Kubernetes IoT Edge Working Group, 4 Sep 2020

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Kubernetes IoT Edge Working Group: Applications at the Edge - Cindy Xing & Steve Wong

Description

Don’t miss out! Join us at our upcoming events: EnvoyCon Virtual on October 15 and KubeCon + CloudNativeCon North America 2020 Virtual from November 17-20. Learn more at https://kubecon.io. The conferences feature presentations from developers and end users of Kubernetes, Prometheus, Envoy, and all of the other CNCF-hosted projects.

Kubernetes IoT Edge Working Group: Applications at the Edge - Cindy Xing, Microsoft & Steve Wong, VMware

This session will address unique considerations of edge applications. Edge workloads on Kubernetes can vary from: - Simple IoT solutions over Raspberry-pi - Video processing, image recognition, or machine learning apps on embedded device or server - Complicated industrial IoT or Teleco solutions requiring high availability, security, and reliability We will survey how edge impacts best practices of app development and operational management - and how Kubernetes and other related open source tools can be utilized for edge workloads. Agenda - Characteristics and requirements of different Edge Applications - Language / runtime considerations - Operational considerations with Kubernetes - App development, build, packaging and CI / CD considerations - Available tooling - How to get involved in the IoT Edge Working Group

https://sched.co/ZeuP

A

Hi this session is presented by the kubernetes iot edge working group. This is about using kubernetes for edge applications, I'm steve wong, a tech lead of the working group. I work on kubernetes and a few other open source projects as an employee of vmware. Today, I'm joined by cindy.

B

Hello, this is cindy singh, I'm co-chair of this kubernetes iot edge working group. Currently I work at microsoft as a principal software engineering manager, especially I'm working on azure iot edge, so we can move to the agenda so before we talk about edge applications, let's first address what edge is about, then we're gonna explain in detail about edge applications, characteristics, operational consideration and toolings, and after that steve will drill down to serverless at the edge in detail.

B

In the end, we'll share contact information and the link to this deck. I hope you will like the talk.

B

So what is edge computing? What's special about it? Edge computing means compute close to where data is originally generated. It has high security requirements, data locality and compliance need to be maintained.

B

The scale of edge devices can be in the range of thousands, millions or tens of millions. There are tons of data, and some of them is real time edge. Computing is very diversified and heterogeneous.

B

Hardware operating system and underlying infrastructure can be quite different. Local resources can be in form of raspberry, pi, server, machine or cluster of machines.

B

From use case perspective, it can be teleco, 5g, industrial, iot, retail, connected cars and etc. Edge needs to coordinate with cloud, and it also requires local autonomy. Ideally, zero touch, operation with observability and remote management are desired.

B

Now, let's talk about edge applications, characteristics and requirements.

B

When comparing edge applications with cloud ones, they have similar natures, for example, high availability, reliability, scalability performance and etc. From securities perspective, the requirements for edge applications are much higher.

B

Imagine when running workload at remote edge locations, one would want to encrypt data before storing it to the disk. Some private secrets are better to be held at hardware level, and each application should attest itself at the runtime.

B

Given edge, devices can be resource constrained together with the needs of cost saving edge. Applications need to be lightweight and controlled in resource consumption.

B

Some edge applications like video streaming gaming or machine learning. They need to process real-time data with high performance requirement.

B

The amount of data processed will be huge, and some of the data has to be local, not leaving the device with edge devices and applications talking to each other identity, provisioning and credential management need to be enabled edge, applications are highly distributed and they run under complicated network topology.

B

Here, let's take a deeper look: a raspberry pi 3 has 1 gigabyte memory, and the intel server has 4 gigabyte.

B

Each application can be written in c, rust, c, plus dot net golan node.js java, a lot of languages operating system can be linux, rubendo, santos or windows or android.

B

Underlying infrastructure can be bare metal, virtual machine, k3s macro k8s q batch or full kubernetes cluster based on different use cases and scenarios. One will have to pick and choose.

B

Now, let's see the operational consideration running a kubernetes cluster at the edge will have a lot of challenges. For example, we need to evaluate what kind of app model to be used in the industry. There are docker, kubernetes, helm and open application model.

B

Each provides apis and command line tools. Docker is straightforward, though it has limited capabilities.

B

Kubernetes is very powerful, but complicated helm chart allows customers to describe the workflow of complicated application deployment.

B

Ohm was targeted to separate application, components and operational configuration and make the deployment easier.

B

Now how about application itself? How will workload be operated? It can run as vm carter, container docker native process or webassemb assembly.

B

Some legacy applications cannot be containerized or really hard to be containerized. The operating system may not support container. In that case, you might have to use host process. Some workload needs to enable higher level security isolation.

B

In that case, you probably want to use virtual machine or cutter container. Web assembly is a new way for process isolation. It has drawn a lot of attention in the industry.

B

Next, let's say where to store those applications, how about application packaging story? How to access the application repository, the application repository can be container registry. The registry need to be deployed and then from edge node perspective. It needs to access this container registry so how to access and where to store the access credentials.

B

If the network bandwidth between edge and cloud is limited, do we need to deploy a local container registry as a mirror from packaging perspective? Is it in form of darker image, tar file or synapse bundle.

B

Edge applications are distributed, then we also need to think about different programming model to enable this highly distributed.

B

Nature reactive model, like a java spring, is widely used in the java community.

B

Recently, google and microsoft announced k native and dapper to allow developers run author and operate event driven actor model in distributed fashion at edge specially for dapper. It enables applications to achieve network routing statement, management and some other capabilities. The model is through a site. Car digital twin is a modeling language.

B

It allows developers to digitalize the connected things, so all those can be leveraged, but you need to think about which one to use ash applications they run remote to achieve remote management and observability metrics logging tracing collections are needed.

B

Applications also need to be versioned and updated when application is in form of container one can verify from cloud or desktop and deploy to edge without worrying. However, in case a remote access and debugger will be helpful at the edge look host level, there are operating system firmware, driver container, runtime, those are host level processes.

B

They need to update it. So then the question is: what will be the mechanism.

B

Obvious to author host build and verify applications. There are tools needed, github and git are widely adopted for source control and parallel programming for code editor and debugger. There are vs code, intellij, eclipse, golen, etc. Personally, I found golan is pretty helpful. If you need to code in go and jenkins normally is used or widely adopted for cicd.

B

Now I'm going to switch back to steve to cover serverless and edge steve.

A

Thank you cindy, as cindy pointed out, absent edge can be in various languages and forms. Next, I'm going to talk about something that has become popular in the public clouds, I'm going to talk about serverless and whether this makes sense at the edge.

A

So what is serverless? Well, the term started out with the announcement of aws lambda in a 2014 reinvent keynote by werner vogels. There's been steady growth in interest. As you can see in this chart, the term serverless group drew a lot of ridicule at first, because obviously there are still servers and what I'm going to talk about isn't just about the original lambda. It was the first.

A

But now all the major clouds have something similar things like cloud functions on google cloud function, compute on alibaba, azure functions and others there's some healthy, open source projects like k, native openfast and others out there and they're designed to support functions as a service that you would run on your own. The cncf hosts a serverless working group. So, let's analyze, what this serverless concept is really about.

A

This is the lambda definition from the docs: let's analyze how it might apply if lambda was hosted at edge. A key difference at edge is that you likely own and operate the host infrastructure.

A

There may be some telco use cases that talk about hosting multi-tenant workloads in cell towers, but for the most part, I'm equating edge with on-premises servers owned by the user. I added some red coloring here uh to what's in the aws docs, to highlight some items that would be impacted by a move to edge at the bottom. This uh statement no charge when not running make sense in a public cloud as a business model decision by by the vendor, but run that same thing on an on-premises host.

A

That's yours and this no charge aspect goes away. You pay for it that without provisioning and managing statement is similar, whether it's an employee or you choose to outsource you're paying for that provisioning and management, and if you have resource constraints, the scales automatically part in red here might not apply to you either. So are we done here?

A

Serverless at edge doesn't make sense, not quite because, while that label server list got slapped on lambdas, there's a little more to it.

A

Moving deeper inside the lambda docs you'll find some additional aspects, some that are super valuable in edge applications. When you use lambda, you invoke code automatically in response to events, you can build up complete systems by wiring, cascading stages of event, consumption and generation.

A

This is really a tool kit that would enable you to build an event-driven architecture and to be fair, I went back and watched werner vogel's talk and it actually led with the value proposition of event driven computing server. Elimination came up later.

A

So, in a simple form, an event-driven architecture looks like this. An event is a piece of information, often used to communicate facts like measurements or commands, statements of intent, functions and services consume events, and it's up to these applications to decide what to do with each event. Events can be ignored, forwarded stored or transformed in a time series stream of events. Whether events are measurements or commands, a new one might make an older one irrelevant if a build-up occurs in a queue.

A

Analysis of events might even emit new events based on a transformation.

A

In a distributed system, an event stream is a communication fabric and when apps and event flows are viewed together, this is actually a programming model event. Driven architecture can be helpful because it encourages some good development practices. I'm talking about loose coupling independently maintainable components and separation of concerns.

A

A distribution layer here can help organize things and maybe even offload things you might otherwise have to write yourself in the application or application components. An example of this would be a pub sub broker, eliminating maintenance of a distribution list from the duties of an event, publisher.

A

So serverless in a public cloud is really about the producers pricing model, not physical reality. From a billing perspective, your service provider makes the server sort of invisible by not charging you for the instance, uh but let's examine this further. Let's have a pretend show of hands here now come on. I know it's a virtual conference, but just do it you've been sitting in that chair too long and you need a stretch break.

A

So I'm going to ask you a question how many of you have gone home on the weekend and had that moment, where you go? Oh crap did I leave my ec2 development instance running. You know there. There are scenarios where not paying for that server instance is attractive when you run in public cloud, so it's great for public cloud, but doing this at edge isn't really serverless at all. Still, let's not throw out the concept because of a bad label at edge.

A

You have a distributed system and these are really how hard to develop, troubleshoot and get to work right event. Driven computing is a really useful pattern at edge and we need every useful tool we can get our hands on.

A

So let's just throw away that serverless label, but we as developers know that the press, the marketing people and some of our bosses, like it's simple. They like a one-liner description of value and I've got one for them for for them.

A

Bear me bear with me here I'm going to do a stream of consciousness thing that will come together at the end. Let's go back in time. We begin in the 60s at that picture. On the top left the ibm mainframe was king. Large organizations could afford one of these installed at headquarters in a showcase, purpose-built room data originated from distant people, traveled by paper, transcribed by key punch operators onto cards.

A

Most people never came close to that computer. Maybe they were allowed to stare at it through a glass window. If they were located in the headquarters, the 70s saw lower priced mini computers installed in many dispersed locations. Most people still never came close enough to touch one but more did than before.

A

The personal computer revolution started in the 80s and eventually most offices, and even some wealthy households had one and shared it. Eventually. It got even cheaper and became something close to a one per person scenario. The trend has been that the compute has been getting closer and closer to the people producing and ingesting the information.

A

The cell phone took it a step further with devices literally at people's hands 24 by 7.. I submit that the next stage is that compute is now going to go ahead of the people. It can't get any closer, so the next step is to go ahead.

A

Apps, like machine learning, medical image processing, are landing at points in the data stream ahead of people to reduce latency and take on tasks. People didn't like or couldn't do very well. I propose that event. Driven at edge is a great concept, but let's re-label it from serverless to peoples.

A

This is about eliminating people in boring jobs like watching analog security, camera monitors and uh putting the compute up front.

A

There are some challenges here. Some frameworks are resource intensive for running on the smallest edge hardware. Canada, I'm talking about you open fast, is perhaps a little bit better, but that doesn't mean don't use these mater, maybe consider landing a framework like k-native at the gateway tier and using something similar to pipe data flows originating on the smaller devices mqtt. As an example, the zero to one latency might be an issue so perhaps configure your system to never go to zero instances.

A

After all, as I pointed out earlier, you're probably paying for the host 24x7 anyway.

A

State is a big issue. You know, lambdas in aws are stateless, but useful. Apps typically need to have some retained. States somewhere edge, isn't an exception to this rule I'll throw out a few thoughts here you know, update in place is dangerous, whether we're talking about device firmware or data, consider uh spending the money to keep at least two copies or full change logs.

A

An example of the danger here is breaking a device when power is interrupted, mid update uh in a 30 minute session. We just can't cover everything.

A

So if you find content like this useful cindy- and I want to invite you to become members of the kubernetes iot edge working group- we aren't writing code in kubernetes in this group, but we are focused on applying kubernetes, with additional open source tools to edge and iot use cases. We've got online meetings every two weeks at alternating times to accommodate members in different time zones, as shown here we encourage a member driven agenda and we oper we're also operating a group channel on the kubernetes slack.

A

Thank you for attending this session cindy and I are going to stick around for a few minutes to handle questions if you have any. uh But if you're watching this recording later, you can also reach out to us on the kubernetes edge iot slack channel.

B

Yeah again welcome to join our group. We are meeting regularly feel free to reach out to us. Thank you for joining us for this session.

A

And with that said, we're finished I'll turn it back over to the cncf organizers.