|
Sqoop中一大亮点就是可以通过hadoop的mapreduce从关系型数据库中导入数据到HDFS,如此可以加快导入时间.一直想了解MapReduce,所以也仔细的阅读了下相关代码,整理成这篇博客.
一.原理:
Sqoop在import时,需要制定split-by参数.Sqoop根据不同的split-by参数值来进行切分,然后将切分出来的区域分配到不同map中.每个map中再处理数据库中获取的一行一行的值,写入到HDFS中.同时split-by根据不同的参数类型有不同的切分方法,如比较简单的int型,Sqoop会取最大和最小split-by字段值,然后根据传入的num-mappers来确定划分几个区域。比如select max(split_by),min(split-by) from得到的max(split-by)和min(split-by)分别为1000和1,而num-mappers为2的话,则会分成两个区域(1,500)和(501-100),同时也会分成2个sql给2个map去进行导入操作,分别为select XXX from table where split-by>=1 and split-by<500和select XXX from table where split-by>=501 and split-by<=1000.最后每个map各自获取各自SQL中的数据进行导入工作。
二.mapreduce job所需要的各种参数在Sqoop中的实现
1)InputFormatClass
com.cloudera.sqoop.mapreduce.db.DataDrivenDBInputFormat
2)OutputFormatClass
1)TextFile
com.cloudera.sqoop.mapreduce.RawKeyTextOutputFormat
2)SequenceFile
org.apache.hadoop.mapreduce.lib.output.SequenceFileOutputFormat
3)AvroDataFile
com.cloudera.sqoop.mapreduce.AvroOutputFormat
3)Mapper
1)TextFile
com.cloudera.sqoop.mapreduce.TextImportMapper
2)SequenceFile
com.cloudera.sqoop.mapreduce.SequenceFileImportMapper
3)AvroDataFile
com.cloudera.sqoop.mapreduce.AvroImportMapper
4)taskNumbers
1)mapred.map.tasks (对应num-mappers参数)
2)job.setNumReduceTasks(0);
这里以我命令行:import --connect jdbc:mysql://localhost/sqoop_datas --username root --password 123456 --query "select sqoop_1.id as foo_id, sqoop_2.id as bar_id from sqoop_1 ,sqoop_2 WHERE $CONDITIONS" --target-dir /tmp/sqoop/foo2 -split-by sqoop_1.id --hadoop-home=/home/guoyun/Downloads/hadoop-0.20.2-CDH3B4 --num-mappers 2
注:红色部分参数,后接根据我的命令衍生的参数值
1)设置Input
DataDrivenImportJob.configureInputFormat(Job job, String tableName,String tableClassName, String splitByCol)
a)DBConfiguration.configureDB(Configuration conf, String driverClass,
String dbUrl, String userName, String passwd, Integer fetchSize)
1).mapreduce.jdbc.driver.class com.mysql.jdbc.Driver
2).mapreduce.jdbc.url jdbc:mysql://localhost/sqoop_datas
3).mapreduce.jdbc.username root
4).mapreduce.jdbc.password 123456
5).mapreduce.jdbc.fetchsize -2147483648
b)DataDrivenDBInputFormat.setInput(Job job,Class<? extends DBWritable> inputClass,
String inputQuery, String inputBoundingQuery)
1)job.setInputFormatClass(DBInputFormat.class);
2)mapred.jdbc.input.bounding.query SELECT MIN(sqoop_1.id), MAX(sqoop_2.id) FROM (select sqoop_1.id as foo_id, sqoop_2.id as bar_id from sqoop_1 ,sqoop_2 WHERE (1 = 1) ) AS t1
3)job.setInputFormatClass(DataDrivenDBInputFormat.class);
4)mapreduce.jdbc.input.orderby sqoop_1.id
c)mapreduce.jdbc.input.class QueryResult
d)sqoop.inline.lob.length.max 16777216
3)job.setInputFormatClass(inputFormatClass); class com.cloudera.sqoop.mapreduce.db.DataDrivenDBInputFormat
2)设置Output
ImportJobBase.configureOutputFormat(Job job, String tableName,String tableClassName)
a)job.setOutputFormatClass(getOutputFormatClass());
b)FileOutputFormat.setOutputCompressorClass(job, codecClass);
c)SequenceFileOutputFormat.setOutputCompressionType(job,CompressionType.BLOCK);
d)FileOutputFormat.setOutputPath(job, outputPath);
3)设置Map
DataDrivenImportJob.configureMapper(Job job, String tableName,String tableClassName)
a) job.setOutputKeyClass(Text.class);
b)job.setOutputValueClass(NullWritable.class);
c)job.setMapperClass(com.cloudera.sqoop.mapreduce.TextImportMapper);
4)设置task number
JobBase.configureNumTasks(Job job)
mapred.map.tasks 4
job.setNumReduceTasks(0);
三。大概流程
1.读取要导入数据的表结构,生成运行类,默认是QueryResult,打成jar包,然后提交给Hadoop
2.设置好job,主要也就是设置好以上二中的各个参数
3.这里就由Hadoop来执行MapReduce来执行Import命令了,
1)首先要对数据进行切分,也就是DataSplit
DataDrivenDBInputFormat.getSplits(JobContext job)
2)切分好范围后,写入范围,以便读取
DataDrivenDBInputFormat.write(DataOutput output)
这里是lowerBoundQuery and upperBoundQuery
3)读取以上2)写入的范围
DataDrivenDBInputFormat.readFields(DataInput input)
4)然后创建RecordReader从数据库中读取数据
DataDrivenDBInputFormat.createRecordReader(InputSplit split,TaskAttemptContext context)
5)创建Map
TextImportMapper.setup(Context context)
6)RecordReader一行一行从关系型数据库中读取数据,设置好Map的Key和Value,交给Map
DBRecordReader.nextKeyValue()
7)运行map
TextImportMapper.map(LongWritable key, SqoopRecord val, Context context)
最后生成的Key是行数据,由QueryResult生成,Value是NullWritable.get()
四.总结
通过这些,我大概了解了MapReduce运行流程.但对于Sqoop这种切分方式感觉还是有很大的问题.比如这里根据ID范围来切分,如此切分出来的数据会很不平均,比如min(split-id)=1,max(split-id)=3000,交给三个map来处理。那么范围是(1-1000),(1001-2000),(2001-3000).而假如1001-2000是没有数据,已经被删除了。那么这个map就什么都不能做。而其他map却累的半死。如此就会拖累job的运行结果。我这里说的范围很小,比如有几十亿条数据交给几百个map去做。map一多,如果任务不均衡就会影响进度。看有没有更好的切分方式?比如取样?如此看来,写好map reduce也不简单!
|
|