Cloud Native Computing Foundation KubeCon + CloudNativeCon China 2018 (Shanghai), 26 Nov 2018

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Evolving Sigma: Adapting Kubernetes at Alibaba - Xiang Li, Alibaba

Description

Join us for Kubernetes Forums Seoul, Sydney, Bengaluru and Delhi - learn more at kubecon.io

Don't miss KubeCon + CloudNativeCon 2020 events in Amsterdam March 30 - April 2, Shanghai July 28-30 and Boston November 17-20! Learn more at kubecon.io. The conference features presentations from developers and end users of Kubernetes, Prometheus, Envoy, and all of the other CNCF-hosted projects

Evolving Sigma: Adapting Kubernetes at Alibaba - Xiang Li, Alibaba

Evolving Sigma: Adapting Kubernetes at Alibaba - Xiang Li, Alibaba Sigma is a cluster management system we use at Alibaba for running online services. The typical container cluster It manages has tens of thousands of servers which is beyond the capability of current Kubernetes. It also manages legacy workloads which are not container based. With the increasing popularity of Kuberenetes, we tried to integrate Kubernetes into our Sigma system. However, there are significant challenges in doing so. How do we extend Kubernetes to support different types of workloads? How do we scale Kubernetes to manage thousands of nodes? How do we keep Kubernetes API while integrating Kubernetes with our infrastructure layer? This talk we will first provide an overview of Sigma. Then we will cover the solutions on how we extend, scale and integrate Kubernetes into Sigma at Alibaba. We will also share the lessons learned in the process.

To learn more click here: https://sched.co/FuJe

A

B

Good morning,.

B

Segment down to Cuba nice and let me introduce myself I'm from Ally engineer previously. I was poking around human eyes and from 2013 I was etcd project and then later in Cabrera's I was leading the team and this year I joined Alibaba and to do the cluster management, the segment ecosystem and then use it as the engine to evolve into the new community.

C

B

My topic today.

C

B

Cover the three aspect, the first exit back wants of Sigma and.

B

B

Armored environment, to talk about the challenges in the.

A

B

Architecture well, actually,.

B

B

Requested to use such kind of resources, or do this true, and we also do.

A

B

And we also have some kind of.

B

And we also have.

B

Some consistency actually a.

A

C

Amount of machines inside.

B

Alibaba and we're serving over 60 use with high quality and storage resources, and we also support 11, November 11th, and this year we actually work on the stack.

A

B

And we found out, we achieved the Tempe GME and we concluded.

A

32,500 transactions per.

B

Second and.

C

Other scale we.

B

Always their butts in hundred a thousand nodes.

B

As I told animation, Alibaba is operating on the level of containers. Well, previously, actually Alibaba was eroded from the virtual machine.

B

All this kind of application was matrix application that time we want to sway the interest applications to be to deploy them so the first stage we use the dilution package each application so that we can have more efficient scheduling.

B

Issues, first of when they provide the basic functions we saw some.

B

We used to make him Speedy's about.

B

32 women released about a ton Zhou.

B

That time we offered a dedicated IP, an isolated.

C

B

Resources, so that's my appreciation was quite smooth or in 2015.

B

B

Kind of images- and we want to describe this images with the standard of formulas. So that time we developed an ally talker and we made.

A

B

So the features of T fours incorporate into the new container, which forms a foundation of the porch that we launched.

A

B

Constant management, 2013 system that are called a.

B

Distinguished features or annotations of each applications only do resources allocation other words, preparation for T for well in 2015. You will want to have a better resources, management and allocation of this kind of resources, and that was quite complicated.

B

Attempts in each business unit- and we made Alice, Wong and Joe system- is a common system in 2017, we'll fund. We made attempts on these three systems and we want to combine this advantages and who made a new system to be the major strands of remember to support our business. So in 2017 we launched the Sigma 2 Pallavi row and including most of all, my voices.

A

B

Launched and we started and work on segments of 3.2, which will work units as a core component.

A

Introduced later.

B

Actually, beyond the concept was a range of chewiness. Actually, it's a more comprehensive cluster management to work for his work. It can be simplified into the following levels and the first it will do the network, storage and computer to that abstract of this, so that we can have the visualization on that aspects and, above that we have the online scheduling, that is to say once you submit to them resources across.

C

B

Word do kind of struggling and we also have offline schedule that to response to scenarios.

B

And what we deal with offline schedule- and we also use various methodologies to do such kind of arrangement.

A

B

Want to dynamically management and regulators kind of resources, as I mentioned, there's a bad job and online service notes the same sensitive to to the resources. So if you want on the same resources, they will have some competition and make awesome. Lanten saying is a way to test before. If there's resources me I'll compete with each other we'll just we want to avoid that, so that issue.

A

Efficiency and stability.

B

We also have such kind of auto scaling.

A

B

We want to skill automatically according to customers need, and in this way we want avoid the overhead of resources to avoid the static resources or resources of wastes. For example, in the peak time of each days, the system, your automatic, will lift up the volume to response to the rise of the volume without human interference or operation, and we also have the resources, accumulation, analysis and above the supportive tools.

B

We'll also have some highly level management, a tool like chargeback well and aims to monitor each business units or each team of costs of each purchase so that we.

C

B

This kind of services, it is a charge function. Besides, we also have a capacity planning.

A

B

Since the job of Sigma and resources, a different and quite complementary, and if they come collocation on the simple, the resources utilization went to equip you increased, it will be increased and you can imagine that.

B

They will deploy on the zero level and we can ensure that resources are.

C

Manageable on each.

B

Level in this slide, we can see that.

B

Utilization rate is not that high.

B

It's quite high.

B

And offline services, we can see the overall utilization which.

B

Is quite accessible and we can see that the relation.

C

B

Sigma and a food-safe and quickly in.

B

Creation is increased by 30 percent of that, for the deployment of machines are contingent. It's only increased.

B

With the control plan and the data architecture.

A

C

On the plane, we have our note agent.

A

C

If you're familiar with a Cuban I take society as API server, everything is communicated to the API service of a previously. We use etcd as they communicated, so everything will be with lattice through the etcd and also we have a scheduler for scheduling decision. We have an API server as well, it's different from the KPS. It can directly control the party. It's like a Daugherty, you can imagine it as some functions.

C

We put certain functions of cabinets into the API server, so it is also a North executor, so the API server undertakes more functions and, as you can see, there is no controller.

A

Treatment was in the table terminology in the.

A

Document images that I- don't you learn to do that Ignace or resource manager controller, will not ask you. Let's see what you saw Julianna tackle with up here, take a seat nonfiction beyond your steel Fe. Unless you spill yeah, that's a good, oh man, that the UFO meant you're, worried they'll, be a similar. Tell me like that's your path, top of the soil in Bolivia, so so you'll remember all including.

A

Exporters in the kids or I think I just wanted everyone know about have to go remain a little basically Normandy style memory. Games of laxatives car needs to go beyond ing little bit through another nickel one tobacco Soviet domination balance do when we see on our models.

A

Any HTML polls that designers, don't wanna open, become yellow just gonna. Be me.

A

Okay, seventy of I see they're already quite quickly that these are no one may not have some political. Each reaction I think realizes when talking about not opportunity. Yet we've got you in here. We actually have ultimate her peignoir. Let's he added this year. She doesn't it gentlemen, I'm gonna go back in recovery, I've, actually a see whenever people get masculine I would think of transmitting alleged intelligence activities on working privacy. Sometime, no sweetie, you time until I call you official.

C

Game they were to see them I'd like to apply this in.

A

The reversibility holiness to the use extracts times we see it or anything as Class C.

B

Minor persons with general be ready.

A

C

A

Your woman's gentle look before any is that next ability to evaluate my feeling for tonight about whether immigrants they orange yeah, we need immediately carry out their duties. The lion video I could be a how to do a single email make sure I knew these gentlemen governor made it. You only be connected to a favor all right into the kind of recovery we're all into smaller pieces or anything happen. Company line.

C

That maybe little game you lose anything. It's unable.

A

To attend creative status to communicate and we're gonna make you recover from memoization generating a fatal. So here, because this you know, that's the denominator company, especially designer, could be jumping, though they're waiting, sensible order to deliver and happen. Will a text or a bunch of you have also lives with you see how comfortable it is, live in India for a value. That is a society that has provided quality to your video. If I stayed home by vector.

C

By rockler collects the students in severe.

A

Customize that I recognized you will see what that mean. I don't know, but you know sing along to an ecosystem Jenni's back in the corner, to terminate this data shows you see this application shows realization how to evade us right into it. What about ecosystems me? An accurate recording, Antarian.

A

C

A

Be a factor overboard, centenary, 300lb on the scenario system, business, architectural project, occasionally, camera consumer survey of American, ingenuity of the rat race of accent is equals to cut off. My own hair begins in Singapore America.

A

Piracy officer, maybe a broken heart I'm, a Superman, even economic, some kind of authority.

A

Don't fight me I won't be several questions when collegiately passes their apartment. I want to do this table choreo.

A

Come take you to see if it's do actually see do to receive in the table to welcome it. We've all made for TV appearance was filled to the lot or being commented about being look at the school, so you're, not single something so you'll be boxing. We're done yet.

A

It's inconvenient place with the Aqua is helpful, will make this happen. Simon.

A

See in living out ways to manage integral calculus unity shop, there is no learning I think you should have skills on that boss to evaluate X to a power.

A

Charger so they.

A

Say to a console: I don't want any.

A

Since that means.

A

A

C

So for this match, it just need to do distrito wait for the regulars of the first match before the match started, so for this image, preheat actually express it to us. They are th, not know what tricky image is download. The image to the local in advance.

C

Stability is very important to us. The cluster itself also needs to has certain functions, and we need to optimize them as well. As the data is very important to us. We actually established such a data pipeline for data connection.

B

B

Concentrate rapier and we're easier to risks, making us be this other thing to know, and we quarks and if they make a house it will not affect the whole system, and we want to address that issue and for the second challenge is about. Some large question now is: are given easily support. Oh.

B

B

Ecosystem and to to make them work better and I want to cooperate with the upstream platform so that the we can benefit all the worker in and reduce the development costs and increase development efficiency. And that's.

B

A limited time do you have any questions.

B

We all the thing to occasional.

B

B

And if you want to change the zr d, our stream- and we want to make the other resources request on the other part of the of the system, I mean that will be as a trend of a given path in the future. But.

C

A

Still want to have taken one container.

C

A

Know that sound.

B

Discussion and they'll talk about events. Well, we have already been mentioned that.

B

Many crafters are only one question.

C

B

Mystic mechanism response to the issue, I believe this is a.

B

Check the house always commercialize the resources.

B

As you mentioned, they're portable correlation among service I believe this seshu were to be adjusted by the middle well.

B

Central package actually will want to think and work out. This question will cover this issue together with a limited time.

B