本文转载自：http://hi.baidu.com/lzpsky/blog/item/99d58738b08a68e7b311c70d.html
hadoop会对原始输入文件进行文件切割，然后把每个split传入mapper程序中进行处理，FileInputFormat是所有以文件作 为数据源的InputFormat实现的基类，FileInputFormat保存作为job输入的所有文件，并实现了对输入文件计算splits的方 法。至于获得记录的方法是有不同的子类进行实现的。
那么，FileInputFormat是怎样将他们划分成splits的呢？FileInputFormat只划分比HDFS block大的文件，所以如果一个文件的大小比block小，将不会被划分，这也是Hadoop处理大文件的效率要比处理很多小文件的效率高的原因。
hadoop默认的InputFormat是TextInputFormat，重写了FileInputFormat中的createRecordReader和isSplitable方法。该类使用的reader是LineRecordReader，即以回车键(CR = 13)或换行符(LF = 10)为行分隔符。
但大多数情况下，回车键或换行符作为输入文件的行分隔符并不能满足我们的需求，通常用户很有可能会输入回车键、换行符，所以通常我们会定义不可见字符(即用户无法输入的字符)为行分隔符，这种情况下，就需要新写一个InputFormat。
又或者，一条记录的分隔符不是字符，而是字符串，这种情况相对麻烦；还有一种情况，输入文件的主键key已经是排好序的了，需要hadoop做的只是把相 同的key作为一个数据块进行逻辑处理，这种情况更麻烦，相当于免去了mapper的过程，直接进去reduce，那么InputFormat的逻辑就相 对较为复杂了，但并不是不能实现。
1、改变一条记录的分隔符，不用默认的回车或换行符作为记录分隔符，甚至可以采用字符串作为记录分隔符。
1）自定义一个InputFormat，继承FileInputFormat，重写createRecordReader方法，如果不需要分片或者需要改变分片的方式，则重写isSplitable方法，具体代码如下：
public class FileInputFormatB extends FileInputFormat
@Override
public RecordReader
return new SearchRecordReader("");
}
@Override
protected boolean isSplitable(FileSystem fs, Path filename) {
// 输入文件不分片
return false;
}
}
2）关键在于定义一个新的SearchRecordReader继承RecordReader，支持自定义的行分隔符，即一条记录的分隔符。标红的地方为与hadoop默认的LineRecordReader不同的地方。
public class IsearchRecordReader extends RecordReader
private static final Log LOG = LogFactory.getLog(IsearchRecordReader.class);
private CompressionCodecFactory compressionCodecs = null;
private long start;
private long pos;
private long end;
private LineReader in;
private int maxLineLength;
private LongWritable key = null;
private Text value = null;
//行分隔符，即一条记录的分隔符
private byte[] separator = {''};
private int sepLength = 1;
public IsearchRecordReader(){
}
public IsearchRecordReader(String seps){
this.separator = seps.getBytes();
sepLength = separator.length;
}
public void initialize(InputSplit genericSplit, TaskAttemptContext context) throws IOException {
FileSplit split = (FileSplit) genericSplit;
Configuration job = context.getConfiguration();
this.maxLineLength = job.getInt("mapred.linerecordreader.maxlength", Integer.MAX_VALUE);
this.start = split.getStart();
this.end = (this.start + split.getLength());
Path file = split.getPath();
this.compressionCodecs = new CompressionCodecFactory(job);
CompressionCodec codec = this.compressionCodecs.getCodec(file);
// open the file and seek to the start of the split
FileSystem fs = file.getFileSystem(job);
FSDataInputStream fileIn = fs.open(split.getPath());
boolean skipFirstLine = false;
if (codec != null) {
this.in = new LineReader(codec.createInputStream(fileIn), job);
this.end = Long.MAX_VALUE;
} else {
if (this.start != 0L) {
skipFirstLine = true;
this.start -= sepLength;
fileIn.seek(this.start);
}
this.in = new LineReader(fileIn, job);
}
if (skipFirstLine) { // skip first line and re-establish "start".
int newSize = in.readLine(new Text(), 0, (int) Math.min( (long) Integer.MAX_VALUE, end - start));
if(newSize > 0){
start += newSize;
}
}
this.pos = this.start;
}
public boolean nextKeyValue() throws IOException {
if (this.key == null) {
this.key = new LongWritable();
}
this.key.set(this.pos);
if (this.value == null) {
this.value = new Text();
}
int newSize = 0;
while (this.pos < this.end) {
newSize = this.in.readLine(this.value, this.maxLineLength, Math.max(
(int) Math.min(Integer.MAX_VALUE, this.end - this.pos), this.maxLineLength));
if (newSize == 0) {
break;
}
this.pos += newSize;
if (newSize < this.maxLineLength) {
break;
}
LOG.info("Skipped line of size " + newSize + " at pos " + (this.pos - newSize));
}
if (newSize == 0) {
//读下一个buffer
this.key = null;
this.value = null;
return false;
}
//读同一个buffer的下一个记录
return true;
}
public LongWritable getCurrentKey() {
return this.key;
}
public Text getCurrentValue() {
return this.value;
}
public float getProgress() {
if (this.start == this.end) {
return 0.0F;
}
return Math.min(1.0F, (float) (this.pos - this.start) / (float) (this.end - this.start));
}
public synchronized void close() throws IOException {
if (this.in != null)
this.in.close();
}
}
3）重写SearchRecordReader需要的LineReader，可作为SearchRecordReader内部类。特别需要注意的地方就 是，读取文件的方式是按指定大小的buffer来读，必定就会遇到一条完整的记录被切成两半，甚至如果分隔符大于1个字符时分隔符也会被切成两半的情况， 这种情况一定要加以拼接处理。
public class LineReader {
//回车键(hadoop默认)
//private static final byte CR = 13;
//换行符(hadoop默认)
//private static final byte LF = 10;
//按buffer进行文件读取
private static final int DEFAULT_BUFFER_SIZE = 32 * 1024 * 1024;
private int bufferSize = DEFAULT_BUFFER_SIZE;
private InputStream in;
private byte[] buffer;
private int bufferLength = 0;
private int bufferPosn = 0;
LineReader(InputStream in, int bufferSize) {
this.bufferLength = 0;
this.bufferPosn = 0;
this.in = in;
this.bufferSize = bufferSize;
this.buffer = new byte[this.bufferSize];
}
public LineReader(InputStream in, Configuration conf) throws IOException {
this(in, conf.getInt("io.file.buffer.size", DEFAULT_BUFFER_SIZE));
}
public void close() throws IOException {
in.close();
}
public int readLine(Text str, int maxLineLength) throws IOException {
return readLine(str, maxLineLength, Integer.MAX_VALUE);
}
public int readLine(Text str) throws IOException {
return readLine(str, Integer.MAX_VALUE, Integer.MAX_VALUE);
}
//以下是需要改写的部分_start，核心代码
public int readLine(Text str, int maxLineLength, int maxBytesToConsume) throws IOException{
str.clear();
Text record = new Text();
int txtLength = 0;
long bytesConsumed = 0L;
boolean newline = false;
int sepPosn = 0;
do {
//已经读到buffer的末尾了，读下一个buffer
if (this.bufferPosn >= this.bufferLength) {
bufferPosn = 0;
bufferLength = in.read(buffer);
//读到文件末尾了，则跳出，进行下一个文件的读取
if (bufferLength <= 0) {
break;
}
}
int startPosn = this.bufferPosn;
for (; bufferPosn < bufferLength; bufferPosn ++) {
//处理上一个buffer的尾巴被切成了两半的分隔符(如果分隔符中重复字符过多在这里会有问题)
if(sepPosn > 0 && buffer[bufferPosn] != separator[sepPosn]){
sepPosn = 0;
}
//遇到行分隔符的第一个字符
if (buffer[bufferPosn] == separator[sepPosn]) {
bufferPosn ++;
int i = 0;
//判断接下来的字符是否也是行分隔符中的字符
for(++ sepPosn; sepPosn < sepLength; i ++, sepPosn ++){
//buffer的最后刚好是分隔符，且分隔符被不幸地切成了两半
if(bufferPosn + i >= bufferLength){
bufferPosn += i - 1;
break;
}
//一旦其中有一个字符不相同，就判定为不是分隔符
if(this.buffer[this.bufferPosn + i] != separator[sepPosn]){
sepPosn = 0;
break;
}
}
//的确遇到了行分隔符
if(sepPosn == sepLength){
bufferPosn += i;
newline = true;
sepPosn = 0;
break;
}
}
}
int readLength = this.bufferPosn - startPosn;
bytesConsumed += readLength;
//行分隔符不放入块中
//int appendLength = readLength - newlineLength;
if (readLength > maxLineLength - txtLength) {
readLength = maxLineLength - txtLength;
}
if (readLength > 0) {
record.append(this.buffer, startPosn, readLength);
txtLength += readLength;
//去掉记录的分隔符
if(newline){
str.set(record.getBytes(), 0, record.getLength() - sepLength);
}
}
} while (!newline && (bytesConsumed < maxBytesToConsume));
if (bytesConsumed > (long)Integer.MAX_VALUE) {
throw new IOException("Too many bytes before newline: " + bytesConsumed);
}
return (int) bytesConsumed;
}
//以下是需要改写的部分_end
//以下是hadoop-core中LineReader的源码_start
public int readLine(Text str, int maxLineLength, int maxBytesToConsume) throws IOException{
str.clear();
int txtLength = 0;
int newlineLength = 0;
boolean prevCharCR = false;
long bytesConsumed = 0L;
do {
int startPosn = this.bufferPosn;
if (this.bufferPosn >= this.bufferLength) {
startPosn = this.bufferPosn = 0;
if (prevCharCR) bytesConsumed ++;
this.bufferLength = this.in.read(this.buffer);
if (this.bufferLength <= 0) break;
}
for (; this.bufferPosn < this.bufferLength; this.bufferPosn ++) {
if (this.buffer[this.bufferPosn] == LF) {
newlineLength = (prevCharCR) ? 2 : 1;
this.bufferPosn ++;
break;
}
if (prevCharCR) {
newlineLength = 1;
break;
}
prevCharCR = this.buffer[this.bufferPosn] == CR;
}
int readLength = this.bufferPosn - startPosn;
if ((prevCharCR) && (newlineLength == 0))
--readLength;
bytesConsumed += readLength;
int appendLength = readLength - newlineLength;
if (appendLength > maxLineLength - txtLength) {
appendLength = maxLineLength - txtLength;
}
if (appendLength > 0) {
str.append(this.buffer, startPosn, appendLength);
txtLength += appendLength; }
}
while ((newlineLength == 0) && (bytesConsumed < maxBytesToConsume));
if (bytesConsumed > (long)Integer.MAX_VALUE) throw new IOException("Too many bytes before newline: " + bytesConsumed);
return (int)bytesConsumed;
}
//以下是hadoop-core中LineReader的源码_end
}
2、已经按主键key排好序了，并保证相同主键key一定是在一起的，假设每条记录的第一个字段为主键，那么如 果沿用上面的LineReader，需要在核心方法readLine中对前后两条记录的id进行equals判断，如果不同才进行split，如果相同继 续下一条记录的判断。代码就不再贴了，但需要注意的地方，依旧是前后两个buffer进行交接的时候，非常有可能一条记录被切成了两半，一半在前一个buffer中，一半在后一个buffer中。
这种方式的好处在于少去了reduce操作，会大大地提高效率，其实mapper的过程相当的快，费时的通常是reduce。