Rust Programming Language FOSDEM Rust Devroom 2018, 6 Feb 2018

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Siddon Tang - TiKV - building a distributed key-value store with Rust

Description

A transactional key-value store powered by Raft

It’s not an easy thing to build a modern Key-Value database which supports the distributed transaction, horizontal scalability, etc. But this is exactly what we are doing and we have built such a database from scratch using Rust. The database is named TiKV. In this talk, I will share how we use Rust to build the storage, to support replication across geographically distributed data networks, to implement an RPC framework, to inject failure for tests, and to monitor the key metrics of the whole cluster.

To build a distributed Key-Value store from scratch, we need to consider many things. In this talk, I will share with you the following experiences when we build TiKV.

Siddon Tang is the Chief Engineer of PingCAP. He has been working on TiDB, the next generation HTAP database, and TiKV, a distributed key-value store. He is an open source lover and has developed quite a few Rust libraries like rust-prometheus, rust-grpc, rust-rocksdb, etc.

Rust devroom @ FOSDEM 2018
https://fosdem.org/2018/schedule/event/rust_distributed_kv_store/

These talks have been recorded at FOSDEM (https://fosdem.org/)
This work is licensed under the Creative Commons Attribution 2.0 Belgium Licence.
To view a copy of this licence, visit http://creativecommons.org/licenses/by/2.0/be/deed.en

A

Hey mask you me sis, to see your hair today. I will talk about how we use rah to build a distributed transaction, a key value database. Before we start, let me introduce myself. My name is Tong nil, a Chinese guy, as you can see, an underachiever I had a pink hair and I have been dirty by the distributed database pidb and the Tai Chi V, and we also open source an hour and have developed some open source, a black write with TB Comalcalco master, codex arch and reform issues, et cetera, /.

A

That's all this is agenda at first I will give you a brief introduction about.

A

Why would you use tubular transactional criminal database? The problem for us, yeah I, will show you the hierarchy. We sought for the database and, alas, I will show you how we combine them war into our database. Ok, that's so when we want to build a distributed, transactional key value database. We found that we need a tool, will meet many many problem and we needed to conquer many many challenges that face. Oh, why is that like?

A

When we must guarantee our database is a data consistency and our database is scalable under stable and has a high performance and has a higher order ability. Of course, we must read our database, follow the acid compliance and many other a prominent we needed to conquer so that night they this award and I may have for us. So what can we do in a traditional Chinese saying a high building or low foundation? So there are many medicines we needed to conquer and but we camp, you need from bottom to the top with different hierarchies.

A

So, let's begin, but at the first, but at first we needed to choose a language to be late. It's a rust here. Okay,.

A

Okay, you guys are familiar with this, so I won't talk about these animal enough. Forgive me! So let you slack you build our database from scratch. The first thing is that, because we are database, the for the first problem is how we can save our data in the local machine and recall there can be no database. So we are choose available story engine. There are many squeal engines in a war neck Rock to me like a lever de Guerre tiger and the eternal row under you can use.

A

Even beauty is by yourself, but here, which shows the rock to be. Why rock to me use a sm-3, as you can see, Alison tree had a high performance for the ride is traumatic for the random ride and had a good performance for the read and the rock 3 medium itself that loss of optimization to speed up the rocks DB and the provided a lot of features about knack in jet ski the fire team is branch.

A

Cue later the user build their service on top easily on top of a rock to DB, and a lot of reason is that ought to be. Is very stable and I you to get user in the my rocks and rock rock TV, so we departed to the rock to be headed up is a project where we call the rock TV is a stapler, a library, and we work our own rust wiper.

A

Here you can have a try if you want to use a rock to be in your router project rocks me is cool, but you can only stay with your data in loco one machine. If the one more train crashed oak, your data is lost, so here we need that. You find a way to replicate our data into Marty machine to guarantee our data safety. Of course, we must guarantee our data consistency. How we can you pair with you the raft anybody here know rock yeah?

A

Oh yes, that cool, ruff ruff is a common-sense distributing the constants of altruism either use. If on the replica Lord, when the client won't even write something to the service, and we were you first, we were use a raft to append the comment to the rough log and the angular radical algorithm to replicated a raft command into the mounting machines.

A

After majority of the load except a rough log, we can sing the rough log is committed and can apply the rock to the same machine then return the read our client using rather can't help us to keep the data. Consistency and doodle replication automatically roughly is good, but it can't fade up in the constable or your data. In one motion we run your data grows huge and huge. So how how we can do here? We can, when you take root large and a large, we can submit our data into market graphic group.

A

Here you can see that we we so treat our water in her as a invading. He ordered hash map, sorry map industry and it's a bit. Our water data include different experience and the each clear and crisp on get to a rough group. Here we use again, we use a market rock q, suppose that data horizontal scalability. So how can I do this example?

A

If we'll all, we have three load ABC and we add a new load T, and we want to transfer more our data reading to the data from C to the team, how we can do. Luckily, we can use a raft interlock every algorithm membership change necklace.

A

We can use a raft of membership change, completely change a load directly to add a new replica of reading to unload of T yeah. We can use a raft how many shiny ribbon load to remove the replica from C, as you can see using rough that we can. We easily transfer the data from Lois agility and the to dock, that your Anna do that it, our horizontal scaling, V, very easy, that's cool and we build our own. The ruff-ruff implication here is very stable and you can.

A

Even if you want to butyl your use, a consensus algorithm in your service, you can have a try and you use it under. You can even using in your production, environment,.

A

Now we have Martin raft, you will support either transaction either horizontal stability, but a lot of problem is that, therefore, that distributed transaction.

A

Let me consider the following case, for example, for the bank transfer. If you want to transfer some celery from account to account be, but I'm not keep both a and the B are in deeper in the load. So here we are in different imarco Marty rough groups under here here we needed to come so the problem that how we can keep the data consistency, quality, market, low crossing market load, crossing Marty rough groups- that's easy! We can use the traditional two-phase commit to PC, but as you can see, that QBC has some problems.

A

So here we use optimizer two-phase, commit inspired by Google pack later and it has a market version, concurrent control and a provider, so we're celebrity or isolation and supporter or committee transaction I. Don't I won't talk more about our distributed transactions here we call it a lot relative that in this topic, so I was to give it tsuki be. If you pay attention to it, you can search apoplectic paper and the more so. No seem to go the better.

A

A lot of problem is that, because we are distributed at ransack, distribute the database and we have many machines and many services and each services needed to commune with each other through like a work. So how we can do this here. We choose the RPC Java C is a PC framework developed by Google and had used in many famous projects. Like you, ladies, like if CD and the charity has many language support later you can use Java.

A

You can use a prop last go long, PHP Python, you are set for many language supported and job say it's based on HTTP queue, so you can get a benefit at you directly lack charity to s-sorry HTTP, so yeah we did I show you you did you see, jobs, Google has provided a state official state job is a library has full API and you Larry Klein streaming server streaming and the duplex screaming they are very powerful but anarchy. This idea, this API, are asynchronous, as you can see you liking. A synchro.

A

Lured called is a letter mayor for us and you can even get stuck into the cone go back here, but luckily in Rasta we can use futures here you simply simple example: when we call the Euler API, you know we return a future and then we can use the future wait and the world a poodle future later. So here we use a lot of futures to provided a synchronous, API for usually use, and it will build. Our own job is a wrapper here.

A

If you want to use your job using your own service to people you're, obviously from work, please have a try.

A

So that's the database take take shift so velara thing that we can. The terrorize can work well, but a lot of things is that how we can guarantee our database work we're here. We can use some monitor two more hours to check our database. Can work well warlord. Currently, chrome issues is thermal, is a famous monitor in the world loud, so we decided to use a pro me shoot some provided for at a counter educates women, his program and a summary below here we only with chemistry, counter cage and his program.

A

There is a simple example, and you can see you can use a rigid counter to reach together counter a premiership counter and increase min the counter. We also build our own premiership decline in the rust. You can have our try and this rough clay is interested in the promotion of a short walk or a hive, or you know already, so you can try. Have it right now.

A

Finally, let's go here: testy you might appealing. Testing is the most important thing in to build a distributed database.

A

Maybe we need a queue because we needed because we are building a database and the way needed our customer trust us, and we must ensure that our to it, they are data in our database cannot be lost, cannot be corrupted so how we can do. How can we do this I think the only way is to do born you test him and the way near that you Pro do many many tests him and for you uncle we needed to do. You need traditional unit testing integration, testing and the human.

A

We can do chaos testing many many more, but I don't cover them all here. But here we mention about a fair point, injection testing. The point that your injection is the sphere is reappointing inspired by a freebsd fear, and you can see that you can pull here here. We, you regret, will reach a selfie of winter like in the function, foo and trigger failure outside when the program ran and enter the function. Panic, oh, that's cool! So this our wrath of your pointer inflammation.

A

If you want to ingest ama failure into a service and do some funny things and you can have a try, so what that's? What and now we can I have making the world hierarchy about and now let's convert comment him together. This is a whole architecture, a war from the pollen to Lata, Rock TB, to save data, using raft here, replicator and a doodle data horizontal and had Emma City and the transfer distribute transaction API you Jabba took for communication and user promise, use for Molitor, that's cool, so least of all the whole thing.

A

Peggy Peggy Lee is lucky to be here. Transactional database and I'll call you the reading in raft. Thank you. Many user have user in Thai, Cavey many user degrees using many in the production environment and, as I can see, I can know that one Thai cubic Lassiter had allowed deploy about 140 machines for one time three cluster in production and from language, hello. One complains that their data is both that's cool, but this is law this in Lord.

A

Our final goal here, I will ask you a question which I don't mention, which I don't ask you before at the first? That's why we were. We want to build a distributed, a transactional, cable database, we're not only for the be we aim higher.

A

First, we want to build a distribute in a relational database. This is a tidy be tidy. B will build my circle layer on top of a high key v2, provided a distributed online transaction processing solution. You can use it high DB and not using my circuit or actually with Jionni with load compatible Lea, and we was a one computer distributor, analytical database. This is a tie Spock we can. You can run the Apache smart on top of attack Yuma directly and that you provide, or anyway provided or distributed online analytic database solution.

A

So what also hear what they are we're going to go? We wanted to go to build a hybrid transactional, analytic, processing database.

A

This is a big future of war, as you can see that we have a tidy be to provide an OLTP and type Spock. You provide OAP and the you can. Even you want to use or rock transaction you can use tagowit directly. This is a whole picture under this is Albert I'll go oh, and an argument goal is to use the one database for your data. Yes, that's what thank you.

A

By the way we are hearing you, we are talking to Tyler you're, going to pay increase signature earth.

A

You see you ask me that that how do how we handle the deadlock in a distributed transaction that look well before the hour per second mole is optimistic or mystic transaction.

A

So, if you make so for transact long to handle a and the key, a and a beer and second behind along to also long 208 and a B, and we were at first, we have sort of a the world of keys, Mac a and a B and then commit a yet so because the committed key is not sorted, so you can't be taught that log yeah yeah order, keep you what committed is yes.

A

A

You asked me that we all the ladies hold up, select short isolation, so how we can support serious success.

A

Sorry si si SSI. Yes now, because you peek, as you can see, we some whole MVCC and everything is only for si. So if we want to use support SSI when we move, we must measure user to write a circle like slack for update expecially to support either SSI. That's a that's. A mediator is the only singer only way we can do here now.

B

C

A

Asked me that mmm we have building or C Webber, but um many other people and sorry I, don't know him but I know he's the water of niño. I had pure pure roster inflammation, but the white way. You still choose that you have C well her because we have we build a CL. We want to use the IBC about one years ago. At that time, there's no stable inflammation and- and there are only over the only stable inflammation, either say you have it's a Seiko libraries.

A

We have to use it evil and but later, if you, the pure wrath information is coming, and we can very a commercial cry to use a there's. Our player.

A

Oh, yes, you asked me which secretary I would do we support I, said I'm, sorry Heidi, be we built on my second layer on top take away, so you can ward us so-called dialogue over type. My circle. We support it, but we don't stop holder. Sorry, we don't support, wait. Sorry, I I have a tooth.

A

We support joining. It's burning lots of how you might circle some holding post critical. We lost somehow Hawaii we lost about closet door, yeah!

A

Maybe warm the circle function. We support it. I know so you can transfer. You can convert your you change your bezel ad from my circle to my career, you're, being away from my cell go to the tiny bees it generally yeah.

D

A

You you, you asked me that that ty kv supporter some API for the transactional yeah, of course I, because, as you can see, there take away much probably in the transactional API so like so like user you'd like add a menu. Three, not sorry like this user. Can you do likely to pay you the begin to start transaction and then you get using courgette and I use course at and again committed transaction date all opposed to supporting on you. Tacky me.

A

You asked me that you thought spikey. We support beam.

E

My transaction.

A

Transaction, oh.

A

My togepi, not yes, oh we have a, we have a penal affiliation, but not in tightly be you see in tidy, be because you can enter a work or the circle you throat ID, be so can't hairy. We can record a solid or the circle or not do record in our yoga implementation, but at a spin or it's not different, it's different from the magical, okay and we can use Urbino under using. You can use a pin or two synchro synchronised order.

A

Your data from the tiny be to our lab to our there are surveys, the neck single Tula, Kafka or you back, search or even another tidy be we can use the pin lock. Yes,.

A

Yes, that's a good question: Mac I! Don't mention present here entirely because relative in this topic and butter, you can I'm sorry I, don't make sure I tidy. Taking me a little only provided, but I try. Second API! You do provide and compensated if yeah. Well, you look casada casal is the same, does concern in HBase and you can push down your some logic to the activity rectory and do some calculation and a retainer which are we already do this, but I don't mission is.

E