Hadoop关于大量小压缩文件的问题和解决方法

philip_kissme

浏览: 16622 次
来自: ...

最近访客更多访客>>

mfcai

duguyiren3476

erichi101

King_xiaozi

博主相关

博客

微博

相册

留言

关于我

文章分类

社区版块

存档分类

hadoop CombineFileInputFormat gzip 小文件

之前一段时间偶尔会收到 hadoop 集群的 nagios 监控报警,具体报警是几个 resource-manager 节点一直负载超过阀值.找了个空闲时间分析了一下 job-history,发现是一个小伙伴的 job在段时间内创建了大量 map-task 导致的,在解决问题后做个笔记备忘

首先分析 job-history 的统计数据

可以发现 map 任务执行的时间很短,但是同时会有大量的 map 任务
与小伙伴沟通后,了解到他的 job 是根据运营侧需求,本周都在应用埋点日志中提取指定按钮的点击计数
应用埋点的日志记录了每个用户的所有访问路径和参数
1) log-agent 通过 logback 将日志记录到本地
2)每小时生成一个 gz 压缩包,并上传至 hdfs 指定目录(根据应用标识+ip 生产目录规则)

根据 hadoop 的 map split 机制我们可以得出如下结论

每个 inputfile 会对应多个 map split(根据 hdfs 的 block zise切分)
每个map split会对应一个 map task
由于每个小时生成的 gz 文件均未超过hdfs block zise(128m)
小伙伴要统计的集群中有三个应用节点,排除凌晨时段没有日志产出的情况,大概一周的日志文件树=24*7*3~=350

解决问题

我们知道 hadoop中是可以利用CombineFileInputFormat来合并大量小文件输入,提高 map 性能的.
但默认实现只提供了CombineSequenceFileInputFormat和CombineTextFileInputFormat,没有压缩文件的支持.
所以这里要实现自定义的CombineFileInputFormat来解决该问题

自定义CompressedCombineFileInputFormat

package ctu.components.amada.hadoop.usertrace;

import org.apache.hadoop.io.Text;
import org.apache.hadoop.mapreduce.InputSplit;
import org.apache.hadoop.mapreduce.RecordReader;
import org.apache.hadoop.mapreduce.TaskAttemptContext;
import org.apache.hadoop.mapreduce.lib.input.CombineFileInputFormat;
import org.apache.hadoop.mapreduce.lib.input.CombineFileRecordReader;
import org.apache.hadoop.mapreduce.lib.input.CombineFileSplit;

import java.io.IOException;

/**
 * Created by pippo on 14/12/23.
 */
public class CompressedCombineFileInputFormat extends CombineFileInputFormat<CompressedRecordKey, Text> {

	public RecordReader<CompressedRecordKey, Text> createRecordReader(InputSplit split, TaskAttemptContext context)
			throws IOException {
		return new CombineFileRecordReader<>((CombineFileSplit) split,
				context,
				CompressedCombineFileRecordReader.class);
	}

}

CompressedCombineFileRecordReader

package ctu.components.amada.hadoop.usertrace;

import org.apache.hadoop.conf.Configuration;
import org.apache.hadoop.fs.FSDataInputStream;
import org.apache.hadoop.fs.FileSystem;
import org.apache.hadoop.fs.Path;
import org.apache.hadoop.io.IOUtils;
import org.apache.hadoop.io.Text;
import org.apache.hadoop.io.compress.CompressionCodec;
import org.apache.hadoop.io.compress.CompressionCodecFactory;
import org.apache.hadoop.mapreduce.InputSplit;
import org.apache.hadoop.mapreduce.RecordReader;
import org.apache.hadoop.mapreduce.TaskAttemptContext;
import org.apache.hadoop.mapreduce.lib.input.CombineFileSplit;
import org.apache.hadoop.util.LineReader;

import java.io.IOException;
import java.io.InputStream;
import java.io.OutputStream;

/**
 * Created by pippo on 14/12/23.
 */
public class CompressedCombineFileRecordReader extends RecordReader<CompressedRecordKey, Text> {

	private long offset;
	private long end;
	private long pos;
	private CompressedRecordKey key;
	private Text value = new Text();
	private CompressTrunk trunk;
	private LineReader reader;

	public CompressedCombineFileRecordReader(CombineFileSplit split, TaskAttemptContext context, Integer index)
			throws IOException {

		/*多个压缩文件合并为一个ombine file, 那么实际的压缩文件就是file中的一个trunk*/
		this.trunk = new CompressTrunk(context.getConfiguration(), split.getPath(index));
		/*trunk在combine 中的起始位置*/
		this.offset = split.getOffset(index);
		/* trunk在combine file中的结束位置*/
		this.end = offset + (trunk.isCompress() ? trunk.getFileLength() : split.getLength(index));

		boolean skipFirstLine = false;
		FSDataInputStream in = trunk.open();

		if (offset != 0) {
			skipFirstLine = true;
			--offset;
			in.seek(offset);
		}

		reader = new LineReader(trunk.open());

		// skip first line and re-establish "offset".
		if (skipFirstLine) {
			offset += reader.readLine(new Text(), 0, (int) Math.min((long) Integer.MAX_VALUE, end - offset));
		}
		this.pos = offset;
	}

	@Override
	public void initialize(InputSplit split, TaskAttemptContext context) throws IOException, InterruptedException {

	}

	@Override
	public void close() throws IOException {
		trunk.close();
	}

	@Override
	public float getProgress() throws IOException {
		if (offset == end) {
			return 0.0f;
		} else {
			return Math.min(1.0f, (pos - offset) / (float) (end - offset));
		}
	}

	@Override
	public boolean nextKeyValue() throws IOException {
		initKey();
		initValue();

		//指定当前记录的读取起始位置
		key.offset = pos;

		int readed = 0;

		//读取一条记录
		if (pos < end) {
			readed = reader.readLine(value);
			pos += readed;
		}

		//如果没有读到任何记录,说明当前 trunk 已经没有更多记录了
		if (readed == 0) {
			key = null;
			value = null;
			return false;
		} else {
			return true;
		}
	}

	private void initKey() {
		if (key == null) {
			key = new CompressedRecordKey();
			key.fileName = trunk.getFileName();
		}
	}

	private void initValue() {
		if (value == null) {
			value = new Text();
		}
	}

	@Override
	public CompressedRecordKey getCurrentKey() throws IOException, InterruptedException {
		return key;
	}

	@Override
	public Text getCurrentValue() throws IOException, InterruptedException {
		return value;
	}

	public static class CompressTrunk {

		public CompressTrunk(Configuration configuration, Path compressFile) throws IOException {
			this.configuration = configuration;
			this.compressFile = compressFile;
			this.fs = compressFile.getFileSystem(configuration);
			this.factory = new CompressionCodecFactory(configuration);
			this.codec = factory.getCodec(compressFile);

			prepareReadableFile();
		}

		/*将trunk解压缩到一个临时目录,并提供inputStream供读取*/
		protected void prepareReadableFile() throws IOException {
			if (!isCompress()) {
				readableFile = compressFile;
				return;
			}

			String _readFile = CompressionCodecFactory.removeSuffix(compressFile.toString(),
					codec.getDefaultExtension());
			readableFile = new Path(_readFile);

			InputStream in = null;
			OutputStream out = null;
			try {
				in = codec.createInputStream(fs.open(compressFile));
				out = fs.create(readableFile);
				IOUtils.copyBytes(in, out, configuration);
			} finally {
				IOUtils.closeStream(in);
				IOUtils.closeStream(out);
			}
		}

		private Configuration configuration;
		//源文件
		private Path compressFile;
		//解压后的文件
		private Path readableFile;
		private FileSystem fs;
		private CompressionCodecFactory factory;
		private CompressionCodec codec;

		public boolean isCompress() {
			return codec != null;
		}

		public String getFileName() {
			return readableFile.getName();
		}

		public long getFileLength() throws IOException {
			return fs.getFileStatus(readableFile).getLen();
		}

		private FSDataInputStream in;

		public FSDataInputStream open() throws IOException {
			if (in == null) {
				in = fs.open(readableFile);
			}
			return in;
		}

		//处理完毕后删除临时文件
		public void close() throws IOException {
			if (in != null) {
				in.close();
			}

			if (isCompress()) {
				fs.delete(readableFile, false);
			}
		}
	}

}

CompressedRecordKey

package ctu.components.amada.hadoop.usertrace;

import org.apache.hadoop.io.WritableComparable;

import java.io.DataInput;
import java.io.DataOutput;
import java.io.IOException;

/**
 * Created by pippo on 14/12/23.
 */
public class CompressedRecordKey implements WritableComparable {

	//记录所在的文件
	public String fileName;
	//记录在文件中所在的位置
	public long offset;

	public CompressedRecordKey() {
		super();
	}

	@Override
	public void readFields(DataInput in) throws IOException {
		this.offset = in.readLong();
		this.fileName = in.readUTF();
	}

	@Override
	public void write(DataOutput out) throws IOException {
		out.writeLong(offset);
		out.writeUTF(fileName);
	}

	@Override
	public int compareTo(Object o) {
		CompressedRecordKey that = (CompressedRecordKey) o;

		int f = this.fileName.compareTo(that.fileName);
		if (f == 0) {
			return (int) Math.signum((double) (this.offset - that.offset));
		}
		return f;
	}

	@Override
	public boolean equals(Object obj) {
		if (obj instanceof CompressedRecordKey) {
			return this.compareTo(obj) == 0;
		}
		return false;
	}

	@Override
	public int hashCode() {

		final int hashPrime = 47;
		int hash = 13;
		hash = hashPrime * hash + (this.fileName != null ? this.fileName.hashCode() : 0);
		hash = hashPrime * hash + (int) (this.offset ^ (this.offset >>> 16));

		return hash;
	}

	@Override
	public String toString() {
		return this.fileName + "-" + this.offset;
	}

}

JOB 配置

	private void buildMapper(Job job) {
		job.setInputFormatClass(CompressedCombineFileInputFormat.class);
		job.setMapperClass(LogMapper.class);
		job.setMapOutputKeyClass(Text.class);
		job.setMapOutputValueClass(UserTrace.class);
		job.setCombinerClass(LogCombiner.class);
	}