Hive的Transform和UDF

2018-01-31  本文已影响0人  这个该叫什么呢

UDTF

UDAF

Hive中的TRANSFORM:使用脚本完成Map/Reduce

转自: http://www.coder4.com/archives/4052

首先来看一下数据:

hive> select * from test;
OK
1       3
2       2
3       1

假设,我们要输出每一列的md5值。在目前的hive中是没有这个udf的。

我们看一下Python的代码:

#!/home/tops/bin/python

import sys
import hashlib

for line in sys.stdin:
    line = line.strip()
    arr = line.split()
    md5_arr = []
    for a in arr:
        md5_arr.append(hashlib.md5(a).hexdigest())
    print "\t".join(md5_arr)

在Hive中,使用脚本,首先要将他们加入:

add file /xxxx/test.py

然后,在调用时,使用TRANSFORM语法。

SELECT 
    TRANSFORM (col1, col2) 
    USING './test.py' 
    AS (new1, new2) 
FORM 
    test;

这里,我们使用了AS,指定输出的若干个列,分别对应到哪个列名。如果省略这句,则Hive会将第1个tab前的结果作为key,后面其余作为value。

这里有一个小坑:有时候,我们结合INSERT OVERWRITE使用上述TRANSFORM,而目标表,其分割副可能不是\t。但是请牢记:TRANSFORM的分割符号,传入、传出脚本的,永远是\t。不要考虑外面其他的分割符号!

最后,解释一下MAP、REDUCE。

在有的Hive语句中,大家可能会看到SELECT MAP (…) USING ‘xx.py’这样的语法。

然而,在Hive中,MAP、REDUCE只不过是TRANSFORM的别名,Hive不保证一定会在map/reduce中调用脚本。看看官方文档是怎么说的:

Formally, MAP ... and REDUCE ... are syntactic transformations of SELECT TRANSFORM ( ... ). In other words, they serve as comments or notes to the reader of the query. BEWARE: Use of these keywords may be dangerous as (e.g.) typing "REDUCE" does not force a reduce phase to occur and typing "MAP" does not force a new map phase!

所以、混用map reduce语法关键字,甚至会引起混淆,所以建议大家还是都用TRANSFORM吧。

友情提示:如果脚本不是Python,而是awk、sed等系统内置命令,可以直接使用,而不用add file。

如果表中有MAP,ARRAY等复杂类型,怎么用TRANSFORM生成?

例如:

CREATE TABLE features
(
    id BIGINT,
    norm_features MAP<STRING, FLOAT> 
);

答案是,要在脚本的输出中,对特殊字段按照HDFS文件中的格式输出即可。

例如,以上面的表结构为例,每行输出应为:

1^Ifeature1^C1.0^Bfeature2^C2.0

其中I是tab键,这是TRANSFORM要求的分割符号。B和^C是Hive存储时MAP类型的KV分割符。

另外,在Hive的TRANSFORM语句的时候,要注意AS中加上类型声明:

SELECT TRANSFORM(stuff)
USING 'script'
AS (thing1 INT, thing2 MAP<STRING, FLOAT>)

Hive中的TRANSFORM:自定义Mapper和Reducer完成Map/Reduce

/**
 * Mapper.
 */
public interface Mapper {
  /**
   * Maps a single row into an intermediate rows.
   * 
   * @param record
   *          input record
   * @param output
   *          collect mapped rows.
   * @throws Exception
   *           on error
   */
  void map(String[] record, Output output) throws Exception;
}

可以将一列拆分为多列

使用样例:

public class ExecuteMap {

    private static final String FULL_PATH_CLASS = "com.***.dpop.ods.mr.impl.";

    private static final Map<String, Mapper> mappers = new HashMap<String, Mapper>();

    public static void main(String[] args) throws Exception {
        if (args.length < 1) {
            throw new Exception("Process class must be given");
        }

        new GenericMR().map(System.in, System.out,
                getMapper(args[0], Arrays.copyOfRange(args, 1, args.length)));
    }

    private static Mapper getMapper(String parserClass, String[] args)
            throws ClassNotFoundException {
        if (mappers.containsKey(parserClass)) {
            return mappers.get(parserClass);
        }

        Class[] classes = new Class[args.length];
        for (int i = 0; i < classes.length; ++i) {
            classes[i] = String.class;
        }
        try {
            Mapper mapper = (Mapper) Class.forName(FULL_PATH_CLASS + parserClass).getConstructor(classes).newInstance(args);
            mappers.put(parserClass, mapper);
            return mapper;
        } catch (ClassNotFoundException e) {
            throw new ClassNotFoundException("Unknown MapperClass:" + parserClass, e);
        } catch (Exception e) {
            throw new  ClassNotFoundException("Error Constructing processor", e);
        }

    }
}
MR_USING=" USING 'java -Xmx512m -Xms512m -cp ods-mr-1.0.jar:hive-contrib-2.3.33.jar com.***.dpop.ods.mr.api.ExecuteMap "

COMMAND="FROM dw_rtb.event_fact_adx_auction "
COMMAND="${COMMAND} INSERT overwrite TABLE dw_rtb.event_fact_mid_adx_auction_ad PARTITION(yymmdd=${CURRENT_DATE}) SELECT transform(search_id, print_time, pthread_id, ad_s) ${MR_USING} EventFactMidAdxAuctionAdMapper' as search_id, print_time, pthread_id, ad_s, ssp_id WHERE $INSERT_PARTITION and original = 'exinternal' "

Hive Python Streaming的原理及写法

http://www.tuicool.com/articles/vmumUjA

上一篇下一篇

猜你喜欢

热点阅读