Apache Cassandra Cassandra Summit EU 2013, 12 Nov 2013

Previous Meeting

⏯

youtube image

►

From YouTube: C* Summit EU 2013: Using Cassandra in a Telco Storage System

Description

Speaker: Pavel Pontryagin, Senior Engineer at Peter-Service
Slides: http://www.slideshare.net/planetcassandra/c-summit-eu-2013-using-cassandra-in-a-telco-storage-system
Data volume grows and in telecommunication area it is painful to support and scale RDBMS systems. This presentation shows how we switched from SQL to NoSQL. This will be an overview of aspects: * how we model schema for call data using NoSQL vs SQL. * what hardware architecture we use * NoSQL vs SQL insert-select performance * how we store graph data using C*.

A

So again, thank you, everybody for eating destroyed presentation. My name is Pai fantaken I'm from Peter serving rushes in Petersburg, and my presentation is about Alice from Porto as either an hour to the communication air. So, let's start this is what we shall discuss today, our discussion without what architecture? Videos, how we model sequel and no signaling and a couple of performance issues.

A

So does it? Let's move on well start from the both signal and the couple words about the Syria. What is the cinema actually is a walk from the sistrum LOL.

A

And each time the person makes a call with greater than 30 degree part is it to a DS. So, by the way, a question in case of ties shall receive.

A

Okay, so empanada nervous every time this home yo connect with our simulator. She walk down names. This entry contains cosmic only party started duration and transformation. This is a typical CDR in case of a mobile operator. He walked busy, and so this fantasy identified a towel which, when used to go so that's we reduced reduction procedure and by the way, this information star formation, available source of information for ohio operator, because it contains usual information of between your buttons and so on, reaching to stop to it animation drum.

A

So that's because I am here, we started about 10 years ago from a hardware, software, complex or mobile operators for marketing purposes, and what work is ten years ago we didn't, it did not know about no sequel. The only choice was talk of additional databases to tell it was absolutely ok. He has one server walled city, our data files into data storage.

A

Actually it was a vm, so for amount of data, 100 million records study, it works, absolutely, ok, don't work perfectly, but time flew and the amount of data in the east. So we added additional expansion enclosures to our system, edit edit edit edit and finally, the scale because edition even closer to give us kissing speed, so the speed limit posts in 10,000 tables the second, let me scale additionally, this kima give us problems.

A

For instance, if one a device dice the whole system and click to upgrade it organized a divinatory, always stop.

A

So we faced with challenges how to end this little history capacity, increase the viability, hospitalization and additionally be a bonus challenges outlet in functionality and, of course, outages lost. So we turned our heads to you: evolve no secret and which seemed very promising for a supposed cuz I animals move on to no sequel ball.

A

Be made this schema, which group works perfectly for us? What is it in here? This is your files in a unified text from article latent, which is placed on crossref s. Go service is distributed for horizontal critical scalable file system, so after file is there, it is visible for all notes in our cluster and on every node V wrong owl, our Lord, which is based on a sauna as a stable, simple as of the tiger, and this dead files really would go to Cassandra cluster.

A

Additionally, there once the main purposes for our system is created, because so on, every node fearon search engine which is available. Why our budget communication horrible. So this architecture solves almost soft sobs. Almost every hour program we can. We can easily scale and the addition of each node gives us in peace mississippian 5000 decades per second, since they use the application factor of 2 and therefore one mode. It's not simply not critically.

A

Ok, and if we talk about price I, think if you compare or comparison assembler think it's almost where everybody is, whether that's it so few words about how we modeled model CDR data. In sequel to get started signal world, we had big event table which was partitioned by started, and we in this each interesting account number or in vivo onion and the typical.

A

It was like give me a call, so this number during the last that so we needed to former in no sequel, and it doesn't in here it is the same formation, but in other way, so we break in two stages. First and he'll hold the hobo in to be calm, family and so I t is a distributed key based on Twitter convoy and when we use the hobo after entry index all the interesting things.

A

In our case it is number so this company is partitioned repetition, video systems, oh so consider this partition by heart and so hokey is a number column, contains composite and supporting supporting voice.

A

The super thick, because in this hood we can rest additional information and during search we do not need to create this person over to host host table. You can use information just on it. Let's get up to speed up Stephen and, let's see a small example: ok switch reduces near effort. We put into this so contaminant, then the indexes we put for number one we could call name which is like them varsity and come worry. Additionally, the inverter same moment we index and the second part from the scope for future calls is also important.

A

So we just witnessed this information. So, let's move on it's a small Commission. How many days we use Hector between so there's no java example. What may say it was so: let's define company.

A

Okie leave it will be by number reset time period and ok, let's get the result, nothing we can do. You won't weaken support formation between or we can community. This is deep. Bro company thought events on us so now part which the last was very challenging.

A

Second, all cool boys are out of the box in the innocent for text searching so they have context indexes and so on and in the center out of the books. Didn't have this functionality. So if you want to eat with when you squeeze like that and non-white or contain something its opposite, so we used data specs enterprise solution.

A

They have it so Lucas and we solve this problem so for this quiz speaker, additional complement, which has the solar secondary index so in degree, numbers like like this quick meeting but is confirm receipt biggest, and we this list to oil company and in case of exact much p, we just really spent confident this whole family is, is.

A

An but it cost another problem how how did estas we try to just insert using Hector, OSS tables or whatever into so account and deposit hematocrit is all down born, so it made us know Taylor if activity they do not need real time in this, and so with. We saw a young family and rewrite two went on to develop things, and we know what we know: the LED synchronize, these two tables, no one. So pretty simple.

A

So now a one more example of this model, this for estimation of where da water warden so again put it in detail, detailed information. We use our current family and, for my agree, we use so cold in this company. Not this case it is called call SWAT. Cell Bruce was partition. Data and Loki is taxed and are heated to minute.

A

We do it through shuttin the role, because if the Rope is extremely wide because, for instance, using the second base station can produce a lot of Records and alway station can be used much more levels, so we're going to write better shorten that doesn't the network so creating a use. Indeed, and because of composite, we can use Queenie, we can create using the block was celebrating what you, even just what it's a key and I think it also can do whatever we want, and here is the jungle.

A

So, let's move on I have talked about functionality and requirements to our complex supposed to add very fashionable now.

A

Did for subscribers on prediction, and so on? So it's obvious that the correlator is itself an H in the car and the subscribers are vertices orphanage, and the points in our case is that all the ages are weighted with is none. Of course, our graph is directed it's important for us and also having their a page type, which is voice for SMS or whatever, and I kept on do I that our system doesn't reproduce in.

A

Norfolk Norfolk jail or whatever life, it is just lost our nation and our outside subsystem Chris. This data analyzed to use mines between celerity and so on and features it like this. So so we just stop information and we do it in CLE. In this way, I have to notice that Julian data, but in every event, it's very important to know how it works physically.

A

In our case, we noticed that each comb is divided in two parts which they can be divided into pass on, one switch and or another switch, and we needed to take it in because because we need to use weights of Ages, so we do it to like this. For instance,.

A

Maybe secret sauce later phone message is swagger and we put it down like this, where you be his countess, yeah Tampa wait'll, you see another decade from another speech and we need to put it down also we it with like this for the same okie with we have another car, but up in there.

A

The last count v right now, the subscriber gisky and the first day. So since discounters, when we tweeted this data and summarize it and use it in our analysis, we choose the maximum of this counters and use it in our future organism. So that's about how many noble grandly the solon cassandra is. This is pretty flexible, so this is smoked salmon. We will kill us first.

A

I have to endure one that the total cost of hardware horrible complexes, a boxing approximately equal to cost of hardware of Cassandra congress and 2nd that a lot of use, expensive, SAS, 50, fifteen, fifteen thousand rpms per second east and for assembly we use cheaper saturdays.

A

So it is, you see her go since our the freedom to the tourism is Harvey scale info Cassandra hundred million the note and easily the scale such Cree here speedo of a queen is also very good. Oh no, it's very plain old school and since I don't want to speak about it and then Cassandra. As you see it's pretty fast, because some is Queenie variable.

A

So that's it about my.

A

Report is tough day CBC patient portal to conserve the candy, not painful, and even.

B

From so okay, that's it in case of questions. Please don't hesitate to ask you. Imagine the very big that you have been using like some distributed file system. Didn't you get the higher latencies with a solution.

A

Maybe my answer my technical.

A

Question is why this is invigorating twitter com fight. There are leather 22 to Melissa kun. You can see other in the trap, k, /, dadnab, snowflake and solution source non straight up two minutes seconds.

C

Are you running solder and Cassandra notes are Cassandra and so they're, not even the same bones or are they deployed as a separate datacenters and all.

A

We use a below to be on Sol, Sol, Sol, others, water.

C

In your face, okay promise: you, when you use it and you using it inicia tokens and nobody knows in our case we did.

A

Not we should we be used to virtual nodes, but.

C

Then the recommendation is, if you deploy disorder on Cassandra, the same of the same denotes with you get the phones over and market I.

A

C

B

Home or 300.

C

Billion rice, for note, is that the statistics inside is that is a per day. Was a total number of us were no more 300.

A

Million, who knows today, para he ghosted anyway about me me me right is the whole record right or just the whole record. It's it's codec was illnesses. We use about seven minutes. If what happened. Additionally, we use some extra mission for bodies.

A

Are you using datastax enterprise.

C

Yes, inquests lost motion is the graphing system. Each other immigration system is that you know which.

A

Driving system is it's not our own, but actually we are employers.

A

A

12.0 weapon 2.0.

A

Educated for new requirements for all in excess of the indexing of your old data next week, your business requires miracles.

C

From last year's data, how do you go back to build official video from server I understand.

A

Actually, our business piece doesn't need to do it. So in this case we just see the Gulf looking.

A

Relatively new wife, yeah I have to undo and it's not immune system. So we have some developed in nano failing.

A

Bean paste, so it's not used to doing no, it's not forbidden, it doesn't know why it's for our antics for Cosmo prediction systems and it can be seen as core for future analytics blue I.

C

So you can't use it over the hard when you're in your current deployment. Now we're not so so. It's only only.

C

Normal special powers, I has a heinous, a failure is no failures. Happy needed are good players. Have you had a changes and pasta.

A

Because the point is that yes,.

B

It's not billing system.

A

In case of a huge total disaster, data center has explored, become a DVD CD, album box, which involved again and actually actually, we can store that as many rows as the gives to ask, but interesting me, Loretta is interested in only half a year, maybe even more in three months. So we can. We do this tomorrow and quickly over the game, we're saying after frequency drop data. No, no, it all depends on the customer. If he says okay, I can wait for it. Just so.

C

You don't have to retain data processing a period of time.

C

It's a lot! So do you need to keep your data for specific period of time because you can use the TTL think a son, so the total purchase date? For you see how.

B

C

In your barn, actually.

A

Be an important potential. Employers also be are playing with the TTL. You have to simulate how fast we have a mechanism, a partition tables. So for this one is one continent, and if you want to bring up space, we just talked as a regular factor, o mighty TT edits to show themselves of them so good. But now gonna do you think you'd ever used a sound repellent in the future.

A

That's a very difficult question, because our company- it's very that wire in my my department, doesn't it's not responsible for doing so other people? Maybe if they see that it is good and it works, they will switch. But I cannot answer positional.

A