Pig 实现关键词匹配

1. 问题描述

收集日志avro数据中有两个Map字段appInstall、appUse分别表示已安装的app、正在使用的app，且key值为app的名称，value值为app使用信息。现在要得到一份匹配上购物类app支付宝|京东|淘宝|天猫的用户名单；MapReduce 解决办法如下：

public static class M extends Mapper<String, Pair, String, Text> {
    Text text = new Text();
    
    @SuppressWarnings("unchecked")
    @Override
    protected void map(String key, Pair value, Context context) throws IOException, InterruptedException {
        Map data = value.fields.data;
        
        String dvc = data.get("dvc").toString();
        Map<String, Object> appInstall = (Map<String, Object>) data.get("appInstall");
        Map<String, Object> appUse = (Map<String, Object>) data.get("appUse");
        
        for(String app: appInstall.keySet()) {
            if(app.matches("支付宝|京东|淘宝|天猫")) {
                text.set(appInstall.keySet().toString());
                context.write(dvc, text);
                return;
            }
        }
        
        for(String app: appUse.keySet()) {
            if(app.matches("支付宝|京东|淘宝|天猫")) {
                text.set(appUse.keySet().toString());
                context.write(dvc, text);
                return;
            }
        }
    }
}

但是，如果要匹配游戏类的app、金融类的app类呢？如果匹配关键词发生了变化呢？显然，我们应该将匹配关键词开放成API，可以自由地匹配正则表达式。这时，pig派上了用场。

2. Bag正则匹配

A = load '/<path>/<to>' using org.apache.pig.piggybank.storage.avro.AvroStorage();
-- A: {key: chararray,value: (fields: (data: map[]))}

B = foreach A generate value.fields.data#'dvc' as dvc, value.fields.data#'appInstall' as ins:map[], value.fields.data#'appUse' as use:map[];
-- B: {dvc: bytearray,ins: map[],use: map[]}

C = foreach B generate dvc, KEYSET(ins) as insk, KEYSET(use) as usek;
-- C: {dvc: bytearray,insk: {(chararray)},usek: {(chararray)}}

在上述代码中，load 数据转换得到bag类型的app-set（insk与usek）；但是，应如何遍历bag中的tuple与正则表达式做匹配呢？答案是UDF。

Apache DataFu Pig 提供了丰富的UDF，其中关于bags的UDF可以参看这里。TupleFromBag 提供根据index从bag提取tuple，支持三个输入参数。依葫芦画瓢，遍历bag匹配正则表达式的UDF如下：

package com.pig.udf.bag;

/**
 * This UDF will return true if one tuple from a bag matches regex.
 * 
 *  There are two input parameter:
 *      1. DataBag
 *      2. Regex String
 */
public class BagMatchRegex extends FilterFunc {

    @Override
    public Boolean exec(Tuple tinput) throws IOException {

        try{
            DataBag samples = (DataBag) tinput.get(0);
            String regex = (String) tinput.get(1);
            for (Tuple tuple : samples) {
                if(((String) tuple.get(0)).matches(regex)){
                    return true;
                }
            }
        }
        catch (Exception e) {
            return false;
        }
        return false;
    }
}

其中，FilterFunc为过滤UDF的基类，继承于EvalFunc<Boolean>，即exec(Tuple tinput)的返回值必为Boolean类型。bag正则匹配的pig脚本如下：

REGISTER ../piglib/udf-0.0.1-SNAPSHOT-jar-with-dependencies.jar
define BagMatchRegex com.pig.udf.bag.BagMatchRegex();
A = load '/user/../current/*.avro' using org.apache.pig.piggybank.storage.avro.AvroStorage();
B = foreach A generate value.fields.data#'dvc' as dvc, value.fields.data#'appInstall' as ins:map[], value.fields.data#'appUse' as use:map[];
C = foreach B generate dvc, KEYSET(ins) as insk, KEYSET(use) as usek;
D = filter C by BagMatchRegex(insk, '支付宝|京东|淘宝|天猫') or BagMatchRegex(usek, '支付宝|京东|淘宝|天猫');

3. 优化

还有没有可以做优化的地方呢？我们先来看看pig中的KEYSET实现：

package org.apache.pig.builtin;

public class KEYSET extends EvalFunc<DataBag> {
    private static final TupleFactory TUPLE_FACTORY = TupleFactory.getInstance();

    @SuppressWarnings("unchecked")
    @Override
    public DataBag exec(Tuple input) throws IOException {
        if(input == null || input.size() == 0) {
            return null;
        }

        Map<String, Object> m = null;
        // Input must be of type Map. This is verified at compile time
        m = (Map<String, Object>)(input.get(0));
        if(m == null) {
            return null;
        }

        DataBag bag = new NonSpillableDataBag(m.size());
        for (String s : m.keySet()) {
            Tuple t = TUPLE_FACTORY.newTuple(s);
            bag.add(t);
        }

        return bag;
    }
    ...
}

需要指出的一点——pig的map数据类型是由Java类Map<String, Object>实现的。从KEYSET源码中可以看出在调用时已经将map遍历了一次，然后在调用BagMatchRegex时又需要将key-set的bag再遍历一次。其实，完全可以只用一次遍历做map-key值的正则匹配：

package com.pig.udf.map;

/**
 * This UDF will return true if map's key matches regex.
 * 
 *  There are two input parameter:
 *      1. Map
 *      2. Regex String
 */
public class KeyMatchRegex extends FilterFunc {
    
    @SuppressWarnings("unchecked")
    @Override
    public Boolean exec(Tuple input) throws IOException
    {
        try{
            Map<String, Object> m = null;
            // Input must be of type Map. This is verified at compile time
            m = (Map<String, Object>)(input.get(0));
            
            String regex = (String) input.get(1);
            for (String key : m.keySet()) {
                if(key.matches(regex)){
                    return true;
                }
            }
        }
        catch (Exception e) {
            return false;
        }
        return false;
    }
}