Apache Cassandra Cassandra Summit 2015, 14 Mar 2016

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Tsinghua University: Two Exemplary Applications in China

Description

Speaker: Jianmin Wang, Professor

In this talk, we will share the experiences of applying Cassandra with two real customers in China. In the first use case, we deployed Cassandra at Sany Group, a leading company of Machinery manufacturing, to manage the sensor data generated by construction machinery. By designing a specific schema and optimizing the write process, we successfully managed over 1.5 billion historical data records and achieved the online write throughput of 10k write operations per second with 5 servers. MapReduce is also used on Cassandra for valued-added services, e.g. operations management, machine failure prediction, and abnormal behavior mining. In the second use case, Cassandra is deployed in the China Meteorological Administration to manage the Meteorological data. We design a hybrid schema to support both slice query and time window based query efficiently. Also, we explored the optimized compaction and deletion strategy for meteorological data in this case.

A

A

So please give me a warm welcome for gentlemen, one exploring two applications in China. Take it away.

B

Thank you very much good afternoon, dear friends, I'm Jimmy well from chiba university you may have experienced, and the heard about the use cases of contenders in us or other places today, I'm pleasure, to introduce the two exemplary Cassandra applications in China in the opposite side of the Pacific metric consists two parts.

B

First, who we are. We are from the yeasty cute for data science, Xinhua University, which was founded in april last year. Our missions include recruiting world-class researchers and the engineering industry and the octamer long-term dedication to see some research and the industry practice.

B

The third is leading China's big date strategy, especially for industrial, big data.

B

Nowadays, a lot of people are talking about big data, where the big data comes form. Firstly, data generated by people, especially our internet, the certainly from the online scenario such as Internet to the offline scenario in the real world versus are generated even melted than people do, such as data from equipment, sensors, RFID, etc.

B

Similarly, tons of semi-structured Salt River log are generated by computers.

B

Let's look at the work load of machine generated data, the existed broadly in several industries, such as agricultural utility interview business machine data, is always formatted as time series with device and data fire, timestamp and value as a regular interval as temporal sequence.

B

All adding TPS attribute to each record different from our the structured data like transactional data, yes, erm machine generated data is a part 24 hours a day and the seven days a week continues dream and again a single application. Its volume can easily reach 1 billion record per second that either high frequency, sensors or millions of sensors.

B

The machine didn't need to be processed under persistent in real-time, and the more important date will be also used for complex query and analytics, such as based on the content, their shifts and the spectrum.

B

We talked about missing, teeth, three ways, but it's not just volume velocity and the variety click industrial business, for example, as a merging of several new trends like industrial internet from us industry, 4.0 from Germany people, realize the importance of the in Dutch a big data which they really hive, but the real challenge come from how to generate value from these data rather than just a process and store the data. Yet I will point the big data in industrial business is more than technology.

B

The answer to this question is how to cut line of business involved easily proactively and the deeply let's look at the big data life cycle of was in generate big data. We need the infrastructure to store the data, support concurrent and the complex query. In the first step.

B

Also, we need to pre precise the data to clean it and the receiver it into a proper size and the granularity and steps three and the for the data are ready for access exploration, modeling and analytics.

B

Today, we will folks, are our work specially on taste, storage with Cassandra.

B

Now, let's look at the industry sensor data management with Cassandra in China's any company first, as you know, China is a big industrial manufacturing country. Sandy company is a global company in the construction machinery industry, which is relatively similar with the caterpillar in u.s. I. Think the only there were more than 200,000 active have highway machinery which are made ben seni in more than 150 countries.

B

Currently, we do not only care about the manufacturing process, but also care about the after-sales services for these smart products, for example, we want to reporter working stated status of these mercenaries.

B

We also want to forecast the passport before it appears and to improve our products by analyzing the historical date records to do that. We need to collect and store the data as much as possible.

B

There were many sensors embedded in the machine today made the incentive group when the mercies are working on the construction police, the sensors collected data and the transfer them to send its center bye, sweetie OG gsm internet. Now these diseases are located in the chancel city.

B

The date records the running status of machines, for example, when the machine is traveling, we collect their speed, their location under the fuel load.

B

When the machine is working, we collected up Tess, it's angle, the pump oppressor and the swap currently sending has more than 5,000 the kind of sensors and the machine will produce more than 50 billion records per year.

B

Semigroup begin to master their sensor date from 2008, along with the increment of date, semigroup release the sequel server with Oracle, but it's not enough in 2012 a table in all co-stars tests of billing records. It will spend several 10 minutes. A DBA run a cycle with work losses.

B

As a result, semi only put six months state into Oracle and dump the older data offline from then we try to have senator riblets Oracle with Cassandra. At that time we use Cassandra version 1.2 euro, a very early version of Cassandra.

B

We analyze the characteristic of the sensor date, design, Chris base and the column, families, gamers and explain experimentally imported and a Tesla system performance. This process last two years in 2014 the new system with Cassandra, begin to release Oracle in production system.

B

Now it works well, I think this is our software stack. We used on to hand to handle the sensor data first and then write them to into Cassandra. We also developed some upper cases to analyze the data with the MapReduce framework up Cassandra.

B

In Cassandra, we design each sensor to a column of family. There are four: we have more than 5000 the column family in total. In one case base you can see for each column, family. We use mousse, indeed, as a rocky use, the scheduled time stamp as a column, name and store the sensor value as a column, value.

B

The reason for our schema design has two facets. The first, the Cassandra system before wasn't too, does not support a multiple primary key or clustering key. It means the peep just for a single attribute.

B

Another reason is advantage to Sudan each column, family or say a sensor in the single column. It is a similar design of columnstore to avoid unnecessary access to the other date, since most of access sensor did instant application against a single sensor.

B

I think is a very traditional sensor. Date management application. However, there were also some challenges. The first is a creating scheme or hand.

B

You may think it is a very simple you're our case. We need to create more than 5000 column in batch when we acute create column, family secret statements. We found it spends too much time at first.

B

Each sequence statement is best less than one second after we create about 300 column, family. The time cost for each increased to 10 seconds at last. For this single statement for create quick for creative chordoma family will spend more than 20 minutes.

B

What's the reason we monitor the memory costs and defend at this time, Cassandra use a lot of a memory family. We find the reason when, when server just to the coordinator receives create column, family requests, it will propagate it to other nodes. Meanwhile, gossip a protocol will find the schemer on each node is different. Then it will try to synchronize the incomplete scheme of modification, frequently.

B

The left, a type of figure, shows the original gossip of per protocol post, the gossip protocol and the massive propagation confronted a schemer inconsistency and initialize a carry out.

B

After analyzed, the reason we do some modification to let the gossip take effects only in two cases. The first is when a node does not receive a propagate message. The second is when a node recover from aphelion.

B

In this way, we can make the call my family creasing time cost to be almost consistent.

B

The second challenge is, did inconsistency. We know Cassandra of light eventually consisting model, so inconsistent exist in the system when we imported into Cassandra in a while. We high speed continuously when we cited the red consistent level with one and the replica number is 2, is not many and that the cluster has found notes after loading 30 billion records.

B

We stop the load process and the try to check whether the date is consistency. We find the Dutch a sim select command on different servers. The results before after finishing our repair operation. Even after finishing the repair operation, each server increased 25 of data. It means that only 75 percentage is written successfully in one replica.

B

When the repair process doesn't work on this situation because for each range and each column.

B

Khagendra need to create a local tree and the transfer it to add a note to compare the replica consistency. So in this scenario there is too many column family and to tate too many data range.

B

B

So Cassandra repair process takes a long time and the not very effective.

B

The third challenge is the heater genius note.

B

Our our experiment is so that when we add true, when we add 8 PC servers into existing two powerful server cluster, the overall cluster performance comes even worse, even though we change the data partition according there.

B

According the note capacities, the cluster performance does not improve. Ideally,.

B

The root cause is the mecanim of replication, which means the node is not only responsible. This bond ball for its own data range, but also need to take care of its neighbors.

B

Once a strong server sends the replica request to a week, server is need to wait for its response.

B

When the read consistency is in the quarter of level, it means that the powerful node has to be idle until the busy week node complete its job in this slide. If the node file is the strong note, H will wait for the wicked server node node one.

B

Another issue is how to allocate the number of virtual node, certainly the bigger number, what you know that we have the most base. We can optimize the location of virtual mode, but cassandra has its upward limit for the water node.

B

On the other side, the number of virtual nodes can lead to the overhead under lookup table.

B

Our solution is to combine the position and the range for optimization together.

B

We have published two research papers to solve this challenge with a quadric optimization model for clustering in your validation and a dynamic programming approach to search for the best position when a new server comes in.

B

After overcoming the previous challenges, the semi-group upgrades their sensor date management platform to cassandra successfully.

B

Notice they can put their three years data online with the Cassandra platform.

B

From this any application, we learn more requirements for the industrial Big Data Platform, the first one is more than one or even ten million data points per second, a very common in this area.

B

The second is: we should reply the complex time series query within one second,.

B

The should we should synchronize the ice data processing with the Claude and deep processing the force is even more. The system should provide adaptive, deep data, comprehension capacity.

B

Finally, we should is cute the trajectory based query efficiently.

B

This is a industrial sensor data analytic backlog pipeline in sunny.

B

To generate the real value from data we build are several analytic applications together with semigroup for their R&D quality control and they're after sale, service divisions.

B

This is a first exemplary application.

B

Let us take a look at the example. This kind of a machine requests stable support to pumping concrete, and thus the legs with cylinder are designed to prevent tip over, and that is like will add, adjust to supporting forth. According to signal from inclination angle sensor,.

B

In order to find out the reason, we collect online inclination angle data from more than certain solvents service, concrete purpose through the concrete.

B

B

Analysis we found some reasons. The first is highly unstable distance with miss this pump six widely, while working. The second type is idle instance, whose inclination angle sensor network change. This means the sensor does not work.

B

Based on such observation, we can drill down to investigate the individual, concrete pump trunk, who we, who has the above situation,.

B

The second application is the fourth diagonals.

B

We are time series analysis and the spatial collegiate we find the leakage problem of master cylinders is highly correlated with a high speed rail construction project.

B

The third application is the spare company components, demand forecasting combining the real-time data for mercy's sale, history, holdings of vehicles, environment data, given the gdp date in a different area, we can reduce half of the young active spare part inventory.

B

This is our application in sunny company.

B

Now, let's come to the calamity data management in Cassandra.

B

There are four steps for processing the calamity date. Firstly collected the weather, did most of them comes from the sensors.

B

All the machine generated data. Secondly, transmit these date into cloud data center, then the end-user's most of them are climate specialists can access the data and the pros use the weather report.

B

Now, let's see the car listings of the calamity did calamity did is a hierarchical data, Chanda meter, logical or the magician use seven kinds of data to predict the weather, as you can see, such as the light in indeed grounded arrow logical.

B

And the others, such as satellite radar typhoon and the model data where model data are the most company complicated one and the produced by powerful server clusters.

B

This model did Heiser, has 71 kind of parameters, for example, T 639, the China model did standard, has temperature win the field win for snowfall and the show up for his parameter? There were about 20 levels, which represents the weather condition in different altitudes and for each level. There are many time points.

B

Totally there were sixteen point: five terabytes climate ated start in 2.8 million files per day.

B

Users in China me too logical, other magician need to broth the clam did it quickly.

B

Therefore, they need a high-affinity three date. Catch system to support it, add user to gather data file within 50 milliseconds.

B

For example, if user want to scanned it in the messer order, 3, which means along the attitude level in the right finger, if we organize the data file in a tree, we must travel in the tree structure in different a lot different level. It's very time consuming.

B

In this project, we choose Cassandra because its scalability and redirect efficiency and because the column is sorted, we can organize the data in time order so that it is easy to scan the date you order along some dimension.

B

We know Cassandra use, iOS three structure and there will come and the wheel compact Isis tables into a new one.

B

We also use a Cassandra's new features about the time-based compaction, so we can delete the previous obsoleted easily.

B

This is our schemer, the claim that it is a multi dimension, hierarchical data set, some dimensions are ordered. We use them as a column. Some of them are not. We use them as column families or keys. We use a classroom tree structure. You can see the final.

B

Calamity date are organized in the file space in the right fingers.

B

The results after one year implementation, the system in china meteorological other nutrition is online, which is 100 times faster than the older system.

B

This is the end of my talk. Thank you for your attention. Thank you very much.

A

Any questions for german.

A

Zest can go back, one slide, one.

B

A

Who wants to take a photo?

A

No question any other questions.

C

There any specific reason to maintain the replica of four instead of three the replication factor maintain it is for. Is there any specific sorry it? The replication.

B

C

The application number four is there any specific reason to maintain. It is four instead of three yeah, oh.

B

Yeah, because this application is very missing critical, so they cannot afford losing of their data. So you use.

C

So it should be a order. Number should be better right, 35, rather than four.

C

C

But I think even even with the four or three right, so the Nord down sionary with the quorum is going to be same right.

B

Yes, there is a story because the first they use that the true replicator 2 replicas and after a half years operation.

B

B

Was destroyed? Yes? What? So, after that, they choose the four numbers replicas, because the machines is not very reliable, I think.

D

I'm a whole question right here about challenge number three that you mentioned about heterogeneous nodes. You said that you use quadratic optimization to find the best positions of the virtual servers. Can you elaborate on that a little bit like the best in terms of what challenge three.

B

D

It's a way back, I think.

B

Chemistry: okay,.

D

C

B

Uh-Huh: okay: hey! How to.

B

Sorry cannot fruit think we have to take that one.

A

Off fine, let's.

B

Reverse we trusted go.

A

B

Yes, we just a thing as a mother. This problem as optimization is structure and considering the replica and the replica position on the different note, and we just calculate a linear. The optimax optometric magic is to solve this problem right right.

A

Let's give our speaker one last round of applause. Thank you very much. Thank.

B

B