HBase在滴滴的实践

1 .Apache HBase at DiDi Kang Yuan 

2 .Agenda 1. About Us 2. DiDi HBase Platform 3. Application and Solution 4. Challenges and future 

3 . About Us • DiDi • In <Silicon valley> 

4 .About Us • DiDi • the world’s leading mobile transportation platform • 20 million rides on a daily basis Express Mini Bus Bus Luxury Car Car pool Taxi Premier Designated Trial run driver Rental Car Taxi For ofo sharing Aged Bike 

5 .About Us • Mission To Redefine the Future of Mobility • Vision To become a global leader in smart transportation and automotive technology, the world’s largest operator of vehicle networks and a global leader in smart transportation systems 

6 .About Us • HBase Team：4 Developers • Kang Yuan • Yang Li • Hanzhi Zhang • Jingyi Yao • Attached to BigData Architecture Department • Cooperate with Hadoop/Hive/Spark/Flink/Druid Team closely 

7 . DiDi HBase Platform • Cluster(3) • Storage Cluster – location A • Compute Cluster – location B • Storage Cluster – location B • Location A: the same place with the hadoop Cluster • Location B: for online business or streaming • Application(50+ business and 160+ tables) • Batching Job result storage • Online writing/reading • Persistence for Streaming Jobs • 99.95% available 

8 .DiDi HBase Platform • HBase Version • Based on 0.98.21 • Region Group patch HBASE-6721 • Thrift2 patch • Multi-tenant Problem • A bad table can put down a cluster • We don’t know who the tables belong to 

9 .DiDi HBase Platform • Region Group 

10 .DiDi HBase Platform • Region Group • Isolate important use cases from others • Easy to manage(web ui, user group, operation tools) • Elastic to assign resources • Different Configuration in one Cluster(for different machine types, business, testing etc) • Easy to compute the cost of the business • Easy to upgrade the regionserver in one Group before do it in the whole cluster • Cost(share pool) = TableSize*x x = cost/GB • Cost(specific group ) = Rscount*y y = cost/RS 

11 .DiDi HBase Platform • Improvement • Web UI to show group • MoveTables bug fix • CreateTableHandler fix 

12 .DiDi HBase Platform 

13 .DiDi HBase Platform • DiDi HBase Service 

14 .DiDi HBase Platform Create Project and Tables 

15 .DiDi HBase Platform workflow 

16 .DiDi HBase Platform Monitor your tables Get your bill, user must care for their cost 

17 .DiDi HBase Platform • Phoenix • Advantage • Easy to use for RDBMS User(jdbc、sql) • Auto salting table for performance and hot spot avoiding • Like a Big Mysql(One sentence to explain to our users) • Disadvantage • Some bug like ordering vector item • Unstable statistic info caching • No good in Join case • So many other hotter system :Presto/Impala/Spark SQL/Kylin • Successful Use • Row Timestamp • Multidimensional Table Schema 

18 .DiDi HBase Platform • Phoenix(more customers recently) • Row Timestamp for metrics • Monitoring table write/read/storage • Easy to compute avg, max, min for metrics • Quick to query recently data • Multi-dimension Table Schema • MR/Spark Job to compute BI reporting data • Many demission combination result like city, gender, age, business type • Primary Key: JobID, date, dimission1, dimission2, dimission3… • Value: dimNameArray, valueArray • This can fit nearly all the Multi-dimension reporting business 

19 .DiDi HBase Platform • Client Access • Multiple Languages Clients • C++, Go, Python, PHP • Thritf2, QueryServer • Security(ACL) 

20 .Application and Solution(Hadoop Monitor) • Hadoop Monitor • Help hadoop to query their fsimage and jobhistory • BI for Hadoop manager • Store data in phoenix 

21 .Application and Solution 

22 .Application and Solution(Gis Query) • GPS 

23 .Application and Solution • GPS • Query Model • Rowkey：ID+Timestamp • Rowkey：Reversed GeoHash+timestamp+ID • GeoHash • A index in two dimensions • Fit HBase rowkey prefect • Point 1: same prefix code result in a nearby place • Point 2: query rowkey prefix can location a region whose area decided by prefix length 

24 .Application and Solution • GPS • GeoHash 

25 .Application and Solution(Online Machine Learning) • ETA(Estimated Time of Arrival) • Origin data collection • ETL • Feature extraction • Storage • Model Training 

26 .Application and Solution • ETA(Estimated Time of Arrival) • Training data by spark, every 30 minutes • Pick up data by city from HBase in 5 minutes • Compute ETA in 25 minutes • Rowkey: Salting+CityId+Type0+Type1+Type2+Timestamp • Columns: Order,Feature • Every day HBase data will be dumped into HDFS for offline training 

27 .Application and Solution(Image) • Traffic in Cloud • High Volume Throughput, Little Read • Read via 8 Thrift nodes • Road traffic info • POI data • Heat-map • Write with Spark Job 

28 .Application and Solution • Architecture 

29 .Challenges and future • More connect with other bigdata framework • Hive Phoenix/HBase Handler • Integration with hive • Use Hive sql to query phoenix • Easy to load data from hadoop cluster • Join with Hive table • Spark Phoenix/HBase Handler