AD and ABAC conditions in Azure Role Assignments:
Let’s take a quick look at how you can manage attributes, use them to filter Azure AD objects, and scale access control in Azure.
The first step is to create an attribute set, which is a collection of related attributes. For example, you can create an attribute set called “marketing” to refer to the attributes related to the marketing department. The second step is to define the attributes inside the attribute set and the characteristics of the attribute set. For example, only pre-defined values are allowed for an attribute and whether an attribute can be assigned a single value or multiple values. In this example, there are three values for the project attribute—Cascade, Baker, and Skagit—and a user can be assigned only one of the three values. The picture below illustrates the above example.
Once attributes are defined, they can be assigned to users, enterprise applications, and Azure managed identities.
Once you assign attributes, users or applications can be filtered using attributes. For example, you can query all enterprise applications with a sensitivity level equal to high.
There are four Azure AD built-in roles that are available to manage attributes.
By default, Global Administrators and Global Readers are not able to create, read, or update the attributes. Global Administrators or Privileged Role Administrators need to assign the attribute management roles to other users, or to themselves, to manage attributes. You can assign these four roles at the tenant or attribute set scope. Assigning the roles at tenant scope allows you to delegate the management of all attribute sets. Assigning the roles at the attribute set scope allows you to delegate the management of the specific attribute set. Let me explain with an example.
Let’s build on our fictional example from the previous blog post on ABAC conditions in Azure Role Assignments. Bob is an Azure subscription owner for the sales team at Contoso Corporation, a home improvement chain that sells items across lighting, appliances, and thousands of other categories. Daily sales reports across these categories are stored in an Azure storage container for that day (2021-03-24, for example); so, the central finance team members can more easily access the reports. Charlie is the sales manager for the lighting category and needs to be able to read the sales reports for the lighting category in any storage container, but not other categories.
With resource attributes (for example, blob index tags) alone, Bob needs to create one role assignment for Charlie and add a condition to restrict read access to blobs with a blob index tag “category = lighting”. Bob needs to create as many role assignments as there are users like Charlie. With user attributes along with resource attributes, Bob can create one role assignment, with all users in an Azure AD group, and add an ABAC condition that requires a user’s category attribute value to match the blob’s category tag value. Xia, Azure AD Admin, creates an attribute set “contosocentralfinance” and assigns Bob the Azure AD Attribute Definition Administrator and Attribute Assignment Administrator roles for the attribute set; giving Bob the least privilege he needs to do his job. The picture below illustrates the scenario.
Bob writes the following condition in ABAC condition builder using user and resource attributes:
To summarize, user attributes, resource attributes, and ABAC conditions allow you to manage access to millions of Azure storage blobs with as few as one role assignment!
Since attributes can contain sensitive information and allow or deny access, activity related to defining, assigning, and unassigning attributes is recorded in Azure AD Audit logs. You can use PowerShell or Microsoft Graph APIs in addition to the portal to manage and automate tasks related to attributes. You can use Azure CLI, PowerShell, or Azure Resource Manager templates and Azure REST APIs to manage ABAC conditions in Azure Role Assignments.
We have several examples with sample conditions to help you get started. The Contoso corporation example demonstrates how ABAC conditions can scale access control for scenarios related to Azure storage blobs. You can read the Azure AD docs, how-to’s, and troubleshooting guides to get started.
We look forward to hearing your feedback on Azure AD custom security attributes and ABAC conditions for Azure storage. Stay tuned to this blog to learn about how you can use custom security attributes in Azure AD Conditional Access. We welcome your input and ideas for future scenarios.
Learn more about Microsoft identity:
This article is contributed. See the original author and article here.
Perhaps no one has been hit harder over the past 20 months than small businesses. To adapt and thrive in this new normal, small businesses need comprehensive solutions that are designed specifically for them and their unique needs.
The post New Microsoft Teams Essentials is built for small businesses appeared first on Microsoft 365 Blog.
Brought to you by Dr. Ware, Microsoft Office 365 Silver Partner, Charleston SC.
This article is contributed. See the original author and article here.
Microsoft Ignite 2021 took place online November 2-5. This fall edition was full of dev news, and if you don’t want to miss anything related to App development and innovation, keep reading!
Session Highlights
What does it take to build the next innovative app?
During the Digital and App Innovation Into Focus session, Ashmi Chokshi and Developer and IT guests Amanda Silver, Donovan Brown and Rick Claus discussed processes and strategies to help deliver innovative capabilities faster.
Developers are driving innovation everywhere, and Ashmi started the conversation strong sharing how she sees the opportunity to drive impact. Then, Donovan presented his definition of cloud native, the benefits of a microservices architecture, and engaged in a discussion with Rick around DevOps and Chaos Engineering. This session also discussed how to transform and modernize your existing .NET and Java applications. Amanda Silver concluded with a demo showing how tools like GitHub Actions, Codespaces and Playwright can help with development, testing and CI/CD, no matter what language and framework you are using.
The sketch below illustrates the cloud native and DevOps segment showcasing the new public preview of Azure Container apps, a fully managed serverless container service built for microservices that scales dynamically based on HTTP traffic, events or long-running background jobs.
To dive deeper into the latest innovation on containers and serverless to create microservices application on Azure, don’t miss Jeff Hollan and Phil Gibson’s session where they demoed Azure Container apps and the Open Service Mesh (OSM) add-on for Azure Kubernetes Service, a lightweight and extensible cloud native open-source service mesh built on the CNCF Envoy project. Brendan Burns, Microsoft CVP of Azure Compute, also shared his views on how Microsoft empowers developers to innovate with cloud-native and open source on Azure in this blog.
Another highlight of Ignite was the Build secure app with collaborative DevSecOps practice session, followed by the Ask-the-expert, where Jessica Deen and Lavanya Kasarabada introduced a complete development solution that enables development teams to securely deliver cloud-native apps at DevOps speed with deep integrations between GitHub and Azure.
Announcements recap
In addition to Azure Container apps and Open Service Mesh add-on for AKS, we also announced new functionalities for Azure Communication Services, API management, Logic Apps, Azure Web PubSub, Java on Azure container platforms and DevOps.
The complete line up of Azure Application development sessions and blogs is listed below:
On-demand sessions:
Blog-posts:
Additional learn resources:
We can’t wait to see what you create!
This article is contributed. See the original author and article here.
I had a customer streaming messages at a high-rate (up to 2000 msg/s – 1KB each) from a protocol translator running on an x86 industrial PC to a cloud-based Mosquito MQTT broker.
That edge device evolved quickly into a more capable and secure intelligent edge solution thanks to Azure IoT Edge and Azure IoT Hub, adding device provisioning (secured with an HW-based identity) and device management capabilities on top of a bi-directional communication, along with the deployment, execution, and monitoring of other edge workloads in addition to a containerized version of the original MQTT protocol translator.
The performance requirement did not change though: the protocol translator (running now as an Azure IoT Edge module) still had to ingest and deliver to the IoT Hub up to 2000 msg/s (1KB each), with a minimum latency.
Is it feasible? Can an IoT Edge solution stream 2000 msg/s or even higher rates? What’s the upper limit? How to minimize the latency?
This blog post will guide you through a detailed analysis of the pitfalls and bottlenecks when dealing with high-rate streams, to eventually show you how to optimize your IoT Edge solution and meet and exceed your performance requirements in terms of rate, throughput, and latency.
Topics covered:
The Azure IoT Edge runtime includes a module named edgeHub, acting as a local proxy for the IoT Hub and as a message broker for the connected devices and modules.
The edgeHub supports extended offline operation: if the connection to the IoT Hub is lost, edgeHub saves messages and twin updates in a local message queue (aka “store&forward” queue). Once the connection is re-established, it synchronizes all the data with the IoT Hub.
The environment variable “UsePersistentStorage” controls whether the message queue is:
When persisted on disk (default), the location of the queue will be:
The size of the queue is not capped, and it will grow as long as the device has storage capacity.
When dealing with a high message rate over a long period, the queue size could easily exceed the available storage capacity and cause the OS crash.
Binding the edgeHub folder to a dedicated partition, or even a dedicated disk if available, would protect the OS from uncontrolled growth of the edgeHub queue.
If the “DATA” partition (or disk) runs out of space:
Do size the partition for the worst-case or reduce the Time-To-Live (TTL).
I will let you judge what’s the worst case in your scenario. But the very worst case is total disconnection for the entire TTL. During the TTL (which is 7200 s = 2 hrs. by default), the queue will accumulate all the incoming messages at a given rate and size. And be aware that the edgeHub keeps one queue per endpoint and per priority.
An estimation of the queue size on disk would be:
And if you do the math, a “high” rate of 2000 [msg/s] with 1 [KB/msg] could easily consume almost 15 GBs of storage within 2 hrs.
But even a “low” 100 [msg/s] rate you could easily consume up to 1GB, which would be an issue on constrained devices with an embedded storage of few GBs.
Then, to keep the disk consumption under control:
If you keep the queue disk consumption under control with proper estimation and sizing, you don’t need to bind it to a dedicated partition/disk. But it’s an extra-precaution that comes with almost no effort.
Btw: If you considered setting UsePersistentStorage=false to store the queue in memory, you may realize now that the amount of RAM needed would make it an expensive option if compared to disk or non-viable at all. Moreover, such an in-memory store would NOT be resilient to unexpected crashes or reboots (as the “EnableNonPersistentStorageBackup” can backup and restore the in-memory queue only when you go through a graceful shutdown and reboot).
What happens to expired messages?
Expired messages are removed every 30 minutes by default, but you can tune that interval using this environment variable MessageCleanupIntervalSecs:
If you use different priorities (LINK), do set “CheckEntireQueueOnCleanup”=true to force a deep clean-up and make sure that all expired messages are removed, regardless of the priority.
Why? The edgeHub keeps one queue per endpoint and per priority (but not per TTL)
If you have 2 routes with the same endpoint and priority but different TTL, those messages will be put in the same queue. In that case, it is possible that the messages with different TTLs are interleaved in that queue. When the cleanup processor runs, by default, it checks the messages on the head of each queue to see if they are expired. It does not check the entire queue, to keep the cleanup process as light as possible. If you want it to clean up all messages with an expired TTL, you can set the flag CheckEntireQueueOnCleanup to true.
Now that you have the disk consumption of your edgeHub queue under control, it’s a good practice to keep it monitored using the edgeAgent and edgeHub built-in metrics and the Azure Monitor integration.
The “edgeAgent_available_disk_space_bytes” reports the amount of space left on the disk.
…but there’s another metric you should pay attention to, which is counting the number of non-expired messages still in the queue (i.e. not delivered yet to the destination endpoint):
That “edgehub_queue_length” is a revelation, and it explains how the latency relates to the rates. But to understand it, we must measure the message rate along the pipeline first.
I developed IotEdgePerf, a simple framework including:
Further instructions in the IotEdgePerf GitHub repo.
I deployed the IotEdgePerf transmitter module to an IoT Edge 1.2 (instructions here) running on a DS2-v2 VM and connected to a 1xS3 IoT Hub. I launched the test from the console app as follows:
dotnet run --
--payload-length=1024
--burst-length=50000
--target-rate=2000
Here are the results:
As anticipated, the “edgehub_queue_length” explains why we have the latency. Let’s have a look at it using Log Analytics:
Let’s correlate the queue length with the transmission burst: as the queue is a FIFO (First-In-First-Out), the last message produced by the transmitter is the last message ingested by the IoT Hub. Looking at “edgehub_queue_length” data, the latency on the last message is 42 seconds.
How does the queue’s growth and degrowth slopes and maximum value relate to the message rate?
which is in line with what you would expect from a queue, where the growth rate:
The consistency among the different measurements (on the message rate, queue growth/degrowth and latency) proves that the methodology and tools are correct.
Using some simple math, we can express the latency as:
where N is the number of messages.
If we apply that equation to the numbers we measured, again we can get a perfect match:
The latency will be minimum when rateOUT = rateIN, i.e. when the upstream rate equals the source output rate, and the queue does not accumulate messages. This is quite an obvious outcome, but now you have a methodology and tools to measure and relate to each other the rates, the latency, and the queue length (and the disk consumption as well).
Let’s go back to the original goal of a sustained 2000 msg/s rate delivered upstream with minimum latency. We are now able to measure both the source output and the upstream rate, and to tell what’s the performance gap we must fill to assure minimum latency:
But… how to fill that gap? What’s the bottleneck? Do I need a faster CPU or more cores? More RAM? Or the networking is the bottleneck? Or are we hitting some throttling limits on the IoT Hub?
Let’s try with a more “powerful” hardware.
Even if IoT Edge will usually run on a physical HW, let’s use IoT Edge on an Azure VMs, which provide a convenient way of testing different sizes in a repeatable way and compare the results consistently.
I measured the baseline performance of the DSx v2 VMs (general purpose with premium storage) sending 300K messages 1KB each using IotEdgePerf:
VM SIZE | SPECS (vCPU/RAM/SCORE) | Source [msg/s] | Upstream [msg/s] |
Standard_DS1_v2 | 1 vcpu / 3.5 GB / ~20 | 900 ÷ 1300 | 500 ÷ 600 |
Standard_DS2_v2 | 2 vcpu / 7 GB / ~40 | ||
Standard_DS3_v2 | 4 vcpu / 7 GB / ~75 | ||
Standard_DS4_v2 | 8 vcpu / 14 GB / ~140 | ||
Standard_DS5_v2 | 16 vcpu / 56 GB / ~300 |
(test conditions: 1xS3 IoT Hub unit, this C# transmitter module, 300K msg, msg size 1KB)
Scaling UP from DS1 to DS5, the source rate increases by around ~50% from a DS1 to DS5… which is peanuts if we consider that a DS5 performs ~15x better (scores here) and costs ~16x (prices here) than a DS1.
Even more interestingly, the upstream rate does not increase, suggesting there’s a weak correlation (or no correlation at all) with the HW specs.
Let’s distribute the source stream across multiple modules in a kind of scaling OUT, and look at the aggregated rate produced by all the modules.
The maximum aggregated source is ~1900 msg/s rate (obtained with N=3 modules), which is higher than the rate of a single module (~1260 msg/s), with a gain of ~50%. However, such improvement is not worth it if we consider the higher complexity of distributing the source stream across multiple source modules.
Interestingly, the upstream rate increases from ~600 to ~1436 msg/s. Why?
The edgeHub can either use the AMQP or the MQTT protocol to communicate upstream with the cloud, independently from protocols used by downstream devices. The AMQP protocol provides multiplexing capabilities that allow the edgeHub for combining multiple downstream logical connections (i.e. the many modules) into a single upstream connection, which is more efficient and leads to the rate improvement that we measured. AMQP is indeed the default upstream protocol and the one I used in this test. This also confirms that the upstream rate is mostly determined by the protocol stack overhead.
Scaling UP from a 1 vCPU machine (DS1) to a 16 vCPU (DS5) machine didn’t help when using a single source module. But what if we have multiple source modules? Would many cores bring an advantage?
Yes. Multiple modules mean multiple docker containers and eventually multiple processes, that will run more efficiently on a multi-core machine.
But is the 3.5x boost worth the 16x price increase of a DS5 vs DS1? No, if you consider that the upstream, again, didn’t increase.
Let’s go back to a single module and increase the message size from 1 KB to 32 KB on a DS1v2 (1 vcpu, 3.5GB). Here are the results:
(test conditions: 1xS3 IoT Hub unit, this C# transmitter module)
The rate decreases from 839msg/s@1KB down to 227msg/s@32KB, but the THROUGHPUT increases from ~0.8 MB/s up to ~7.2MB/s. Such behavior suggests that sending big messages and at a lower rate is more efficient!
How to leverage it?
Let’s assume that the big 16KB message is a batch of 16 x 1 KB messages: it means that the ~4.5MB/s throughput would be equivalent to a message rate of ~4500 msg/s of 1KB each.
Then message rate and throughput is not the same thing. If the transport protocol (and networking) is the bottleneck, higher throughputs can be achieved by sending bigger messages at a lower rate.
How do we implement message batching then?
We have two options:
The default is MaxUpstreamBatchSize = 10, meaning that some batching is already happening under the cover, even if you didn’t realize it. The optimal value for MaxUpstreamBatchSize would be 256 KB / size(msg), as you would want to fit as many small messages as you can in the batch message.
How does it work?
Set the edgeHub RuntimeLogLevel to DEBUG and look for lines containing “obtained next batch for endpoint iothub”:
Looking at the timestamps, you’ll see that messages are collected and sent upstream in a batch every 20ms, suggesting that:
Are the application-level and built-in batching equivalent?
On the UPSTREAM side, you have batch messages (i.e. big and low-rate) in both cases. It means that:
The latter point is quite interesting: as the message batch counts as 1 device-to-cloud operation, the batching helps also reduce the pressure on the d2c throttling limit of the IoT Hub, which is an attention point when dealing with high message rates.
On the SOURCE side, the two approaches are very different:
Eventually, we can state that:
Let’s test that assumption.
On a DS2v2 VM, the max output rate of the source module is ~1100 [msg/s] (with 1KB messages).
As expected, such source rate does not benefit from higher MaxUpstreamBatchSize values, while the upstream does, and it eventually equals the 1100 [msg/s] source rate (hence no latency).
The application-level batching is effective on both the source and upstream throughput.
On a DS1 (1 vCPU, 3.5GB of RAM, which is the smallest size of the DSv2 family) you can achieve:
Is latency always a problem? Not actually. As we saw, the latency is proportional to the number of messages sent (N):
When sending a short burst of messages, the latency may be negligible. On the other hand, on short burst you may want to minimize the transmission duration.
As an example, let’s assume you want to send N=1000 messages, of 1 KB each. Depending on the batching, the transmission duration at the source side will be:
Shortening the transmission duration from a 1.1s down to 0.1s could be critical in battery powered applications, or when the device must send some information upon an unexpected loss of power.
The performance results discussed in this blog post were obtained using this C# module, which leverages the .NET Azure IoT Device SDK.
How do other SDKs (Python, Node.js, Java, C) perform in terms of maximum message rate and throughput? The performance gap, if any, would be due to:
As an example of the latter, the JAVA SDK differs from other SDKs as the device/module client embeds a queue and a periodic thread that checks for queued messages. Both the thread period (“receivePeriodInMilliseconds”) and the number of messages de-queued per execution (“SetMaxMessagesSentPerThread”) can be tweaked to increase the output rate.
On the other hand, what is the common denominator to all the SDKs? It’s the transport protocol, which ultimately sets the upper bound of the achievable performance. In this blog post we focused on MQTT, and it would be worth to explore the performance upper bound by using a fast and lightweight MQTT client instead of the SDK. That’s doable and it’s explained here.
A performance comparison among the different Device SDKs as well as a MQTT client will be the topic for a next blog post.
A bash script to spin up a VM with Ubuntu Server 18.04 and a fully provisioned Azure IoT Edge ready-to-go: https://github.com/arlotito/vm-iotedge-provision
A framework and a CLI tool to measure the rate, the throughput and end-to-end latency of an Azure IoT Edge device:
https://github.com/arlotito/IotEdgePerf
This blog post provided a detailed analysis of the pitfalls and bottlenecks when dealing with high-rate streams, and showed you how to optimize your IoT Edge solution to meet and exceed your performance requirements in terms of rate, throughput, and latency.
On an Azure DS1v2 Virtual Machine, we were able to meet and exceed the original performance target of 2000 msg/s (1KB each) and minimum latency, and we achieved a sustained end-2-end throughput of 3600 KB/s with no latency, or up to 7300 msg/s (1KB each) with some latency.
With the methodology and tools discussed in this blog post, you can assess the performance baseline on your specific platform and eventually optimize it using the built-in or application-level message batching.
In a nutshell:
Special thanks to the Azure IoT Edge team (Venkat Yalla and Varun Puranik), the Industry Solutions team (Simone Banchieri, Stefano Causarano and Franco Salmoiraghi), the IoT CSU (Vitaliy Slepakov and Michiel van Schaik) and Olivier Bloch for the support and the many inspiring conversations.
Recent Comments