Will Kubernetes Be the Answer to Edge Computing?

Key Takeaways

[Music] hey it's Alex Williams the new staff welcome to the new sack makers a podcast where we talked about application development deployment and management at scale the coop com+ cloud native conferences gather adopters and technologists to further the education advancement of cloud native computing the vendor-neutral event features domain experts and key maintainer behind popular projects like kubernetes prometheus envoy corps dns container d and more [Music] okay here we are for another episode of the new stack makers here at KU com+ cloud native con in San Diego it's november 2019 and i'm and join this the specter of this conversation it's a discussion which i think is interesting it's looking at really how we think about the edge and how we think about kubernetes at the edge and how we think about what we often hear is IOT and i try to avoid that term but i but i think of it more in terms of well we'll get to that but first let me before i get to that point i'd like to introduce our guests steve steven wong is a software engineer cloud native business unit and vmware and diablo sonic is a senior software engineer red hat he's a member of the apache software foundation a committer at the Eclipse Foundation I set this topic guys it's not a car to know kubernetes at the edge and I think of I I'm fascinated by this idea of you know how everything we think of as you know something real is actually something made of something else in many ways they constitute some it constitutes many kind of different composition of many different types of things and so we could actually classify and label it as something that's other than what we think of it as like a table or a car whatever it might be and one of the fascinations to me has always been about distributed computing is how we now have all those compute storage networking capability for applications but they're mostly focused inside of a data center or even in a cloud is because it's really kind of a combination of multiples of data centers and as we discussed cloud native technologies and kubernetes we've started to migrate towards this discussion about what is the context of the world beyond the data center and how do we bring the clusters to those places out at the edge where the objects and need to be thought of much differently than just connecting to a computer or a SmartWatch dream on a phone so I'm how is kubernetes entered this discussion how is cloud Niihau cloud native technologies introduced well sure I think one way to look at ads you first define it you know and your car analogy is spot-on because there's two aspects of it you know a lot of the luxury models of cars now have as many as 37 computers inside them now they might be tiny computers but you've got things like the entertainment system the anti-lock brakes the traction control even the power doors and windows often have a little controller associated with the motors that drive those now when you say these are objects there's two things that you need to do with these objects one their sources of data potentially that somebody somewhere in the world would like to know number two they often need configuration and updates and you know kubernetes originally was developed as a system using a control plane to control vast numbers of containerized applications in a public cloud under circumstances where you didn't want humans doing that with their hands on you did it remotely but the architects of kubernetes started by building a foundation for this rich and as it turns out extensible api that can control things at massive scale through an API where you state the desired outcome that I want these things to be in this condition in the perfect world and then things called controllers work in the background to relentlessly drive the state of the real world they measure they take note of what you said you wanted measure what this thing is now and relentlessly drive it towards your stated goal and they don't do it with just one loop they enabled it so it was paralyzed Abul so you could have thousands of these things working simultaneously driving the world to this because it was extensible it means that it's not necessarily confined to controlling containerized applications two years ago I found on YouTube preparing for a talk I'm giving here a slide deck where somebody used the exact use case of a temperature sensor declaring whether you wanted this temperature sensor to measure in degrees Fahrenheit or Celsius you know and I think it was just to prove a point simply that you know this is how you do one of these custom resource definitions but kubernetes has the potential to be a toolkit to automate managing these things out there at edge and by the way it isn't just consumer objects it's potentially a lot of the people Dion and I represent a working group within kubernetes a fair amount of the people who come there are trying to do remote office retail stores fast-food retail outlets things like that and kind of control a little shop that is by no means a data center you know it doesn't have racks of computers but that also comes under here yeah isn't that the first front really those those industrial objects and perhaps maybe he gives some context for kubernetes addition maybe even the shortcomings that need to be addressed to be able to make these irrelevant just to extend on the Steve's points so so kubernetes is there but I think cloud native is really a key enabler in all these edge computing so we'll stop maybe a little bit later about hypes and everything and I think everything will be labeled edge at some point but I think so if adding new people try to put compute resources on you know did you have it for a long time for four decades right right but but what what change is now and the same thing like when we care the transition from machine to machine and suddenly everything was calling IOT right so so now every no system will try to label themselves as edge but to me edge is really just the extension of the cloud so we need a cloud native architecture on the back insecurities but we need also all the tools for developers that they use to build applications that can run only in cloud to be able to deploy it on all these edge sites as well so it's it's a whole so to say continuum of platform and and developers tools that will enable to do these things on a massive scale that people want to do it doesn't this get into kind of your space that you are an expert in I would assume to say in terms of connectivity and messaging yeah yeah and that's a lot of your work at the clip yeah so it eclipsed confirmation for the last couple of years between we try to build basic fundamentals for for a cloud IOT cloud platform that will be open source and based on open standards and and it's going pretty well but even even with that we see we see that the things are changing so so for example even if you have a very good connectivity platform for the cloud right we can support today millions of devices connecting to the platform but but it's just getting grower getting more and more and the solution and one of the things that edge computing is trying to solve is to bring those compute resources closer to the source to the things that the device is instead of all of them connecting to the central clouds so what we want to do next is exactly that so try to really take that on the good conscience that we have today and and try and try to distribute these architectures so that we can have a local connectivity for the devices may be some local processing there so saving the bandwidth improving the latency or all of these things that you usually talk about in the context of edge computing if I can step in I can give an example I agree with Dion that for example take an example of edge devices being security cameras that are maybe collecting 30 or more frames per second of video you don't really want to upload all of that video to the cloud if what you're really interested in is collecting license-plate numbers or the names of the people who walked in the door so there's a use case where you that data the frame-by-frame videos coming from that camera has a gravity that would like to pole processing towards it machine learning image recognition and tagging out towards where the data was collected because there's less that can go wrong there's lower latency and you get your results sooner and frankly you save bandwidth you didn't have to send the massive amount of information up to the cloud if you could process it and greatly reduce it by three orders of magnitude by reducing it out at the edge and there's ultimately those things those machine learning apps and things that are associated with these sensors really are containerized apps like kubernetes is used to managing in a way you could argue that edge computing really is cloud computing you've just inverted the place where these containerized apps are going to run oh it does help that if it does help if you can integrate management of dispatching those containers running them monitoring their health with managing of the sensors and and the actuators out there at the edge along with managing the network interconnects and I'm not saying kubernetes fully has this down but that's the point of our working group to get discussions going to do a barn raising it's the best foundation we have today and we shouldn't build on top of that to provide a better experience in the future how's the container technology gonna have to adapt to that you know these are I mean if we're talking about edge you know let's let's just you know talking about an industrial environment for example I mean are these containers the same containers that are running in a datacenter that are running out at the edge they smaller in size are they lighter weight what are the what are the kind of considerations yeah did they add a couple of things did that people should be aware of I think yeah a resource constraint is always a big big concern because one thing that you don't have on the edge is auto scaling which is a big promise of the cloud computing right you need more resources you click upload buttons you know put your credit card and you know problem solved on the on the site you have probably you know a limited number of nodes and it's very expensive to get additional resources there so developers need to be more more careful about how they write those up you some services need to be critical higher priority some services are you know a less priority could be evicted so those are some of the concerns but in the end we have those tools so those containers those containers applications should be able to run in the same way in the cloud or in the edge it's just a deployment option that it should take some consideration if you you know thinking of putting them there a couple things that maybe are different I Deon's point is valid that frankly even in a big public cloud people do the wrong thing and take a VM and literally move it into a container and it's this one terabyte container and that's always a mistake but you can cheat if you're willing to pay for in a public cloud and it flat out blows up if you try to do that at edge so you got to do it the right way but it shows up what do you mean well you just don't have rack after rack of computer so you have a finite amount and when it runs out it just doesn't run if you demand this monstrous amount in AWS your bill goes up but it runs if you just pay a lot of money for it it just doesn't run out at edge because there's nothing to back it up well why not okay so let me play devil's advocate then why not a micro VM instead of well no you just there it isn't even just the money thing it's often the physical size even the nature of processors you know one difference in the containers are that I think there's a bigger demand out at edge for things like arm containers you know a standard docker container image is actually CPU specific and for it's not even because the CPUs are cheaper but you've got situations where you don't have air conditioning and you maybe are constrained to a small physical size if you're building a traffic light controller or something and you just picture them as they're running on something more akin to a cell phone than a rack mount server and I think one more thing to answer that question is developer because they they want to have a seamless I think if you're gonna do edge successfully to attract developers and all developers are now in the cloud native and they want to use the same tools in same practices they use today to develop those applications in my mind it should be like that like you should be able to run writers micro-services as you write them right and then you can deploy them wherever you want you should should it be like a cloud only deployment or or should you be able to move some of the micro services closer to to on some nodes close or or some clusters closer to due to the devices and in the end the data that's what you know cloud cloud native computing gives us and that's why we need our cube relatives and and all these things we have today to be able to do that tomorrow I'll tell you another aspect you need to keep in mind that's kind of the next layer above the container runtime and the container images themselves and that is the registry it comes from you know typically people architect these apps in a public cloud to pull the images out of say docker hub or something and you can it's essentially guaranteed you can get there from here when it's running in a public cloud but a lot of these edge applications have very intermittent connectivity to the Internet the bandwidth that's there can go from zero to poor and never to good and it means that you have this consideration in many of these where you want to build up a cache or an image registry on-premises so that your pre-populating and pulling it from there you know let me give you an example of a retail store suffers a power outage all the computers reboot and they come up and the internet connection is lost or it's so poor that if every every computer in a big retail store goes and tries to pull images at once it's sucking it through a little soda straw and it's just gonna take hours and none of the cash registers in the store can come out that would be bad so you need a solution the cash is that kind of stuff and there's even some like I've heard of cruise ship lines and things that are actually using containers and they only dock in port once over a few days but they can to work properly if you properly design Heather's restaurant chains for using you know clusters and inside the actual stores themselves Mike you know I want to get back to what you know you bring up to is this intermittency issue and that gets back to the connectivity issue and the messaging issue and I know you've been working on some projects in particular you event-driven messaging and activemq Artemis for example and how is this architecture better suited you know to to these new types of environments so there's a lot of yeah I think messaging in general in IOT is a basic communication pattern between because that's the normal way of how things communicate sending messages back and forth right so so what we've done so far is trying to for example in couple this project try to put some more semantics on top of those communication it's going on so you know trying to narrow down basically because general messaging is too broad and people reinvent every time and make civil applications because everybody's solving the same problem in a different ways so we try to do a little bit more to constrain them in terms of you know in IOT and and in these kind of environments you have two basic semantics so sending some kind of telemetry from the from the edge to the cloud and sending commands back back to two devices so so you try to create a couple of API it should be it should allow people to more easily reuse them and create something that's interchangeable and integratable instead of everybody solving that their own problem on their own but another thing as you said is what we're doing right now is trying to do a technology that we coined so far as a virtual application networking and and and that's kind of trying to solve these this multi cluster or or hybrid cloud communication in a better way and and that problem is is particularly visible in India India generally where you have a lot of all these edged sites are usually in the Private Networks not which not reachable from from the outside world so we want to make sure that that our service is running on the edge and on the cloud can communicate in in a duplex way we want to connect to services running in two different private private clouds so what we did is is try to apply networking principles on delayed layer seven completely on the on the application layer layer so that mitigates a lot of problems in trying to set up VPNs firewalls and all that kind of kind of things manually so so that's one of the ways the layer 7 case because this this gets into like the autonomous networks versus networks that are you know the traditionally that we know that might be running in a data center and calling out yeah to you know to a known for example now the nodes may necessarily have to communicate with each other right right I I think it's fair to say that in the bulk of these most of the traffic by volume is leaf node to leaf node and you really don't want one leaf node here pinning up to a public cloud and coming back because it's not for people who are not really familiar with what is a leaf node well if you've got an edge picture it like a tree where you say branch from the central cloud out to the branches and that the tips are leaves and you might have hundred thousand leaps and I'm saying that in many of these applications a lot of these Leafs want to communicate with something nearby some other sensor nearby like a motor is trigger needs to talk to a sensor to decide when to turn on and they're at the same physical location in the same city in the same block but you don't want the sensor signal to go way up to a cloud in Seattle thing you know from Pittsburgh then come back down to the motor in hits Berg that's just not going to work out well for you for a lot of reasons so that's what I mean by leave to leave so you want some sort of networking facility where maybe a central control plane is desirable but it needs to keep out of the data plane so that you don't inject latency and you don't have an unnecessarily long path security is also a big issue here like when he raised the issue of l7 and typically up at l7 HTTP people use things like TLS certificates and some aspects of that don't necessarily work well at edge because really when you look at these certificates they get installed in a device that you presume you have physical security over somebody can't go there and steal it but with a lot of these edge devices if somebody was to grab the device and the certificate with it and that gave them a ticket on to an internal corporate network that is really bad I hear rumors that they've been people burned by not thinking of these aspects of the design but by going to specialized protocols designed and incapable of getting generalized generally routable around the world they might actually be that might actually be a feature rather than a deficit because you know they'd be incapable of getting to places generally you don't want them to be capable of some of this actually it's interesting like Dion brought up MQTT and arguably one of the hottest things here at cube con is this idea of service mesh envoy proxies where the idea is that the average app developer isn't a networking expert or a security expert and if they don't know well they're gonna make mistakes and you suffer really Ike would contend that that old MQTT that's been around a while was whether they knew it or not or called it that at the time did that same thing of throwing you know relieving the app developer of needing to know networking it's a publish/subscribe thing where when you want to publish this sensor information you just give it a label you don't have to keep track of how many people are subscribing that's kicked off on the underlying communication platform and you know some of these tools actually were very well done and are very close to what seems to feels like new things like sto and envoy up in this data center world but they're also close enough that I think a lot of these concepts of service mash in a central control plane may become backfill and put a facade layer on this kind of legacy stuff in the IOT space deliver great value then it gets into you know the next generation of connectivity and messaging doesn't it again some of the some of the concepts I've heard about is that you know there's we have these computer storage and networking capabilities in these massive data centers now the computer almost sometimes seems you know endless and the networking capabilities are what we never possibly could be dreamed of I would never we'd never could have dreamed of and and there's you know storage is just an afterthought almost right so there's these vast resources that are available in these you know these hyper data center environments and what is then going to be I mean one of the questions I have I guess is kubernetes architectures evolved for developers and building out applications for you know for for phones and for you know you know civil computers or whatever how do we think about how the data center will evolve because one of the this is where I one of the kind of the concepts I think is a just about functions right now you know the role of functions and what role functions will play it helped me kind of deliver almost intelligence out to the out to these nodes so then the nodes can then communicate with each other based upon that you know the intelligence are being sent to them you know so one of the things we're trying to solve with this is maybe which maybe you can help here is if if we move the service discovery and adjustability to the right layer seven we get much more flexibility of where our services live and it is you're like papa you know like a publisher's I've kind of manner we can we can layer on top of that much more complex messaging protocols then it's a simple mode balancing and request reply that we here with HTTP and DNS is a solution seen in the service message so there's a lot of you know a lot of things that could be done to mitigate those those kind of kind of issues and then when you have this addressability solved then you can you can do what what you're saying moving functions moving services all around and and making these distributed data centers one important aspect there is that you know so we've been talking so far about the edge from the IOT perspective and and and retail perspective but there's an a big group of people interesting in this and these are the service providers telcos trakai is peace but they look at the edge from the completely different angle they see their own point of presence kind of infrastructure that ology exists and say why don't we use that to host some of these services and and things like that but they'll have these constraints they can put put it out yeah yeah the telcos are showing up to our working group meetings in huge numbers I thought I I think maybe that might be literally all of them in the world I mean I've certainly seen all of them I know of showing up to the meetings that's and it's not a concept of yeah if you have this dream of how does the data center and you're picturing something like the AWS cloud or they Azure cloud or the Google cloud how does it change I think what changes is there's more clouds hosted out towards edge locations you know what if you envision the world that the data gravity pulls demand for hosting things not in a couple or a handful of places in the world but in every major city of over 200,000 people and I think really that's what's going to happen and change that you're going to have a lot more data centers now there's no reason they can't be powered and managed by kubernetes but a picture world where there's more of them and it's not just me saying this I think I encountered an analyst forecast saying that by 2022 70% of the processing workloads that process data generated at the edge will actually be running at the edge and it was predicting that even though we've had this hockey stick growth of new workloads being deployed to these big central public clouds for the past five years that we're about to enter a world in LTE five is a big enabler of this of big deployments out in edge locations now the telco the telcos sometimes have the ability to host multi tenant clouds out at any location question is you know what will the developers want to do you know so right I think that's gonna be the biggest question is because they can say yeah they have those they also had the dial tone they also have the infrastructure for you know for for the internet but you know you know what happened I mean well you use the old dial tone analogy which is interesting and it's a throwback I'm I'm old enough to remember that I don't know if Dion is but one of the concepts is that you know in the 90s when the world moved from dial-up modem to broadband it was a big bandwidth improvement but the real thing that disrupted society was the fact you were always on I mean I remember when you went through your household saying does anybody need to make a phone call or does anyone expect one and it goes off the air and once you are always connected it rapidly changed to every member of the family has their own device or a couple and you expect them to be always on well LT five is like that where it actually is going to be a big enabler of say three orders of magnitude more devices and it's modeling of connections makes it in it makes it economically viable to have some really cheap device that just squirts out a signal saying this light switch switched and it only has to send it once every two weeks and you don't have to like pay for you know a data account on a monthly basis we're gonna have to finish up here in a minute so John I would like to just get from you and then Steve maybe you could conclude what are what are the projects you're working on now what what is you know what is the conversation you're having koukin what is it following kook on that you that's on the horizon yeah so Steve and I represent the the cube net is ith working group here and and we're trying to build a community of people you know interested in in general this this topic and as you see from this conversation the topic is so huge yeah and so wide that you can attack it from any angle yeah and so far we're having really good discussions but everywhere you look you scratch a little bit and then you open a completely new field so being there networking being the developer tooling and things like that so me particularly I'm continue working mostly in the connectivity so so working for the IOT connectivity and and working this scupper project virtual application networks that you try to enable distributing this community distributing this connectivity closer to the edge and making a better experience for for this edge edge edge world in in terms of messaging and networking so that's my mostly been my focus going on and what also sorry would like to do and what I'll talk about today mostly so far we were talking about infrastructure about platform how to get communities there we still don't talk much about developers and I think that's a very important topic so I would like to encourage people to start taking those discussion more seriously how we have a set of those playground yeah how we get up on the edge reviews or all these kind of stuff in this new environment so I'm here more not to represent my company but to represent the IOT edge working group and I guess my campaign is to facilitate the whole community doing this and we're gonna have to probably ramp from one topic to the other we just finished a security white paper I think we're talking at Q Khan this time about something to coalesce the unique protocols and communication protocols found out an edge one of the things that I'm campaigning is this idea of actually extending the kubernetes control plane through C RDS to put a facade on legacy stuff one of the issues with this whole space for industrial IOT in particular is that people buy that kind of stuff on 30 20 to 30 year life cycles they don't replace it like a rackmount server and the data center probably has a lifetime of three to five years and you change it out so if you had to rip and replace all your existing legacy that's not a great solution a lot of that stuff has been out there and uses proprietary api's unique to one vendor communication gateways but I think there's a real opportunity to go put a kubernetes control plane and data collection mechanism on top of that stuff and enable the end users of these things to do new things without ripping out all of the old stuff I think that's really a key lesson to be learned even with containerization and kubernetes that yeah there's greenfield apps but the real big value thing is when you can bridge greenfield with legacy and get the whole thing to work because nobody does a big bang roughly the entire that well thank you very much to the to you for joining us today thank you it's been a conversation I look forward to continuing I'm sure this is just the start of these conversations on how we think about the world beyond the data center into these environments where we're looking at you know how we think about networking how we think about you know these developer spaces that can be created and how we think about these other issues like you're talking about in terms of you know how we think about containers in these environments and such so thank you very much Steve Wong software engineer cloud native business unit VMware and Dion Boston a senior software engineer Red Hat member of the pipe patchy Sound Software Foundation and a committer at the Eclipse Foundation thank you both for your time thank you thank you [Music] the coupe com+ cloud native con conferences gather adopters and technologists to further the education advancement of cloud native computing the vendor-neutral event features domain experts and key maintainer x' behind popular projects like kubernetes prometheus envoy for dns container d and more [Music] listen to more episodes of the new stack makers at the new stack io / podcasts please rate here review us on iTunes like us on YouTube and follow us on soundcloud thanks for listening and see you next time [Music] you