Cloud Native Computing Foundation KubeCon + CloudNativeCon Europe 2020 - Virtual, 4 Sep 2020

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Tutorial: KubeEdge Hands on Workshop -- Build Your Edge AI App on Real Ed... Zefeng Wang & Zhang Jie

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.

Tutorial: KubeEdge Hands on Workshop -- Build Your Edge AI App on Real Edge Devices - Zefeng Wang, Huawei & Zhang Jie, China Unicom

Here a hands-on KubeEdge workshop is proposed which ideally will run through the full length of the conference. This workshop is intend to invite participants to get hands on experience building a real edge computing solution with KubeEdge, end-to-end. Starting from deploying and provisioning an edge node(e.g. raspberry Pi), followed with device modeling and connectivity setup, then building a video stream machine learning based solution. Through this exercise, participants will get first hand experience to understand the orchestration engine build on top of kurbenetes, understand the edge computing node setup mechanism, learn the device modeling concept for IoT Edge scenarios. And develop a state-of-art AI based video stream processing flow, all in a 30 minutes session.

https://sched.co/Zep5

A

All right, yeah, I think my screen is sharing right.

A

I'm sorry there's some uh voice.

A

A

Is very limited comparing to the central cloud data center, while there are a lot of devices and also many data being produced and processed dna in the middle, it's basically the uh age of the core network, or uh we understand it's uh kind of age, of the the operator's network.

A

So in this location, the uh the applications might be different, like people want to run their uh live streaming while casting there or run some gaming like vr games, server, side there and also doing some online education.

A

So in this location, uh if we also look at the ai perspective,.

A

There are a lot of like transcoding um applications running there, so around thirty percent, ai inferencing will be down there because the way the more the most of the inferencing applications are running in the uh and the user side. All right, the essential cloud is a is we, I think we are quite extremely familiar today. The uh applications can be uh quite different, so we can run ai training as well as some some well knowing the uh inferencing in the data center and also we can do transcoding.

A

So one common understanding that is uh from the central cloud to the age, the uh the hash rate and the underlying hardware architecture are quite different. So in the central cloud today uh they are men uh more based down the x, 86 right and also, but uh on the edge they are different. So we know that the arm based hardware need less power. uh Supply supplements right and also there are some of the uh smaller, smaller xav6, otherwise can be run in there.

A

So the hashrate is quite different and also the underlying architecture is quite different.

A

So uh the major consideration why we build a age platform on top of kubernetes? Why we choose cloud native technology to build? Is that actually the? I think the benefits are quite well known already? The kubernetes are using containers as the core pieces of a application.

A

We know that containers we just need to build once and they can run everywhere and also the uh the container image have different multiple layers. So we can easily optimize the uh the base image layer to reduce the image size, as well as the system of overhead of the application and also kubernetes today.

A

We all believe that it's already a defender de facto standard of the application of instruction, especially in the cloud if we can use the cloud native technology to build the age computing platform and also build the uh the uh platform across cloud and age. We can get a very good experience that we manage the applications in the cloud and on the age with the uh gaming experience and also for kubernetes itself. um We know that all the applications are only talking to the api server right and also they have.

A

The uh managers themselves have the uh declarative api and also it supports multi, multiple groups of api and different api versions and also have the extensible functionality. Like crd, we can easily extend the api if we need some more uh new abstraction on the uh to run the applications on the cloud on the age and also for kubernetes itself. We can easily to add new components like we use the body operation operators for some customized applications in cloud today, but there are still some challenges.

A

So one big challenge is that, especially for industry and iot, the age nodes are, the results on the age are quite limited. So for a vanilla, kubernetes, even the lightweight distribution. We know that it needs less at least one gigabyte right and also like. Even even we try to uh install a cubelet on some of the agent notes.

A

The plan itself needs more than 100 megabytes.

A

While we know that on some gateway node on the edge in total, it may have like one 128 megabytes or 256 megabytes. So the operat, the operator operating system itself also needs some memory and also if we run the container runtime it takes.

A

For example, container d: it needs at least 17 megabytes, so the resource limitation is one of the biggest challenge to manage the applications on the edge and also the age and the cloud between cloud and age.

A

The network environment is quite different, so in many cases the um the age may locate in a private network or behind a firewall. It has very limited bandwidth to the central cloud and also the latency. Latency is quite a different in different cases.

A

It's not guaranteed latency, comparing to all the nodes, uh communicating with each other inside a data vendor or instead a uh inside a cloud, a cloud region right, so the network is quite unstable and also one of the big motivation of managing the applications on the age is that we want to deal with some of the business serving on the age right it it make it able to process some of the uh requests on the edge directly to save the safety entrance latency and besides that, we also want to uh the age, have some of the autonomy capabilities, so they can manage the application themselves.

A

I mean when some disaster operate and recovered. The age should be able to recover the applications themselves and not not rely on the central cloud to push to recover all the applications and also the age may.

A

In some cases the age may get offline for quite a long time, so it definitely uh it would break, if always rely on the central cloud and also on the edge. We know that the agent nodes have very different uh fine architecture.

A

Some are based on the arm and some are based on x86 and also especially for the iot devices and the industrial device. They are very different devices and also they have very different device protocols.

A

So, for the end users, if they want to develop the applications on the age, they need to deal with the multiple different device protocols. It adds the uh flexibility of developing a application, so kubernetes and cloud native technology today not yet have a very uh simplified solution of speeding up the application development.

A

Oh that's also the uh another big challenge.

A

So, uh with that background analysis, so cuba h, we think that first of all, we definitely want to build based on kubernetes, as well as the uh general cloud native net technology and besides that, for especially to the cloud and age coordination.

A

Cases we we think that we have four aspects of the enhancement. First of all, is uh cuba provide a seamless cloud and age coordination mechanism?

A

So, basically, even if the age nodes are located in prior uh private network or behind a firewall or if it doesn't have a public ib address it's okay, so uh the bi-directional communication terminal, it itself will uh deal with the uh the upstream and downstream uh synchronization, and it supports both both metadata and the user data and also for age autonomy.

A

Can we introduce the node level uh meter persistency?

A

So, even if it's disconnected to the cloud, the engine notes doesn't need to wait for the uh list watch response from the master it can recover by itself, so it it's uh getting ready faster and it saved the bandwidth between cloud and age by the way uh for the uh general architecture. There are two uh types in the in in the ecosystem, so one idea is that uh running the whole kinetic cluster or maybe a lightweight kubernetes cluster on the edge another one is putting the nodes on the edge.

A

So cubed started with running nodes on the edge, because we think that it's easier to do the.

A

Resource optimization for or the very limited use cases and it's easier to coordinate with the central cloud, okay for the uh resource restriction. So what cuba h optimized is that it removed some of the uh unrelevant functionalities like the cloud provider, storage integration, because on the edge we definitely have a very different storage type.

A

So also because we introduced the uh node level metadata persistency, so we kind of replaced the static part functionality to make it read from the uh local metadata so uh like in cubase. 1.3 itself needs around 70 megabytes when it started up and also uh cubase supports the standard uh ci interface.

A

So uh it's easy to uh integrate with container d, so container d need around uh another 17 megabytes. So then, uh if we add, the uh system offer overhead of the operating system, uh probably a agent can for the end user. They just need like 112, and they can have a lot of resources to run their own applications and also for the cio cryo project itself needs around 30 megabytes.

A

Then we can uh even run a run tube edge on an even smaller engine up. So that's quite available to especially the resource resource constraint situations and also for the age device. Communication coverage introduced a extensible framework to make users able to define their own device protocol so and also for the underlying framework.

A

Cuba provides a device method layer, so it helps convert from the actual device protocol to a general toggle. Currently, it's mqtt so for an application developers. They just need to rely on mqtt and they can talk to all the devices on the edge, so that uh helps a lot for the end users to build their applications on the edge, especially when dealing with multiple devices that consists of different protocols.

A

All right, so this is a very uh general overview of the cubase texture, so uh it it have uh basically three apart. So one is uh on the cloud, the uh central uh in the public cloud data center and also one is on the h and then otherwise about the devices.

A

So in the cloud, if we look on the left, it's quite similar to the vanilla kubernetes, so we know that companies manage all these resource resources as objects right. Any status change would finally reflect on the object status.

A

So what cuba h did is that we have a cloud core to deal with all the stuff on the edge and we do the shadow management inside kubernetes, so for kubernetes it can find some node marked with a label uh that is located on the edge.

A

But actually we know that the scheduler just to see the node objects, and they just see there are some parts to to get scheduled, so they will do as usual, and the cloud core will watch all the uh resources that are relevant to the age and it will help to dispatch all the uh updates to different agents.

A

So today it contains uh from a user user perspective, it contains uh age, controller, device, controller and cloud hub, but inside there are more other other modules I will introduce later the uh biodirectional communication terminal. Today we are relying on the messaging over websocket because we know that websocket just supports both apps streaming under downstream, with uh with one connection and also, we support the messaging over quick, the websocket. The implementation today is more reliable and has better uh performance, so it's uh the websocket is default now, all right, so at the age.

A

The most importantly, are the h chord. So, uh besides, like a lightweight equivalent, we also have a uh node-level metadata, persistency and also we have the uh where we actually have a heart. That's uh symmetric to the cloud hub to deal with the messaging stuff and the bio directional communication, and also for the container runtime integration. We support a standard cri, so it's easy to uh coordinate with either darker energy or cryo or other runtime, and also uh we support, uh because we support the cri.

A

We support a seminar and also for the csi uh it's easy to integrate with the ca csi on the edge uh cubase itself. We actually, we introduced some csi plugin to hook the request footage to so users can just integrate the existing third-party csi implementation and they can have the same experience.

A

Managing the storage on the edge and one of the big challenges is that how if we have a very big deployment and highly distributed so we may have uh our age nodes located in multiple age, say each site, so that means that the different age nodes may locate in different private networks.

A

The network environment is quite different with the uh those in the data center right in the data center in the cloud a to kubernetes. It's very it's a actually a very simple network. Each node can talk to each other directly based on the node ip because they are located inside of the same uh vvc right if you, if we run it in the uh public cloud, but then in the age computing cases, it's quite different that different the aging nodes may locate in different private network.

A

So first, therefore, the agent can not talk to each other because they even don't know which ip address is reachable to that node and also for the port. Networking the big challenge is how we, uh how we install the ci implementation, because almost all uh all the ci implementation today relies a relies on the data center networks and also on top of that, how to deal with the uh service and the other like uh less so uh today. The uh uh the the basically idea is that uh qh I want to.

A

You, you reuse, the service mesh.

A

A

So so, basically.

A

Kubernetes cluster in the cloud, and also you have some notes on the edge you want to manage, and also you have some of the devices you want to manage from the uh the the end.

A

But actually this whole texture is very flexible if, for example, in the manufacturing cases, the the factory may not want to have the data in the public cloud, they want to hold on-prem deployment, then they can actually just install the uh kubernetes. They basically currently are the depart market uh marked with cloud in their uh central uh central central room in the factory and also they can have the uh edge nodes in in different uh workrooms, and also the uh the networking between the the cloud core and edge core they can.

A

They don't need to uh have the uh the wired network. They can just rely on uh wireless like wi-fi or 4g, or we know that 5g is definitely better, but if even it's 4g is ok, ok and also they can have the uh the devices on their product so line.

A

uh The whole architecture is quite flexible, really depends on how the end user want to uh deploy that so the um cloud core.

A

So the cloud core itself is actually just doing the uh shadow management of the up uh the stuff, the resources on the age.

A

We know that for a standard, vanilla, kubernetes node, each cubelet will manage and will update the uh the known resource as well as part, that's running on that node to the master right and on the uh for each computing device uh in cases. The difference that is, that the actually the edge nodes are located on the edge so cloud 4 cloud core here will hook all the updates and then distribute it to the remote nodes.

A

Inside cloud core, uh currently, we have uh the five part, uh basically five modules. We we actually uh uh want to reduce the number of processes, make to make it easier to deploy and manage so edge controller device, controller and sync controller are in one process. They can be deployed as a part and also the cloud hub can be deployed as supporting a kubernetes.

A

So age controller is a shutter management for actually the kubernetes core apis, especially the node parts, and also configure map or services right and and also the uh the secrets and ending points, so the device controller uh so actually for the uh industry and the iot cases, cubbage provided a set of the device api device crd to simplify integration with a third part device protocol.

A

So the device controller. Here it is responsible to deal with the uh device crts, the lifecycle of the device crts and also it is responsible for the shadow management of the uh real iot devices on the edge.

A

For the scene controller, it's actually a a helping controller. It's it's responsible to deal with the uh consistency. It will uh periodically check the status in kubernetes and those successfully persisted on the age. If there's any uh inconsistency detected, it will trigger the.

A

Reconcilement and also this is a very useful and useful in the uh disaster recovery, so this system doesn't need to uh check the whole things they can. They just need to check the current status between cloud and age and cuba cs. Driver here is actually a plugin to hook the storage providing.

A

A

A

A

A

Sorry uh there were some network issue on my side, I'm back now.

A

Let me share my slide.

A

Okay, also, the uh admission web hook here um is uh dealing with the uh the crd validation and also in the long term. We will integrate the best properties.

A

For example, we know that kubernetes by default it will mark all the uh node unreachable.

A

uh If the node is disconnected, it will mark it to unreachable status and evict all the parts, but in the each computing cases we know that the difference there is that even some of the nodes are offline.

A

It's kind of a expected status, so users should be able to customize the uh eviction uh period of time or how long they want to uh the they want their applications stay. On that note, that's one of the uh that's one of the uh scenario.

A

And also uh like other scenarios like that, uh so even for example, the uh a set of node is offline, but uh the uh part may uh the one of the agenda may get crushed when it's offline, the uh the nd part needs to be moved to a healthy node in that each site.

A

Then, when it come back the there are a data con inconsistency between clan and age, and in that case we cannot simply uh synchronize all the uh status from cloud to age, because uh it will uh cause the uh uh the position again. So uh in the longer term we will enforce the best practice in the uh uh the relevant age computing cases uh inside the admission admission web group.

A

So uh here is a a little more detail about how cubase make it happen to able to manage a a remote net node on the uh on the age.

A

uh So uh so this is when uh I think you all guys are very uh familiar with that. It's actually a a diagram, how vanilla, kubernetes spin up a application.

A

So uh so, like the workload controllers, we know that it's dealing with the workload api uh lifecycle management, so it basically fetched the uh workload like deployment updates and also update the number of corresponding parts and also for scheduler. It's actually just uh keep watching the unscheduled parts and schedule them to some of the notes, and also for cumulate each couple letter just to watch the uh pause schedule down to its cover, uh corresponding node and manage the uh uh the the uh pod light cycle after schedule.

A

Right, uh if you look at all the labels marked with zero, it basically tells us the uh when the system uh spin up right, the controllers and the schedulers and cubelets will create watch to the api server right with the uh filter uh with their own filters right also, uh then, uh for example, we say that a a end use a user uh created one deployment with a kubocado command that we're going to uh create with the uh deployment yaml and also the uh underlying uh cd interaction at cd interaction.

A

uh It's basically just a persistent thing: the uh the metadata updates right so uh once the uh deployment yaml is submitted, it will be uh stored in the hcd and then ips server can note notification from netcd the deployment is created and then the deployment controller getting notification. That deployment is just successfully created and then the deployment controller.

A

uh We know that it's actually uh creating the replica sets and then their replica sets uh create uh the number of parts right and then the parts get persisted and then uh the pod api server forward the notification that uh power port successfully successfully created to the scheduler and scheduler will uh schedule part according multiple uh scheduling policies and finally update the uh part spec right to to fill up with a node name.

A

We call that a binding right bind the part to a note and then uh the uh the specific uh node the cubelet will get the update. That part is about okay, so it it will spin up that part. On that note, create containers, create, create uh provision, the uh container network and also the container storage.

A

Right, and uh so what cuba h uh hooked in uh this lifecycle for parts uh is actually uh is actually the step outline here.

A

So uh how could it work um in this case? Actually uh after the schedule uh scheduled a part? It's not directly the cubelet uh guarding the notification notification, but the cloud core.

A

So actually the age controller got the notification, so actu, actually age controller will get all the uh port updated uh notification, but it will filter uh it deal with only the part that that are scheduled to the age nodes right, because the age controller has know everything, so it can easily to filter the agent notes by name and get the corresponding part, and then it will uh build up with a message and allow the cloudhub to send the notification to the corresponding node.

A

So here the reason why we have a cloud hub and hub here is that we know that the age nodes made locating a private network, it doesn't have a public ib address or it's maybe behind a some of the uh firewall right. So actually the the when the system starting up the age core is actually create starting the uh communications.

A

It will request to the cloud core to build a a biodirectional communication terminal, and then it will, uh the the the eternal will keep there and once the pod updates come, the cloudhub knows that and it will send the message to the corresponding uh h all right. So when the h core of the clone hub got the notification first of all, it will let the meta manager know so. Metamanager here is basically a node level, metadata persistency.

A

So for the metadata storage backend here on the edge node, we are using actually using a sqlite currently because it's it's kind of only actually. Caching, the nodes uh informations on the node, uh so circularity is a very uh lightweight and simple uh storage for a database for uh storage in uh like the metadatas.

A

If any uh inconsistency happen or if it's broken, we can still uh synchronize from the central cloud from the kubernetes api server from the lcd, because there they have the whole set of the uh the the data there right. So at the end, uh when the persistency uh succeeded, the method manager will forward the uh the event to the hd. uh Hdd is actually a light weighted uh coolette and the hdd will spin up the uh part on the note okay.

A

So so, actually, if you uh ignore the uh the cloud core steps in the middle, it's actually uh quite similar, comparing to the vanilla cumulative behavior in the cloud.

A

So basically we we have added mu, a few more steps to forward it to the notes on the age.

A

Okay, so on the age side, we actually uh also keep the uh idea that we we want to. We want to run the whole stuff inside the process, so make make it easier to manage.

A

So currently we have uh the uh uh the following uh modules and for different user cases, users are able to disable the uh model if the uh it's not useful. In that case, for example, if so much user uh like running cubase in a cdn cases, they may not need um the device management functionality, so they don't need that all right. So uh for the hub, it's a a symmetric um implementation.

A

Comparing to the cloud hub. It's uh it's dealing with the biodirectional communication.

A

Okay, so meter manage here, is actually a node level metadata persistence right. The hd is the lightweight cubelet uh implementation. So basically we we, uh we are actually uh reconstructing of vanilla, tubular modules to uh remove the uh like uh cloud provider, uh storage, implementation, yeah and device training here is actually to deal with the device of status updates between cloud and also persisted on the edge node for the applications to uh to process the event. Bus is a client implementation to uh talk with the uh mqtt broker and also the uh age mesh.

A

Here is a like, I said it's a a multi-layer functionality to deal with both service and as well as the cri. So the difference between each mesh and the cni here is that c9, so the uh qh in the qh project. We don't want to rebuild dci, even even though the current cni implementation can not deal with cross private network communication.

A

We want to make it there is that the existing third party ni they can still deal with the cross noted communication inside one network.

A

When it's cross-subnet discussion uh communication each mesh will help to forward like use the, for example, p2p terno, or we rely on some gateway nodes to forward the packages.

A

For the monitoring, currently, we support the uh the standard, the metrics interface.

A

All right so that that's basically uh how cubase works and we I will quickly go through the uh a few updates about uh uh latest uh cubage uh release, so we just have our 1.4 release by the end of last week. So since, since last week time we met we, we up, we improved the component config.

A

So now you can easily use a declarative, uh basically the kubernetes style of api to config, uh qbh components and also we provide the minimal config and also the full uh default config command to help generate uh config file with defaults for the node setup. We uh provide the age node or to automated uh registration and also automatically uh enforce the tls between agents and the cloud hub and cloud core, and also we introduced the automatically uh certificate rotation. So users don't need to uh worry about the certificate exploration in the edge computing cases.

A

uh We also have a uh installer tool, it's called the ke enemy. So again we are reusing the idea uh we have in the vanilla kubernetes in the data center. uh The enhancement recently is we added the uh centos support and especially, we added the rest. Raspbian support so so now user can easily uh use one command, k, admin and a join to set up a node in the uh raspberry pi.

A

Also for the uh club part, the higher will be availability improvement is that we we implemented active standby mode for uh also integrated with randy's gates.

A

Another thing I want to highlight here is the cube cubed itself, the cloud core. Even it did it just have one active instance.

A

At the same time, it can manage one up to one thousand, eight notes, so uh we think it it's a it's a it's a it's a good status for now and where the community are keep working um to improve their scalability, okay and then another thing is that we we improved the uh reliable message: delivery between um between cloud and age to to make the uh the whole system uh more stable and for uh container runtime.

A

We, uh the community, verified the integration of cryo and the cutter containers, and also both the x86 and arm architecture, especially arm v7 and v8, are verified and for uh the uh observability uh users are now able to uh use two cutoff logs to collect all the logs uh from the pods on the engine. To the 2d cloud, basically to to the where you run the uh the command and also for the matrix, we, we supported kubernetes metric server for collecting both those in the cloud and uh those in the age, so basically uh cubase.

A

The idea is that we want to extend vanilla kubernetes to make one cluster able to manage both uh node cloud cloud, nodes and agents.

A

All right, another uh big improvement in 1.4 is the uh device management, so uh we actually uh modified a bit about the api, so added more uh uh improve the the uh capability of customizing, the uh uh the industrial protocol and also added the functionality for users to process some data on the age. For example, we know that in many cases the the the iot device may produce a lot of massive data and we need to use some filter to to reduce the data and fetch the one we want.

A

So uh so, with the data section and the data pipeline integration, it's easy to to do that, and also for the uh the basic uh basic uh properties. We move the property visitors uh from those model to to actually the device instance api. So uh because we we think that we found that in many cases the device properties may be different in different uh on different device instances all right for the community uh release development.

A

So currently, uh cuba is uh uh doing uh release every uh three months, so basically it's uh we're following the uh semantic versioning, spec and uh community. uh So actually we we're also doing quite kubernetes way. We do.

A

We will uh review the feature planning in the beginning of each release cycle and also we will do the uh code phrase uh when it has two ways: it's two ways, two weeks away to the final release of each uh recycle and the accept only bug, fixes and document updates on the until the release out to make the release stable- and another thing I want to highlight is that, because we know that the uh the underlying hardware architecture reality so currently, both x86 and r, are supported as a native architecture through whole release lifecycle for the code.

A

The code is actually uh very general right from the build to test ship and the run. We have the arm-based ci and also we have the cross, build cross, testing scripts and test cases.

A

So a little bit about the community updates. Currently we have around 400 contributors all over the world and more than 130 contributors submitted the code, and we also have the uh over 30 organizations working inside the community.

A

It's covering the cloud provider, I.t service provider, hardware, hardware, vendors and also the iot vendors and the economic, economic academics, and also uh cuba community, currently is moving to the sig based governance model. So recently we started two new things: one is device ltc uh that say is more uh focusing on the uh device management uh cases on the age and also uh responsible to uh support the and improve the extensibility of integrating more uh industrial protocols.

A

Another is the sig mec. It's it's focusing on uh improved cubase to become a uh relevant reference. Architecture of uh of uh building a infrastructure for 5g and mec, and currently uh the the sig is collaborating with uh since ftlg telecom user group and also the rfh projects uh we have. We have a uh two relevant blueprints inside. Our acronym acronym is the project under rfh all right. uh So this here we have a few uh uh contributors here, so we can found that uh the uh community member companies include the cloud providers.

A

I t service providers and we also have the telecalls and also the uh hardware and the iot vendors coverage itself currently have more than 20 adopters and the uh so. The biggest use case is the uh china highway uh provincial uh electronic tolling system. It's currently managing uh more than 50 000 nodes, so very impressive, all right uh yeah. This is more uh a little more details about the uh use case, so, basically uh uh the so actually in china.

A

The highway uh system is very uh complicated and also the the totaling system uh they have is highly distributed and have in total. There are. There are more than 100 000 nodes to manage so currently they use the cubage and in the cloud, the kubernetes to manage the whole infrastructure.

A

So they can do uh training in the cloud to uh to to get the model to recognize the uh car number plate uh and also uh ship the uh models to the influenza inferencing model to the age, and the aging notes will recognize the uh the car plate and also uh uh trigger the electronic tolling all right. So uh one of the uh progress currently is that the system is uh dealing with over uh 300 million data records per day and for a the car passing through the tow station.

A

It improved from 15 seconds to two seconds and the fourth truck it reduced from 29, almost 30 seconds to 3 seconds so very impressive, and currently the uh the end user are still uh doing more work to adopt a deeper uh qh to introduce the vehicle corporate uh system and also uh provide the the central uh video cloud and improve the uh ai algorithm to make the uh whole system smarter all right. So a future plan.

A

We still want to improve the uh the uh observability and the maintainability so and simplify the like remote debug from the cloud to the edge and also improve the uh cross subnet communication. So currently the implementation is relying on the gateway nodes in the future. We want to implement with like p2p tunneling technology and also like other technologies like sd1.

A

And also for applications, we will we are going to provide decentralized, the security, basically the identified identification for authentication for the applications on the age from community-wide.

A

uh We think uh we are currently uh applying for cncf incubation, it's still under voting currently, and we want to uh uh improve a better uh contributor experience and also the uh uh have more uh contributor events, and we uh also organize more uh uh the uh events and also uh we may do more cross-community collaboration.

A

So these are the use for useful uh resources about the community. All right, uh then, I will share with the uh demo uh because of the time I uh record a recorded uh recorded with the video.

A

A

Hold on, I said.

A

All right, so the first demo is the uh device management. uh That's actually uh with this device developed against the device. Crd v1 alpha 1, but it's uh I think it's it's. uh It's uh simple enough to easy enough for the uh the audio attendees to get the idea that how qh api the device crd works and can manage the devices on the node, uh because this one in this case we are actually uh connecting the uh device with the wires.

A

So we we actually, uh the mapper here is uh converting the mqtt to the to the uh uh to the what to the I o of the uh raspberry pi. So it's it's a little bit different to the real world that we're actually converting to from a mqtg to a industrial uh protocol, all right so in this. uh In in this deployment, we use a couple to get notes and then we can found that one.

A

We have one nodes on the age: it's a raspberry pi and that one that time was based on the uh uh qh17.

A

Yeah and then we have the uh traffic light apart. This one is actually a device mapper implementation to deal with the uh traffic light management converts from the uh crd uh change it to the uh to the light at the gpio uh gpio output and here by default, it's off right. So let me just go few seconds back.

A

Sorry uh here, okay uh yeah, so uh so we introduced that the device train uh section in the device crd status, so the training. Basically they have two parts, one is divide uh desired. Another one is to report it.

A

So uh simply you can change the desired from off to on and it will turn on the uh corresponding light light to on.

A

uh In this case, we we actually uh uh coded the uh dpio details in the mapper. So from the crd side you it's very simple: okay,.

A

All right save it, and then you can find the red light is on and also we can change to change the green light on and just save. Then the green light design, and also we can change like the yellow light.

A

Or we can just turn off the uh red light and also turn off the green light here. uh This one is very simple: actually uh because there's more uh iot relevant it's and the actually, the the the load is how you design your mapper and how you uh deal with the uh the device protocols in details. So I'm not going to throw too much details here. uh If you have any questions or you want know more details, you can also enter the examples repository on under qh github.

A

You can find the audi code there.

A

All right another one demo here is the the uh facial recognition all right. So here uh I'm also using the uh raspberry pi as a h note, and for the camera I'm using the uh the latest. It's a 8, 8 million uh camera 8 million px. I think yeah and yeah. Let me go through a little bit quicker. Okay, so all right uh here uh on the right is actually a a uh uh cuddle uh terminal and on the left, it's I'm.

A

Turning on the uh the remote view of the raspberry pi node, because we we want to view the video there okay, so I trend the uh model in the cloud and uh it's based on the uh cpu, so raspberry pi we can. We have the cpu there. We can rely on cpu to do the uh facial emotion, recognition there.

A

So here we can found the uh uh the way I trend the application and also upload it to and also uh sorry all right, yeah still, this price kind of covering all right. So uh we because we also want to uh preview the uh the emotion recognition. So I set up the environments to display the video window and also I mounted the the uh corresponding hardware and on the reservoir pi, so made the container able to fetch the video frames.

A

All right, then, uh let's apply uh yama.

A

Okay, uh you can find out the video uh window is out, so you can find it's. uh It's actually uh recognizing the uh uh the emotion uh uh rapidly.

A

Okay, so, uh and also a little bit uh details inside the uh image. Actually, uh we have the uh the trend models inside there and the online ai framework. We we used uh paddle paddle here in this time. Pedophile is a open source uh ei framework, it's uh it's kind of uh equivalent to the tensorflow.

A

Actually, uh you can find out the uh so uh this image uh it's actually the image size is around 500 megabytes and most of the space is used by the base image and for the model.

A

uh Here, it's a it's actually a very small model. It's only about um 80, uh 80 megabytes, yeah, all right, that's more about the demo, because it's online uh tutorial, so um we don't have a very good interaction experience about trying out by yourself, but uh please feel free to ask any questions uh here or uh even joining the uh community.

A

To ask your questions. We are the most of the community vendors uh members. We are available on the slack. Let me share the community.

A

Resources all right, and while this resource sharing will check, if there's any.

A

Questions all right also, maybe I need to stop sharing and then answer the question.

A

A

Yeah, uh let me repeat the question: first, uh the from alexander hi is have a website. Client server in regards to the cloud hub, in other words, is who initiates the website connection all right. So uh so, actually, uh cloudhub is the uh website server, because in the cloud we can easily get a a load balancer or a floating ip to expose as a service and for each edge node. When it spinning up, it will uh send a request to the cloud hub to initiate the websocket connection.

A

So it means that uh hub is the website, client and the uh the cloud hub is the server.

A

All right another question from um from alex thanks for the tutorial I was wondering, may you tell us more about the crd abstraction provided for device uh in the cubic thanks evidence? Thank you for your. Thank you for your question. This is a very good question, so actually uh we we changed it a little bit on the device crd in latest 1.4 release and latest update. You can actually find out the uh either on github or from a cubase community website. The cumulative community website is cuba, dot, io and cuba.

A

Github is actually uh github.com cube age and you will find the uh detail uh crd obstruction. There.

A

All right, uh we are running out of time, so I will end my uh talk now. I can see a few more questions there. uh Thank you for your uh joining. uh Please answer. Ask your question in the slack channel and I will be uh answering there and also there we have a few more other uh community members there all right. Thank you for your listening.