Tag Archives: Cloud

Achieving Enterprise Agility at the Retail Edge

At any conference, you start to feel fatigued near the end of the day. Between the sprawling session locations, the constant engagement, and the sometimes overwhelming number of people and places to navigate, attention begins to wan by the last sessions you attend. Presenters in the unfortunate end-of-day time slots are between attendees and dinner. I was pleasantly surprised to be reinvigorated after attending my second to last session of the day. Presenters Connie Yu and Joana Cruz, of Target, managed to bring humor, excitement, and energy to their topic, titled: Achieve Enterprise Agility at Edge Locations by adopting Kubernetes, Spinnaker and Unimatrix.

My background is in software defined infrastructure orchestration at scale, both in a data center environment, as well as in my company’s private cloud deployment. I was curious to see this type of orchestration in use by a retailer to manage their in store deployments. Connie and Joana structured their session as a humorous conversation on how the stores could adopt supply chain methodologies to improve operations. By doing so, Target went from delivering software to 1850 stores just 3 times a year to multiple times daily. How? Let’s dive in!

Target’s Strategy:

Have you ever installed an operating system and all of the software you want to use on it from scratch? Most consumers buy an off the shelf system that has already been configured for their baseline needs. In an enterprise environment, IT is (usually) responsible for installing and configuring the “stack” — the operating system and all necessary software, in addition to modifying settings to secure and optimize the system, and they have to do this for thousands of systems, not just one. If you had to hand build thousands of systems, given your experience having done it just for one, then you know, it can be tedious, and error-prone work. The time spent to automate the process is well worth the investment as it saves countless hours and makes deployment repeatable, predictable, and scalable. To do this, you need the right tools, and you also need to take the time to consider the architecture to ensure it will meet the requirements.

There are unique infrastructure constraints in a retail store setting. The primary compute resources are thin clients, which have less I/O, Memory, Storage and CPU than you would find in a traditional data center. For this reason, some of the conventional architectures, like server:client segmentation, have to be redesigned to run at the edge. Target presented three primary components of their edge computing solution.

Key components of Target’s edge computing strategy

Microservices are lightweight and loosely coupled which makes the architecture a good fit for distributed computing with limited resources, as is the case in the edge retail environment. Containers allow application dependencies, packages, and software to be fully encapsulated so that they run in an isolated environment side by side with other containers — a logical choice for consolidating services on your infrastructure without disturbing functionality. With microservices, they are typically implemented using REST which often leverages HTTP, and securing all of the data flowing between the services on different endpoints is crucial to maintaining security — no store wants to become the next data breach cautionary tale.


Orchestration, in brief, is the practice of automating the configuration and management of your infrastructure. Whether you have three systems or thousands, managing them is far easier when you have the right tools. Target chose the open source orchestration platform Kubernetes, as their orchestration agent.

Using Kubernetes for orchestration

The primary reasons cited for this choice were: self healing, state maintenance, and high availability. Often, the task of running and scheduling of workloads is segmented onto two different servers. In this case, Target containerized the applications and co-located them on the same host and had a cluster of servers that provided the functionality, thereby achieving consolidation and redundancy.


The deployment tool of choice for Target was Spinnaker. I had never heard of it before. If you’re like me you had to Google it. I’ll save you the trouble, here is the Cliff’s Notes version: It is an open source tool that supports multiple cloud providers and integrates with many existing orchestration tools (like Kubernetes).

Using Spinnaker to deploy

Spinnaker uses the concepts of applications, clusters, and server groups to construct and manage deployment piplines. In short, it allows you to automate which target devices you want to install your software on.

Edge Cloud

One of the points Connie and Joana mentioned was that Target had few corproate datacenters, but 1,855 stores all with their own infrastructure. They wanted to manage each store like its own cloud, and so they needed an edge cloud provider.

Target named their edge cloud provider Unimatrix, which is a nod to Star Trek. Unimatrix commands and controls all of the Kubernetes deployments at the edge.

Edge Cloud Provider: Unimatrix

Unimatrix reuses the existing Kubernetes API, and allows deployment by pilot, mass roll out, or even by store regions. Maintaining the API contract and only adding additional microservices to augment custom requirements allowed Target to continue using the existing Spinnaker and Kubernetes integration. An additional component is the Unimatrix agent. The agent takes the role of the operator. It applies the application spec to the Kubernetes clusters. As the spec is applied, it reports back to Unimatrix, which syncs the state to Spinnaker so the developer can see the deployment progress. You can read more in depth about Unimatrix on the Target blog.

Demoing the Use case: Deploy a promotion

Now that we know all of the components of the system, we can see the process in action.

Process flow: deploying a promotion
  1. All promotion change features implemented as microservices
  2. Push code to Git
  3. Security tools triggered to ensure
  4. Build process ensues
  5. Code goes into docker registry
  6. The Spinnaker pipeline is triggered, and pushes the application spec to the Unimatrix central server
  7. The Unimatrix deployment server begins to sync application spec to the agents in the defined pipelines

As all of this is happening, the state of the deployment is maintained and can be visualized in Spinnaker:

Deployment progress in Spinnaker UI — the green is complete, red is still in progress

And that is how it’s done!


So their teams did all of that work. What was the result?

  • On boarded > 90 applications in the first 6 months; which included IoT platforms, new microservices, and video machine learning capabilities.
  • Running in production at Target edge locations

How did they accomplish this success? The answer — defining very clear objectives from day one:

  1. The ability to deploy to stores in a targeted fashion (pun intended)
  2. Enable application developers to do the deployments
  3. Have clear measurements of progress
  4. Reusing tools allowed developers to rapidly adopt the technology
  5. Flexibility during the solution building — Target accelerated development by using open source, and created custom solutions only when necessary

Not to be biased, but…this was my FAVORITE GHC session. I loved the humorous exchanges between Joana and Connie, they were well rehearsed, knew the material, the flow was logically presented and really took the audience on the journey of WHY developing this technology was necessary, how they implemented the solution, and the return on the investment. Also, might I add, the speakers had the best red tennis shoes, very on brand.

Living on the Edge

With the growing number of intelligent devices (and by “intelligent” I mean: An internet connected device running some form of software that collects and performs actions on data), bringing intelligence to the Edge has become a more pervasive topic of discussion. What is “the Edge?” Computing devices physically located at the point of use running software that is running locally instead of in the cloud. The cloud? Remotely hosted computing services such as storage (think: Google Drive) and infrastructure (virtual machines) — these services depend on a network connection. This sets the stage for the session: Developing Embedded Intelligence: Opportunities on the Edge. It was led by a panelist of experts in the field:

Brenda Zhuong of MathWorks (moderator), Miriam Leeser of Northeastern University, Micheala Blott of Xilinx Research, Mary Ann Maher of SoftMEMS LLC , and Yan Wan from the University of Texas at Arlington.

Panelists for the Embedded Intelligence session

The high level considerations that were illuminated by the panel of experts were:

  • Improving intelligence through data analytics
  • Gaining efficiency by taking advantage of the hardware capabilities
  • Reducing dependency on the (internet) network by hosting data and services where they will be used.

We then got into some real-world use cases that demonstrate the power of harnessing intelligence at the Edge.

Gain Efficiency in Embedded Intelligence; Michaela Blott

Reduce cost through custom arithmetic

Urban Aerial Vehicle (UAV)-based airborne computing for future Internet of Things; Yan Wan

UAV-based Emergency Communication

Developing an Artificial Kidney; Mary Ann Maher

Artificial Kidney


  • Seed neural networks with data coming from the actual environment so you can determine when sensors are malfunctioning — not all data is good data.
  • Understand power constraints — battery life has been a limitation
  • With all of the data being collected, privacy becomes paramount — always encrypting data at rest and in transit.

Some best practices…

If you plan to develop in this space, or are already working on a project, there are methodologies you can put into practice that will improve your solution. Here are a few:

  • There is so much sensor data, it is not possible, nor would it be advisable, to send all of it back to the cloud, which means the capability to do some level of analysis at the edge and send only the relevant data to the cloud is
  • Take advantage of cloud computing to simulate sensor data to rapidly design but then deploy at the edge.
  • Tagging quality data so you can establish a confidence in the sensor data with corresponding visualization (graphical representation) is crucial.
  • IoT devices tend not to have as much computing power, leveraging distributed computing models to perform calculations
  • Implement redundancy of sensors and computational devices in mission critical systems

Where do we go from here?

Computer vision and artificial intelligence are changing the landscape of computing. We have to improve the accuracy of vision sensors so that the data is reliable. The cost of quality vision sensors is prohibitive at the moment, and before development can really take off, it must come down. Research into correctness proofs is needed to verify algorithms developed for mission critical Edge workloads. Safety, especially for use cases like autonomous driving, is not a 90% correctness solution — it’s okay if your cell phone drops its connection, it is deadly if your car stops in the middle of an intersection. Portable IoT devices like ultrasound and water testing can bring life-saving technology to network constrained environments, and have the potential to revolutionize medicine.

The panelists were asked to make predictions for the next three years in Edge Intelligence (If any of these things come to pass, you heard it here first!) Some of them being:

  • Speech recognition as a human computer interface will become much more prevalent
  • Facial recognition will be integrated into more technology
  • Deployments of sensors and devices in rural areas will help to eradicate diseases
  • More innovative hardware platforms for IoT use cases

Recommended reading: