Hello everyone, my name is Wei Zhang. I'm from Alibaba Cloud Kubernetes services team. Today I'm glad to share serverless Kubernetes, to include service containers, introduction, architecture, and several common use cases. Okay, and we'll start. The first part is introduction of the service container. Now, does everyone talk about Serverless three in public cloud. We all love the function computing, like n dimensions and lambda. We can see that developers love serverless the most, I think. The reason of that is, I think it is a service that make a developer's life might be much easier. We think the reason contains the three specs. The first is agile development. In service way, developers don't need to purchase a service. No need to provision service. No need to manage your back-end service. All you need to do is just a real long workload and they don't need to upgrade operating system and system software. We can see that the productivity has become much higher, and the time to market of the application is much shorter. The second reason is scaling. We think in serverless way, the scaling is hyper-scalable. Application, better with serverless infrastructure, where it scales automatically as the user grows older and the usage increases. From a developer point of view, the capacity can be unlimited, so scaling in serverless architecture is much simpler. The third reason is cost saving. In serverless architecture, there is no need to maintain fixed resource pool. It's pay-by-demand. It charges at startup, when workload is started, charges at end, when workload has ended. We will never pay for the idle resources. We found cos saving by. When we change the application architecture to service 70% when we change the application architecture to serverless. In this section I will talk through the service container. Service container is one kind of serverless architecture. It have all advantages of service architecture, agile development, and easy to sale, and cost saving. Beside of these advantages, it's based on container, not function or any other form. Container make our application good once and it'll run anywhere, there is known limitation on language or the library, we can choose whatever language or library we want to build our application. And the most important part is ecosystem. The whole cloud that you have in the Kubernetes ecosystem is built on container. We can benefit from anomalous language has, and still connect with other, and many other Kubernetes framework. We don't need to worry about vendor lock-in because all this is built on top of containers. Today almost all public cloud providers have their own service container service available. This is part of the landscape of the public cloud. In service container catalogue, Alibaba Cloud has serverless Kubernetes and AWS container instance. AWS has Fargate and on based on Fargate, they have ASK Fargate and ECS on Fargate. Azure has ACI, Azure container instance. Lastly we will deep dive into our service called container product, this here, ACK on ECI and ASK. ECI is the means elastic container instance. It just simply means that we can create a container without managing all this infrastructure. Before ECI, we have to purchase ECS virtual machine to deploy a container. With ECI, no ECS purchasing all, and the provision as needed. A container is a first-class citizen in public cloud. Each ECI is completely isolated, it runs in a sandbox execution environment. The CPU is run run from 0.25 to 64, and the spot GPU instance. All this is paid by On the [INAUDIBLE]. Using spot instance, we can save on cost, mostly by 90%. ECI startup time is about 10 seconds with the image cache, image pulling timer also can be served. And also, to remind [INAUDIBLE] ECI and ECS share the same IaaS pool, this can guarantee that enough online resources are provided to ECI users. Standalone ECI is just like a single powered incorporated cluster. If we want to use the deployment with multiple replicas or step by step for certain jobs or [INAUDIBLE] jobs, we shall use a code that is on ECI. At the Alibaba Cloud, ASK on ECI, we have flexible ways to collect [INAUDIBLE] ECI. The first one is using the ECI pod in [INAUDIBLE] cluster. The second one is the [INAUDIBLE]. We call these ASK. In ASK cluster, all [INAUDIBLE] is an ECI instance, there is no load needed to manage it. It is completed. That is architecture. You basically on this architecture run [INAUDIBLE] [INAUDIBLE] under ECI Cloud is succeeded in one cluster. Normally user can put a long run workload on ECS load and put a short run or scaled workload on ECI. This way we can optimize the total cost efficiency. [INAUDIBLE] Cluster capacity is determined by the number of nodes in cluster. By adding ECI into cluster, the capacity of cluster becomes unlimited, which makes, it can be much more easier, because users don't need to worry about the capacity planning. Using kind of a hybrid cluster, ECI Pod and the ECS Pod can be interconnected. It means ECI, it's easy to communicate with ECS Pod. [INAUDIBLE] Also other cluster IP service. In this way is ECI Pod is just like a [INAUDIBLE] pod. In service [INAUDIBLE] cluster, we have load and we need to manage it. Users can just run the container application without the [INAUDIBLE] infrastructure. It's quite easy to use. In ASK, we will never see the common user issue of workloads. For example, load not ready does not exist in ASK cluster. In other way, ASK make a comparison much easier for end user. The scanning ASK is very faster. [INAUDIBLE] ECI. It is compatible with the [INAUDIBLE] API like [INAUDIBLE] Java service, [INAUDIBLE] and [INAUDIBLE] and the other operators. We can also install [INAUDIBLE]. The pod can be mounted disc [INAUDIBLE] all sorts of forms. In ASK we provide connected service under industrial workload. Function, that is, we will expand this model. ASK also integrated well with the ARMS and SLS products. It's easy to monitor the [INAUDIBLE] and [INAUDIBLE] into SLS. This is the architecture of Service Corp. That is as we can see, the central part is ASK schedule which is responsible for what part of changes and then schedule the product by calling user API to quit ECI instance. Then ASK scheduler synchronizes the ECI instances that is [INAUDIBLE]. Using this architecture, we [INAUDIBLE] predator there compared to [INAUDIBLE] under ECI resource there. This makes it ideal for elastic scanning in public cloud. The next part will show some basic functionalities to use the ECI Pod in [INAUDIBLE]. It's quite easy to quit a GPU instance. GPU driver is preinstalled, such as tensor flow and code are torque it. We can just configure the GPU specs in a rotation and there's Number of GPU in the GPU instances created. Each GPU instance shares similar specs as you see as the GPU. Spot instance is also called a preemptible instance, which means its power can be preemptible. After one hour is crucial time based on a [INAUDIBLE] in situation. Spot instances can minimize the cost in many sub areas such as big data, image and media coding, scientific computing, a Scrabble website and the tester. Such as the same as you see as a spot instance, we both SpotAsPriceGo and SpotWithPriceLimit strategy. With the last workload, what can manage the replicas distribution of ECI on demand and spot instance in one deployment. This can save cost for long run workload. The parameter is therefore set at workload. In this example we prime the engine, so we'll have two replicas that run in normal, is on demand, and the rest for replicas runs as a spotter ECI. Recently, ASK have a manager connected with serving functionality and provide automatic scanning when the part that have low traffic, it will scale to the minimal instances with lowest cost. When traffic become normal part of their schedule, normal instance, more connected with serving functionality is also supported in ASK. Okay, the final part that we will see some common use case of Assembly [INAUDIBLE] that is. Weibo is the largest social media company in China. It's just my computer in United States. You could have unpredictable hot events which bring hundreds of times of traffic in a short period of time. In order to handle this kind of people traffic using ECS scanning normally at least needs five [INAUDIBLE], but using ASK and ECI we will reduce the scanning time by 70%. It can scan 1000 parts in one minute and schedule zero when there is no traffic. Some new [INAUDIBLE] Weibo, many online education companies use ASK or ECI to handle people traffic. When people traffic coming, just scale workload on ECI. So there is no need to maintain large amounts of ECS notices in complex cluster. Big data computing is also common use case for ASK. What can schedule spark workload jobs on ECI? There's no need to [INAUDIBLE] a cluster capacity by this story. Today, many users use a scale for CI CD pipeline, use scheduled check ins and level runner jobs in ASK cluster. The total coaster is much lower than using a classic [INAUDIBLE] cluster because the job is paid for on demand, not the load. Okay here. Here I have some links for getting started. Try ASK and ECI, it's very simple, just need two steps. Firstly is to create a [INAUDIBLE] complex cluster. Secondly is to create a part inside a cluster. Okay this is all my slides of this section. Welcome to [INAUDIBLE], so this contains the service code that is in Alibaba cloud. Hello everyone, here is a demo of [INAUDIBLE] that is, in this demo we will show you how to create a service cluster and show how to create a part in service cluster. As this is our ICK concern, we can create a normal basic cluster in this web concern. Also we can create a service cluster in here. Creating a service cluster is very simple. First of all we give our lemma. Now we type service. Which was averaging. Then lastly we choose the [INAUDIBLE] version and that's a VPC. We can create a new VPC or use existing VPC. Which was in the zone. NAT Getaway means when we able NAT Getaway, we can let our pardon accessing public Internet from the goodwill. So if our product, there's no need to access the Internet or just disable in this way. Public access means we want to access the web, okay, server for my Internet. So we need to expose, I guess, with the AIP. The PublicZone means a service discovery in cluster, which just as similar with in ACK single cluster. Before we just run a simple job, there's no need to use discovery, so we also, don't get PrivateZone here, okay? Then we Agree with Terms and I see Create a Cluster, so as we can see, create some of this cluster is very, very simple. There is no need to trace in or provision and notice that all works. They just given amateurs reaching and cretae a cluster. Okay, so it took about one minute or less time to create a Sims custom. And it was a 2 minus that suffix cluster can be totally free. We don't, there is no charger or free, if we don't choose yeah, NAT Gateway or Public Access, so it's a totally free. It only charges when we create a part in service cluster, okay? As again, we get back, which is a initializing status. We wonder what for this subcluster with has the user exceeds in 100 region ACK cluster. We can copy the configure into a sensor 2 terminal. As we can see, I mean this sort of cluster there is no node order. So it's a different with normal complex cluster. In normal complex cluster, there must have one order if all running a part, so but in certain cluster, we just show our part. We don't need to care about the underlying builders. Now, first demo is just that draw a simple M5, it means CM-5 contains a service and a deployment replicas of this deployment user 2. And it's quite a simple, it's just the engines imaging, applied, Okay, get pod, As you can see see the power is assigned to a virtual coder to builder. It becomes running, status, another impending. Okay, both of these becomes running and now, we check the date is again continued how one vertical data loader. So this back C shows us now, we first have taught the virtual load. Measurement data series, we could, okay, this is quite a simple demo that I would delete it. No, okay, we then can resched this deployment. Reschedule 2. OK, we also can check this part in web concern. In here. We got into the cluster. And check the workload. IPods. They are become running [SOUND]. They hand the true loaders through this cluster. OK, we will clear that resources. The next demo is the spot instance. The unified is still very simple. It is a deployment of visit replicas two and though we settled the specs is two under four gigabyte. We use this spotted strategy as the price go, G means we don't set as a price limits, so it's not a total air prize. As a marketing flows [SOUND]. OK, the two recognizes, but I'm running. Actually this two parter instance is. Spot instance is not undermined. So basically the. Coaster obviously spot instances very low. So each, so spotting cities we suggested it can be used in medicine area like a jobs, their city or sparkle pistol or big data computing in medical fees. The next demo is awaken. Install a ham charge. He just looked how nasty. Workload. Just quoted here, very simple. OK, I have one. We need to connect [SOUND]. OK. Could update [SOUND]. The industrial workload or controller isn't in pending status. Amazon, OK. Then we will create an industry work loader object in this cluster at infested share for the young file. This is young file the. It is by employee first is our basically primal definition. This is quite a demo and we find a new industrial walk alone object. Do reference and have reference to the engines deployment are the minimal replica services arrow and the maximum is it true and I told her last walk loader replicas in six and they have are in a secure. You see I spotted in this unit all pod have a rotation races spot as price growth. It means you this is nasty unit or product is Spartan and they have basically to represent replicas is normal easy pod. Yes, quoted [SOUND]. OK. Let's clean up before in juices blocks. As we can see, they have a true is that normal part, and for eci spot instance we can secure the elastic workload. [INAUDIBLE] too. [SOUND]. Scared with us through a cloud. [SOUND]. 2, [SOUND]. Okay, let's put for instance have at replicas. We can scale it down. To, fine. [SOUND] Okay. [SOUND]. They just have one, no more is iPod will scalar bank. [SOUND]. 2,6 [SOUND]. They have a four DCI spot. [SOUND] This is a demo or we lost to walk around? [SOUND]. Besides the obvious demo there, we can also easily too. John spark jobs in [SOUND] service cluster. A hammer chart installation is a very simple. Okay, sorry. We not doing, okay. We change it to this. It could happen. [SOUND]. After we install spark Hammer Charter, we can literally concludes sparkle jobs like [SOUND]. Like this one okay. [SOUND]. Turn to, concern that. This is a demo of service complex cluster. As you can see, it's quite a simple and welcome to try it. Thank you for the time to watch this demo. Thank you.