【Pig源码分析】谈谈Pig的数据模型
1. 数据模型
Schema
Pig Latin表达式操作的是relation,FILTER、FOREACH、GROUP、SPLIT等关系操作符所操作的relation就是bag,bag为tuple的集合,tuple为有序的field列表集合,而field表示数据块(A field is a piece of data),可理解为数据字段。
Schema为数据所遵从的类型格式,包括:field的名称及类型(names and types)。用户常用as语句来自定义schema,或是load函数导入schema,比如:
A = foreach X generate .. as field1:chararray, .. as field2:bag{};
A = load '..' using PigStorage('\t', '-schema');
A = load '..' using org.apache.pig.piggybank.storage.avro.AvroStorage();
若不指定field的类型,则其默认为bytearray。对未知schema进行操作时,有:
- 若join/cogroup/cross多关系操作遇到未知schema,则会将其视为null schema,导致返回结果的schema也为null;
- 若flatten一个empty inner schema的bag(即:bag{})时,则返回结果的schema为null;
- 若union时二者relation的schema不一致,则返回结果的schema为null;
- 若field的schema为null,会将该字段视为bytearray。
为了保证pig脚本运行的有效性,在写UDF时要在outputSchema方法中指定返回结果的schema。
数据类型
Pig的基本数据类型与对应的Java类:
Simple Pig Type | Example | Java Class |
---|---|---|
bytearray | DataByteArray | |
chararray | 'hello world' | String |
int | 10 | Integer |
long | 10L | Long |
float | 10.5F or 1050.0F | Float |
double | Double | |
boolean | true/false | Boolean |
datetime | DateTime | |
bigdecimal | BigDecimal | |
biginteger | BigInteger |
复杂数据类型及其对应的Java类:
Complex Pig Type | Example | Java Class |
---|---|---|
tuple | (19, 'hello') | Tuple |
bag | DataBag | |
map | [open#apache] | Map |
Pig的复杂数据类型可以嵌套表达,比如:tuple中有tuple (a, (b, c, d))
,tuple中有bag (a, {(b,c), (d,e)})
等等。但是一定要遵从数据类型本身的定义,比如:bag中只能是tuple的集合,比如{a, {(b),(c)}}
就是不合法的。
Pig还有一种特殊的数据类型:null,与Java、C中null不一样,其表示不知道的或不存在的数据类型(unknown or non-existent)。比如,在load数据时,如果有的数据行字段不符合定义的schema,则该字段会被置为null。
2. 源码分析
以下源码分析采用的是0.12版本。
Tuple
在KEYSET源码中,创建Tuple对象采用工厂+单例设计模式:
private static final TupleFactory TUPLE_FACTORY = TupleFactory.getInstance();
Tuple t = TUPLE_FACTORY.newTuple(s);
事实上,TupleFactory是个抽象类,实现接口TupleMaker<Tuple>
。在方法TupleFactory.getInstance()
中,默认情况下返回的是BinSedesTupleFactory对象,同时支持加载用户重写的TupleFactory类(pig.data.tuple.factory.name
指定类名、 pig.data.tuple.factory.jar
指定类所在的jar)。BinSedesTupleFactory继承于TupleFactory:
在BinSedesTupleFactory的newTuple方法中,返回的是BinSedesTuple对象。BinSedesTuple类继承于DefaultTuple类,在DefaultTuple类中有List<Object> mFields
字段,这便是存储Tuple数据的地方了,mFields所持有类型为ArrayList<Object>()
;。类图关系:
Bag
创建Bag对象有下面几种方法:
// factory
BagFactory mBagFactory = BagFactory.getInstance();
DataBag output = mBagFactory.newDefaultBag();
// if you know upfront how many tuples you are going to put in this bag.
DataBag bag = new NonSpillableDataBag(m.size());
与TupleFactory一样,BagFactory也是抽象类,也支持用户自定义重写;getInstance方法默认返回的是DefaultBagFactory。DefaultBagFactory有newDefaultBag、newSortedBag、newDistinctBag方法分别创建三类bag:
- default bag中的tuple没有排序,也没有去重;
- sorted bag中的tuple是按序存放,顺序是由tuple default comparator或bag创建时的comparator所定义的;
- distinct bag顾名思义,tuple有去重。
三类bag的构造器如下:
public DefaultDataBag() {
mContents = new ArrayList<Tuple>();
}
public SortedDataBag(Comparator<Tuple> comp) {
mComp = (comp == null) ? new DefaultComparator() : comp;
mContents = new ArrayList<Tuple>();
}
public DistinctDataBag() {
mContents = new HashSet<Tuple>();
}
BagFactory的类图:
DefaultAbstractBag作为三种类型bag的基类,有一个字段mContents用于存放tuple,NonSpillableDataBag直接实现DataBag接口。DataBag的类图:
3. 实战
现有avro日志数据(见前一篇),其字段:
dvc
表示用户手机标识;appUse
与appInstall
同为avro Map类型,其key为app名称(app name),value为Map<String, Object>
,包含了一个表示使用时间的字段timelist
(类型为ArrayList);具体格式如下
'dvc': 'imei_123',
'appUse': {
'app name1': {
...
'timelist': [...]
},
'app name2': {
...
'timelist': [...]
},
...
},
'appInstall': {
'app name1': {
...
'timelist': [...]
},
...
}
现在,想要得到每个用户的app列表及app的打开次数,以格式dvc, {(app)}, {(app, frequency)}
输出,即用户 + app列表 + 使用次数类表。如果用MapRduce做,得分为以下步骤:
- 以(dvc, app)为key值,计算value值为使用次数;
- 以dvc为key值,合并同一用户的不同app,value值为(app, fre);
- 以dvc为key值,计算appinstall的app列表;
- 将步骤2得到的数据与步骤3得到的数据做join,然后输出。
可以看出用MapReduce略显繁复,如何来用pig来实现呢?我们可以对appUse:map[]
编写EVAL UDF,让其返回(app名称, timelist的长度) :
public class AppTimelist extends EvalFunc<DataBag>{
private static final TupleFactory TUPLE_FACTORY = TupleFactory.getInstance();
private static final BagFactory BAG_FACTORY = BagFactory.getInstance();
@SuppressWarnings({ "unchecked" })
@Override
public DataBag exec(Tuple input) throws IOException {
Map<String, Map<String, Object>> m = (Map<String, Map<String, Object>>) input.get(0);
List<Object> result = new ArrayList<Object>();
DataBag output = BAG_FACTORY.newDefaultBag();
if(m == null)
return null;
for(Map.Entry<String, Map<String, Object>> e: m.entrySet()) {
result.clear();
String app = e.getKey();
long size = ((DataBag) e.getValue().get("timelist")).size();
result.add(app);
result.add(size);
output.add(TUPLE_FACTORY.newTuple(result));
}
return output;
}
}
pig将Java的ArrayList转成DataBag的类型,所以要对timelist进行强转操作。
对appInstall:map[]
编写EVAL UDF,返回(appList):
public class DistinctBag extends EvalFunc<DataBag> {
BagFactory mBagFactory = BagFactory.getInstance();
@Override
public DataBag exec(Tuple input) throws IOException {
if(input == null || input.size() == 0) {
return null;
}
DataBag in = (DataBag) input.get(0);
DataBag out = mBagFactory.newDistinctBag();
if(in == null) {
return null;
}
for(Tuple tp: in) {
DataBag applist = (DataBag) tp.get(0);
for(Tuple app: applist)
out.add(app);
}
return out;
}
}
上面提到过,若没有给EVAL UDF指定返回值的schema,则返回结果的schema为null,如此会造成类型的丢失,在后面的操作中容易报NullPointerException。
// AppTimelist.java
@Override
public Schema outputSchema(Schema input) {
try {
Schema tupleSchema = new Schema();
FieldSchema chararrayFieldSchema = new Schema.FieldSchema(null, DataType.CHARARRAY);
FieldSchema longFieldSchema = new Schema.FieldSchema(null, DataType.LONG);
tupleSchema.add(chararrayFieldSchema);
tupleSchema.add(longFieldSchema);
return new Schema(new Schema.FieldSchema(getSchemaName(this
.getClass().getName().toLowerCase(), input), tupleSchema,
DataType.TUPLE));
} catch (Exception e) {
return null;
}
}
// DistinctBag.java
@Override
public Schema outputSchema(Schema input) {
FieldSchema innerFieldSchema = new Schema.FieldSchema(null, DataType.CHARARRAY);
Schema innerSchema = new Schema(innerFieldSchema);
Schema bagSchema = null;
try {
bagSchema = new Schema(new FieldSchema(null, innerSchema, DataType.BAG));
} catch(FrontendException e) {
throw new RuntimeException(e);
}
return bagSchema;
}
统计app列表:
define AvroStorage org.apache.pig.piggybank.storage.avro.AvroStorage;
define DistinctBag com.pig.udf.bag.DistinctBag;
A = load '..' using AvroStorage();
B = foreach A generate value.fields.data#'dvc' as dvc:chararray, value.fields.data#'appInstall' as ins:map[map[]];
C = foreach B generate dvc, KEYSET(ins) as applist;
D = group C by dvc;
-- extract applist from grouped D
E = foreach D {
projected = foreach $1 generate applist;
generate group as dvc, projected as grouped;
}
F = foreach E generate dvc, DistinctBag(grouped) as applist;
store F into '..' using AvroStorage();
统计app使用时长:
define AvroStorage org.apache.pig.piggybank.storage.avro.AvroStorage;
define AppTimelist com.pig.udf.map.AppTimelist;
A = load '..' using AvroStorage();
B = foreach A generate value.fields.data#'dvc' as dvc:chararray, value.fields.data#'appUse' as use:map[map[]];
C = foreach B generate dvc, flatten(AppTimelist(use)) as (app, fre);
D = group C by (dvc, app);
E = foreach D generate flatten(group) as (dvc, app), SUM($1.fre) as fre;
F = group E by dvc;
G = foreach F {
projected = foreach $1 generate app, fre;
generate group as dvc, projected as appfre;
}
store G into '..' using AvroStorage();
二者做join即可得到结果。