HBase的WAL机制

简述

WAL(Write-Ahead-Log)预写日志是Hbase的RegionServer在处理数据插入和删除的过程中用来记录操作内容的一种日志。在每次Put、Delete等一条记录时，首先将其数据写入到RegionServer对应的HLog文件的过程。 

客户端往RegionServer端提交数据的时候，会先写WAL日志，只有当WAL日志写成功以后，客户端才会被告诉提交数据成功，如果写WAL失败会告知客户端提交失败，换句话说这其实是一个数据落地的过程。 
在一个RegionServer上的所有的Region都共享一个HLog，一次数据的提交是先写WAL，写入成功后，再写memstore。当memstore值到达一定是，就会形成一个个StoreFile（理解为HFile格式的封装，本质上还是以HFile的形式存储的）。

相关类

HLog类

 
RegionServer内WAL文件与Region的关系图

WAL的实现类是HLog，当一个Region被初始化的时候，一个HLog的实例会作为构造函数的参数传进去。 
当Region在处理Put、Delete等更新操作时，可以直接使用该共享的HLog的append方法来落地数据。 

Put、Delete在客户端上可以通过setWriteToWAL(false)方法来关闭该操作的日志，这么做虽然可以提升入库速度，但最好别这么做，因为有数据丢失的风险存在。

package org.apache.hadoop.hbase.regionserver.wal;
public interface HLog {
    interface Reader {
    /**
     * @param fs File system.
     * @param path Path.
     * @param c Configuration.
     * @param s Input stream that may have been pre-opened by the caller; may be null.
     */
    void init(FileSystem fs, Path path, Configuration c, FSDataInputStream s) throws IOException;

    void close() throws IOException;

    Entry next() throws IOException;

    Entry next(Entry reuse) throws IOException;

    void seek(long pos) throws IOException;

    long getPosition() throws IOException;
    void reset() throws IOException;

    /**
     * @return the WALTrailer of the current HLog. It may be null in case of legacy or corrupt WAL
     *         files.
     */
    // TODO: What we need a trailer on WAL for?
    WALTrailer getWALTrailer();
  }

  interface Writer {
    void init(FileSystem fs, Path path, Configuration c, boolean overwritable) throws IOException;

    void close() throws IOException;

    void sync() throws IOException;

    void append(Entry entry) throws IOException;

    long getLength() throws IOException;

    /**
     * Sets HLog's WALTrailer. This trailer is appended at the end of WAL on closing.
     * @param walTrailer trailer to append to WAL.
     */
    void setWALTrailer(WALTrailer walTrailer);
  }

  /**
   * Utility class that lets us keep track of the edit with it's key.
   * Only used when splitting logs.
   */
  // TODO: Remove this Writable.
  class Entry implements Writable {

  }
}

HLogKey类

1、当前的WAL使用的是Hadoop的SequenceFile格式，其key是HLogKey实例。HLogKey中记录了写入数据的归属信息，除了table和region名字外，同时还包括sequence number和timestamp。 
timestamp是”写入时间”,sequence number的起始值为0，或者是最近一次存入文件系统中Sequence number。 
2、HLog Sequence File的Value是HBase的KeyValue对象，即对应HFile中的KeyValue。

package org.apache.hadoop.hbase.regionserver.wal;
/**
 * A Key for an entry in the change log.
 *
 * The log intermingles edits to many tables and rows, so each log entry
 * identifies the appropriate table and row.  Within a table and row, they're
 * also sorted.
 *
 * 
Some Transactional edits (START, COMMIT, ABORT) will not have an
 * associated row.
 */
// TODO: Key and WALEdit are never used separately, or in one-to-many relation, for practical
//       purposes. They need to be merged into HLogEntry.
public class HLogKey implements WritableComparable {
}

WALEdit类

客户端发送的每个修改都会封装成WALEdit类操作，一个WALEdit类包含了多个更新操作，可以说一个WALEdit就是一个原子操作 
包含了若干个操作的集合。

package org.apache.hadoop.hbase.regionserver.wal;
/**
 * WALEdit: Used in HBase's transaction log (WAL) to represent
 * the collection of edits (KeyValue objects) corresponding to a
 * single transaction. The class implements "Writable" interface
 * for serializing/deserializing a set of KeyValue items.
 *
 * Previously, if a transaction contains 3 edits to c1, c2, c3 for a row R,
 * the HLog would have three log entries as follows:
 *
 *    :
 *    :
 *    :
 *
 * This presents problems because row level atomicity of transactions
 * was not guaranteed. If we crash after few of the above appends make
 * it, then recovery will restore a partial transaction.
 *
 * In the new world, all the edits for a given transaction are written
 * out as a single record, for example:
 *
 *   :
 *
 * where, the WALEdit is serialized as:
 *   <-1, # of edits, , , ... >
 * For example:
 *   <-1, 3, , , >
 *
 * The -1 marker is just a special way of being backward compatible with
 * an old HLog which would have contained a single .
 *
 * The deserializer for WALEdit backward compatibly detects if the record
 * is an old style KeyValue or the new style WALEdit.
 */
public class WALEdit implements Writable, HeapSize {
    private final ArrayList kvs = new ArrayList(1);
}

LogSyncer类

Table在创建的时候，有一个参数可以设置，是否每次写Log日志都需要往集群里的其他机器同步一次，默认是每次都同步，同步的开销是比较大的，但不及时同步又可能因为机器宕而丢日志。

同步的操作现在是通过Pipeline的方式来实现的，Pipeline是指datanode接收数据后，再传给另外一台datanode，是一种串行的方式； 
n-Way Writes是指多datanode同时接收数据，最慢的一台结束就是整个结束。 
差别在于一个延迟大，一个并发高，hdfs现在正在开发中，以便可以选择是按Pipeline还是n-Way Writes来实现写操作。

Table如果设置每次不同步，则写操作会被RegionServe缓存，并启动一个LogSyncer线程来定时同步日志，定时时间默认是1秒，也可由hbase.regionserver.optionallogflushinterval设置。

LogRoller类

日志写入的大小是有限制的。LogRoller类会作为一个后台线程运行，在特定的时间间隔内滚动日志。通过hbase.regionserver.logroll.period属性控制，默认1小时。

HBase容错处理

 
数据量比较大的时候，启动会比较耗时，大部分都是在做split log操作。 
WAL记载了每一个RegionServer所对应的HLog。

RegionServer1或者RegionServer1上的某一个region挂掉了，都会迁移到其他机器上去处理。重新操作，进行恢复。