转自:http://db3.iteye.com/blog/1072778
ObjectInspector是Hive中一个咋一看比较令人困惑的概念,当初读Hive源代码时,花了很长时间才理解。 当读懂之后,发现ObjectInspector作用相当大,它解耦了数据使用和数据格式,从而提高了代码的复用程度。 简单的说,ObjectInspector接口使得Hive可以不拘泥于一种特定数据格式,使得数据流 1)在输入端和输出端切换不同的输入/输出格式 2)在不同的Operator上使用不同的数据格式。
这是ObjectInspector interface
public interface ObjectInspector extends Cloneable {
public static enum Category {
PRIMITIVE, LIST, MAP, STRUCT, UNION
};
String getTypeName();
Category getCategory();
}
这个interface提供了最一般的方法 getTypeName 和 getCategory。 我们再来看它的子抽象类和interface:
StructObjectInspector
MapObjectInspector
ListObjectInspector
PrimitiveObjectInspector
UnionObjectInspector
其中,PrimitiveObjectInspector用来完成对基本数据类型的解析,而StructObjectInspector用了完成对一行数据的解析,它本身有一组ObjectInspector组成。 由于Hive支持Nested Data Structure,所以,在StructObjectInspector中又可以(一层或多层的)嵌套任意的ObjectInspector。 Struct, Map, List, Union是Hive支持的4种集合数据类型,比如某一列的数据可以被声明为Struct类型,这样解析这一列的StructObjectInspector中就会嵌套了另一个StructObjectInspector。
现在我们可以从一个小例子看看ObjectInspector是如何工作的,这是一个Hive SerDe的测试用例代码:
/**
* Test the LazySimpleSerDe class.
*/
public void testLazySimpleSerDe() throws Throwable {
try {
// Create the SerDe
LazySimpleSerDe serDe = new LazySimpleSerDe();
Configuration conf = new Configuration();
Properties tbl = createProperties();
//用Properties初始化serDe
serDe.initialize(conf, tbl);
// Data
Text t = new Text("123\t456\t789\t1000\t5.3\thive and hadoop\t1.\tNULL");
String s = "123\t456\t789\t1000\t5.3\thive and hadoop\tNULL\tNULL";
Object[] expectedFieldsData = {new ByteWritable((byte) 123),
new ShortWritable((short) 456), new IntWritable(789),
new LongWritable(1000), new DoubleWritable(5.3),
new Text("hive and hadoop"), null, null};
// Test
deserializeAndSerialize(serDe, t, s, expectedFieldsData);
} catch (Throwable e) {
e.printStackTrace();
throw e;
}
}
private void deserializeAndSerialize(LazySimpleSerDe serDe, Text t, String s,
Object[] expectedFieldsData) throws SerDeException {
// Get the row ObjectInspector
StructObjectInspector oi = (StructObjectInspector) serDe
.getObjectInspector();
// 获取列信息
List<? extends StructField> fieldRefs = oi.getAllStructFieldRefs();
assertEquals(8, fieldRefs.size());
// Deserialize
Object row = serDe.deserialize(t);
for (int i = 0; i < fieldRefs.size(); i++) {
Object fieldData = oi.getStructFieldData(row, fieldRefs.get(i));
if (fieldData != null) {
fieldData = ((LazyPrimitive) fieldData).getWritableObject();
}
assertEquals("Field " + i, expectedFieldsData[i], fieldData);
}
// Serialize
assertEquals(Text.class, serDe.getSerializedClass());
Text serializedText = (Text) serDe.serialize(row, oi);
assertEquals("Serialized data", s, serializedText.toString());
}
//创建schema,保存在Properties中
private Properties createProperties() {
Properties tbl = new Properties();
// Set the configuration parameters
tbl.setProperty(Constants.SERIALIZATION_FORMAT, "9");
tbl.setProperty("columns",
"abyte,ashort,aint,along,adouble,astring,anullint,anullstring");
tbl.setProperty("columns.types",
"tinyint:smallint:int:bigint:double:string:int:string");
tbl.setProperty(Constants.SERIALIZATION_NULL_FORMAT, "NULL");
return tbl;
}
从这个例子中,不难出,Hive将对行中列的读取和行的存储方式解耦和了,只有ObjectInspector清楚行和行中的列是怎样存取的,但使用者并不知道存储的细节。 对于数据的使用者来说,只需要行的Object和相应的ObjectInspector,就能读取出每一列的对象。
这段代码再清晰不过了,ObjectInspector oi控制了对列的Access
for (int i = 0; i < fieldRefs.size(); i++) {
Object fieldData = oi.getStructFieldData(row, fieldRefs.get(i));
if (fieldData != null) {
fieldData = ((LazyPrimitive) fieldData).getWritableObject();
}
assertEquals("Field " + i, expectedFieldsData[i], fieldData);
}
这段代码的作用是把一行deserialize,然后再serialize
Object row = serDe.deserialize(t);
Text serializedText = (Text) serDe.serialize(row, oi);
由此不难看出,只要有了不同的SerDe对象,可以很容易的将一条数据deserialize,然后再serialize成不同的格式,从而非常方便的实现数据格式的切换。
理解了上面的例子,就不难理解为什么所有的Hive ExprNodeEvaluator 和 UDF,UDAF, UDTF 都需要 (Object, ObjectInspector) pair了。 数据存储细节和使用的分离,使得Hive不需要针对不同的数据格式对同一个UDF, UDAF 或UDTF实现不同的版本,这些函数看到的只是WritableObject!
下面是表达式evaluator的interface:
/**
* ExprNodeEvaluator.
*
*/
public abstract class ExprNodeEvaluator {
/**
* Initialize should be called once and only once. Return the ObjectInspector
* for the return value, given the rowInspector.
*/
public abstract ObjectInspector initialize(ObjectInspector rowInspector) throws HiveException;
/**
* Evaluate the expression given the row. This method should use the
* rowInspector passed in from initialize to inspect the row object. The
* return value will be inspected by the return value of initialize.
*/
public abstract Object evaluate(Object row) throws HiveException;
}
initialize中需要初始化ObjectInspector,返回输出数据的ObjectInspector(它负责解析evaluate method返回的对象);而每次evaluate call传进来一条Object数据,它的解析由ObjectInspector负责。
接下来是GenericUDF抽象类:
public abstract class GenericUDF {
/**
* A Defered Object allows us to do lazy-evaluation and short-circuiting.
* GenericUDF use DeferedObject to pass arguments.
*/
public static interface DeferredObject {
Object get() throws HiveException;
};
/**
* The constructor.
*/
public GenericUDF() {
}
/**
* Initialize this GenericUDF. This will be called once and only once per
* GenericUDF instance.
*
* @param arguments
* The ObjectInspector for the arguments
* @throws UDFArgumentException
* Thrown when arguments have wrong types, wrong length, etc.
* @return The ObjectInspector for the return value
*/
public abstract ObjectInspector initialize(ObjectInspector[] arguments)
throws UDFArgumentException;
/**
* Evaluate the GenericUDF with the arguments.
*
* @param arguments
* The arguments as DeferedObject, use DeferedObject.get() to get the
* actual argument Object. The Objects can be inspected by the
* ObjectInspectors passed in the initialize call.
* @return The
*/
public abstract Object evaluate(DeferredObject[] arguments)
throws HiveException;
/**
* Get the String to be displayed in explain.
*/
public abstract String getDisplayString(String[] children);
}
它的机制与evaluator非常类似,初始化中敲定ObjectInspector数组,它们负责解析输入,返回output数据(即evaluator method返回的Object)的ObjectInspector;每次evaluate call传进一个Object数组,返回一条数据。
Hive支持LazySimple, LazyBinary,Thrift等不同的数据格式,同一个查询计划中,可以在operator上切换数据流的格式。比较常见的是在Mapper端使用LazySimpleSerDe,Mapper输出的数据使用LazyBinarySerDe,因为binary格式比较节省空间,从而减少repartition时的网络传输。 如果你想看查询计划的每一步到底使用了哪一种SerDe格式,只要用"Explain Extended"就可以查清楚了。