Cloud Native Computing Foundation KubeCon + CloudNativeCon China 2019 (Shanghai), 5 Jul 2019

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: TiKV Best Practices - James Zhang, PingCAP

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

TiKV Best Practices - James Zhang, PingCAP

TiKV is an open source distributed transactional key-value database, also a sandbox project of Cloud Native Computing Foundation (CNCF). Built in Rust and powered by Raft, TiKV provides high availability, strong consistency, ACID compliance, and horizontal scalability.TiKV supports externally-consistent distributed transactions and also implements a coprocessor framework to support distributed computing. In this talk, we will introduce some best practices of TiKV such as how to control data distribution and what is the recommended deployment in cross-DC scenarios. We will include data balancing topics like scale out and scale in, and how to control the speed of balancing. This talk will also show you how to identify a hotspot issue and how to fight with it. Last but not least, we will introduce how to fine-tune performance under different workloads.

https://sched.co/NrmC

A

So good afternoon and welcome until you become 2019- and this is about iqb best practices and introduced about on TV. So from the description. Actually, the speaker would be John simple, but I can't hear rather than Johnson Punk, because he is the CEO. And this morning we discussed about the topic that is tanki V, best practices and in order to avoid being sleepy to let our audience to be quite sleepy. So I come here to introduce the open source community status and the background of tanki.

A

We best practices, so I will quickly to go through the details and what is tanki being why we are doing this project since program as well as in.

B

C

A

And the second half would be given by Johnson poem about state I could be hot cold. Technical details will be introduced by mr. John simple and how to design it and.

B

A

Come from pink cap I myself is the co initiator founder of pink hammer, so you may ask for wouldn't think so anyone have heard about it. Okay, maybe most of you.

C

Have heard about team camp.

A

Well, as you can see from the list, this is CNC have company's contribution, the midst of the contribution, and you can see globally the global contributors in the project, the top 10 among the top 10 one is start. Adventist quarter, pink and iBM is falling behind. So this is our company, it is an open-source company enterprise and behind it and they tied to be the Tonchi be. Actually there are all the projects that we.

C

A

So talking about the topic today, actually 10kv, it is on key value database.

A

At that time. My idea is, if you have know about the up hunch, if you are quite familiar with it, you will know the HBase and HBase in C&C event is the database and it is the project of the database. However CN CF, actually it does not have a design for scalable the database. A large scale database to support our CEO, keep value, storage.

B

A

That is why, at that time we come up with this project, but if you think about another, HBase, well, actually its of no significance. So at that time, when we were doing this projects this program, actually there are some important decisions. The first is: we need to support the.

B

Transaction if.

A

You are quite familiar with HBase or concealer, and you will know been consistence and their transactions does not perform quite well, especially HBase. If we want to operate.

B

A

Will be full success or for failure, and this is quite hard to realize, however, int key be compared to with HBase. We have a big difference, variance in key value operation. We can support the a CID features support. So this is the first aspect.

A

The whole design is fine spanner and, as you know, they go go distributed. Database right and the second point to make is another difference from the HBase.

A

If you're quite familiar with HBase you, where no scientists data is stored in HDFS, but actually we just live it because we do not want the whole stack to rely on British storage. So here at the bottom, where use raft and rocks DB.

B

A

Realize it and to improve the overall performance and to reduce the latency. Actually, this project was initiated in April, 2016 and.

B

A

Now three years have passed well, actually, our company, the peak of its original objective, is to build a relational distributed database. That is, to recreate spinner.

A

C

A

The very beginning- maybe you do not know that and the very beginning, HBase supports that Taiki be, but it has various kinds of problems. So, since 2016, what happened? New, Cavalli's storage soul- and we do not rely on HBase and later we found we just got burned.

A

C

A

Work well in relational database for tiny B, but it also can work in other ways. For example, if you have the bottom layer, storage and if it can support key value and transaction so not based on this system, you can build different distributed systems. The first sample can use time KB to build another distributed.

A

That is metadata so that we can change the metadata without concerning about the inconsistency of the data or the availability. So this is one scenario. Another scenario is that we can, on top of the kV, we can have a Redis interface so that the visitors can access and it can be horizontally, extend so some of the communities they made be quite familiar with this process because they aren't doing it right now.

A

I need to enjoy high popularity, so our vision is to use tiny V as a building block called a distributed platform, and that is why we contributed to CN CF, and we hope that the community can, based on time, KB to be new things.

A

Architecture, a very brief, simple architecture of tie, can be a mountain mr. Johnson poem for further details, I'm not in that bright on it actually.

B

This is very cloud native project and its language.

A

Is rust and for the whole time KB it is grants community the one of the largest open-source project, including rust, its programming language and it's initiated its founder, actually is working in our enterprise on the 20 project and the GRDC. Actually, we use it from all their times.

C

A

Also, the maintainer fish library so online. If you want to use GM PC and then the same time, it is gr PC and most probably, you will use our products.

A

Another thing also where the maintainer and the storage weighs rods. Db.

B

A

Storage, the bronze. We also we're also the.

B

A

Of these two things, so our relation with CN CF after one year, sandboxing finally several weeks ago or become the.

B

A

Efficiency F, they can incubator of CNC everything. So this is the first incubator level database and this project is also led by channels team and it is initiated in China. So we've been quite proud of it and.

B

A

Is tine kV its official website? You can visit it so essentially it is.

C

A

Has many relations were together with time? Tv shows.

A

B

A

Relation between Thai, kV and tiny B, and it has many users.

A

In addition, besides the call of Tycho's be, if you want to participate in the contribution of this project, for example, some friends would ask me: I want to contribute code to try KB but I'm, not in database expert I'm, not an expert. So how can I participate in this project? Well, actually, I can also think about building a database. It's not only about the database code because it has some surroundings or contributions.

A

For example, now we have and present it's official terminal end is: if you operate time kV you can only use go library to access it, but in fact the entire KP community. You can also build a pure rust. Actually, this is quite weird why we don't have the rest, the user end, because you know the entire.

B

A

B

A

If you want to make the hard call, it will be.

B

A

Troublesome but now we have some improvement actually, but finally, which is rest so until now we have some available resources and community to help us to do rest and right so.

B

A

If anyone had any one of you is interested in rust, you can also participate in call contribution of time KP, but also you can join in through ransom.

B

A

Example: two secret plans to the whole community.

A

C

Entering to the same.

A

C

Friends provided the coast about 90% of land from our company about 90% was from.

C

Now I would give the floor to my friend.

C

Of the Chi TV project, so he would just say, give.

B

C

Related technology about the crew of this projection. Okay, thank you for introduction and introduced why we have the Chi TV for jumped soil who's. What the right, TV, projector.

B

B

C

From two aspects to introduce.

B

C

Project first framework and the main part is about the practice we will, that is to say, when we use them. I will easily project to what problems we may encounter and how to adjust. That.

C

B

Refined the transaction.

C

B

Whatever I want chemical.

B

Itself, actually is a.

C

New so-called it based based on.

A

B

This part is actually the circle.

A

B

By the circle to.

C

B

Your information to.

C

Chi TV, of course you can be yourself.

A

B

The system based on secret, for example, you can see about this.

C

Other protocols.

C

How to ensure that data.

B

C

A copy of the theta cannot restore korra's single.

A

C

That is to say, we.

B

Cut off the data into different and.

C

Each wrench is actually a rough to.

B

B

C

If a leader of the raft we have a leader in.

A

Originally, what looks like.

B

C

C

Single device, if you have some constraints, if you notice waiting of like unit status.

C

Was split into two cells.

C

We just merge the neighboring regions.

B

C

B

Data, so that the weekend rolls.

A

And a black thing.

B

C

Is the foundation of our ability.

C

A

C

We need to have.

C

The copy in the original device in this way.

B

You know Sammy, let me lucky for you just want to.

C

Actually me to just kill this device and put.

C

Another device.

C

Kaity is not only one project, it also complained climb, it also included clothes and other related. So.

B

C

Developed by herself.

B

C

Comes to the second apart: what about practice? The first step is about.

C

C

B

C

How to address that.

C

Kitey kitey they cluster performance. What can we do.

B

C

For example, this is a single.

C

C

C

A

C

A

C

B

Data is distributed.

C

Is well distributed.

C

Either control.

A

C

To control have precision control of these leaders to make sure that they will not appear other.

A

C

B

C

Migrate, the leaders in other works in the same rack.

A

C

Want to actually.

A

C

Luke I've, even known.

A

Pd address.

C

In this way, we can automatically I'd.

A

C

Node, that is quite simple right.

A

On compared to with the.

C

Traditional database.

A

To migrate Oh to scale it, actually, it has a big difference. Why is that? That is what we are doing and what we.

B

A

Working on and working on because we want.

C

To liberate or liberalize the.

A

Productivity, our workforce so.

B

That we cannot wait to not.

C

Need to waste a lot of time and.

B

How about hope they ought to know and then just turn it off.

A

Actually, it is okay, but actually ten summaries, for example, way and turn off the machine. The.

C

Whole standing deletes.

A

The data and then another machine another hose to the replica and all that.

B

Is broken down.

C

B

It is risking the wrong because we need.

A

To ensure that Ian Brown.

B

C

A

Of the knows, when they exist, is.

C

A

Data is secure, but if all the machines the holes break down now, you cannot ensure that the residents also maintains the data for.

B

Example previously 130 times three replicas.

A

B

Limited it has sony to us, but if.

A

It is not time to add the third replica on the other host and the other hosts break down. So.

B

Now you only have a my replica, but.

A

B

A

It so this time you just remove it and don't close the deleted stone now, because the replication works of this doors, regions.

B

Is still going on and you.

A

Should, after the divisional status,.

B

You can stop this time so.

A

B

Is first which, which so you need to delete yeah for.

A

Example, this tie Cavey yep and what you need to do, that is to store.

B

And so what did you do.

A

Actually, it is now doing a migration.

B

B

Store it only has two: replicas only needs one more action, that is to.

A

Complement the two replicas into three replicas, but how come we know when it is ready when interview.

C

A

C

So the day you can see the state of.

A

B

After implementing this order,.

A

It will become offline.

B

A

You can see from our site so at this moment, don't close the deleted store now, because the replications.

B

A

Region is still going on so after the deleted stole status, trance.

B

Trance ooh Tom, stone after it becomes.

A

Tom stone when it is labeled as gemstone. That means all the replication works have been finished. So you can stop this time. Cavey.

A

Next is about hot spot in.

B

The distributed.

A

System hot spot, it is an very important.

B

Issue why we can see this image.

A

All the retail work is wonder.

B

Why you shouldn't note that means you will.

A

Resort to some problems, the first is the single number will become the bottleneck of the whole class and.

B

A

After the data comes to the KP.

B

You need to carry it out because.

A

You have large amount of data, so this.

B

Is not recommended or.

A

If we discover this problem, how should we address it.

B

A

Kind of reasons want to cause such problems. Francis the object, I time is now.

B

A

Will fight that the time.

B

Is incremental.

B

A

Those will be in the same region.

B

A

Auto incremental ID so similarly.

C

A

Be written wavy in this region.

B

For example, if we do not know whether.

A

You have the auto incremental ID or update now update time, so if we can use a simple way to find how to identify, if we have problems of.

B

Hotspot, so if we.

A

B

We see this picture.

A

Granting hotspot? Because we have the.

C

Real-Time monitoring, your arms, don't and.

A

Applies you bring me this website, it is for.

B

B

To implement raft and.

A

State to apply.

C

The state of so.

A

Many they are for right.

C

B

A

B

Balanced to be different, pine.

A

Kv, for example, one is too high on the other, is too low. That means we encounter the problems of hot spot.

B

A

Is the same, there are two symmetric.

B

A

First is can't.

B

A

His storage read, cool, handles the kV read and if it is CPU is not balanced, then you will also encounter a problem of hot spot.

B

Well, I didn't introduce coprocessor.

A

Just now, actually it does not support transactional, but also they interface and the framework for calculation.

B

C

A

I'll give you an example, for example five-time KP. Actually, originally it is storage block ITP, so there will be some complex algorithms calculation with unit, for example, to do some some work and.

C

B

A

Process, if you.

B

Make a huge cannot understand data so.

A

That it is hard for you to calculate and it is more reliable because first, it's been with, cannot suffer.

B

A

Time TB becomes a bottleneck for computing.

B

A

In the tank we can support vehicle processing framework and you tell it what how to process this data so magic can accomplish this.

B

A

Example, it's going to calculate the average.

A

Return be after getting this result. It can do second side. Calculation.

B

A

We can't find we have core processor, CPU monitoring.

B

A

If you yourself only use transaction, oh hi, KB interface, maybe you can only notice them read, hold CPUs.

A

After talking so much about policy and time fee, if there is an.

B

A

C

A

C

Issue, actually, we have because it.

B

A

Record with me to update the region in the past, a period of time is traffic of the read and write its read and write flow.

A

After sending it to the.

B

Pda it can calculation I presents.

A

The classes which are easy.

B

C

Channel busy returning to automatically.

B

Classified they read.

A

And write flow.

C

Or traffic I'm.

B

To ensure that it.

C

Would not be concentrated.

A

In one small region for example, way you only have 100.

B

A

We can't constantly beat the 100 megawatt I'm gonna do megabits, so you will encounter the problem.

A

B

We have the automated process and we.

A

Have a manual process.

C

A

The PT control- this is also a 12 of us.

B

A

Out the hottest.

B

A

Example, if it is very small, with a small number of baton and.

A

Actually, it cannot.

B

Be caught America Li, but.

A

We will seek to realize this function in future. Yes,.

C

A

So if you genius, we will work on this phone and present the current version you can manually.

B

A

B

The segmentation I do not need to do any more. So after the calculation.

B

A

Can balance the hot regions between time KB's.

B

A

Pd will learns from the collected they can distinguish hot read.

B

Or write regions.

A

B

Example, anticipate this region by hand and to.

A

Have the split region, the operator.

B

Absolute region so.

A

B

C

A

C

A

The policy whether.

B

A

Is the precise consolation or whether to split in which science? Well? It's? Okay, it is just two strategies and you should really split the region by hunt.

A

B

A

Just now we talked about the transfer.

B

A

That means two.

B

Lakes, the leader and the.

A

Second, we also have.

B

A

Pia for certain Pia or certain datum.

B

A

Also provide some manual balance for this operator and Trust for Pia. So, according to the.

B

A

You can just operate it directly.

B

And the next part is about.

A

The performance tuning, how we identify the bottlenecks and how to fix those fortune acts.

A

Because you have got to start and then thread pool.

B

A

They are high, they of the bottleneck. The rev stalls were full of the apply.

B

So, for example, your disk.

A

I/O, if it is the bottleneck um or if the CPU becomes the bottleneck,.

B

A

They are all monitored and this slide shows the precise amount, hopefully time to be or tie client, and you can see there are many thread pools, so Constantino monitoring witch-hunt becomes the bottleneck. You can adjust it through the parameters on the right side.

A

If your disc becomes the bottleneck, you can also adjust the manner for the.

B

A

B

To have a high.

A

Level compression so that he can.

B

A

Space for the distal and if the second usage is high, can use the compression.

B

A

Low CPU cost the compression per level.

A

This is reading teeth, hi kami. We have two folks, we just can't be read and did this cure rate? The parameters are on the right side.

B

A

To the morning.

B

Tell you can observe this parameters? The last.

A

One is rocks, TV.

B

A

B

A

Through adjusting my blog cache, the size of the box, you can add, or you can adjust, they block cache I.

A

B

So in this light.

A

You can see you.

B

Accurately, thank.

A

The DP, its accuracy and its success rate.

B

A

The cash you need to leave some space for your system, the system.

C

A

Cash because, according to this processing from ma'am table and get from blog cash and then.

B

Resolve enough memory for the paid cash so directly.

A

To the paid cash it is compressed and then you decompress.

B

A

Thank you due to the time limit. This is the end of my introduction and I just.

B

Go through it very quickly without too much details, but.

A

If you have any questions, feel free to ask me after the conference. Thank you. Thank you for listening.