java

关注公众号 jb51net

关闭
首页 > 软件编程 > java > Lucene 索引删除策略

Lucene 索引删除策略源码解析

作者：沧叔解码

这篇文章主要为大家介绍了Lucene 索引删除策略源码解析，有需要的朋友可以借鉴参考下，希望能够有所帮助，祝大家多多进步，早日升职加薪

Lucene

从今天开始，我们要开始介绍Lucene中索引构建的流程。因为索引构建的逻辑涉及到的东西非常多，如果从构建入口IndexWriter来开始介绍，是很难说清楚的。所以接下来按化零为整的方式 ，从构建相关的各个组件开始介绍，尽量每一篇文章都是可以独立阅读，依赖的前置知识都是我已经介绍的内容。不管就算如此，还是会有部分内容可能需要结合整体流程才能明白，对于这部分的内容，大家可以先留个印象，以后介绍相关联的内容时，我会再重新指出。

今天我们一起来看看索引文件删除相关的。

IndexCommit

Lucene中，需要持久化的索引信息都要进行commit操作，然后会生成一个segments_N的索引文件记录此次commit相关的索引信息。

一次commit生成segments_N之后，就对应了一个IndexCommit，IndexCommit只是一个接口，它定义了可以从IndexCommit中获取哪些信息：

public abstract class IndexCommit implements Comparable<IndexCommit> {
  // commit对应的segments_N
  public abstract String getSegmentsFileName();
  // commit关联的所有的索引文件
  public abstract Collection<String> getFileNames() throws IOException;
  // 索引所在的Directory
  public abstract Directory getDirectory();
  // 删除commit，后面会看到，删除其实减少commit关联的索引文件的引用计数
  public abstract void delete();
  // commit是否被删除了
  public abstract boolean isDeleted();
  // commit关联了几个segment
  public abstract int getSegmentCount();
  // segments_N文件中的N
  public abstract long getGeneration();
  // commit可以记录一些用户自定义的信息
  public abstract Map<String, String> getUserData() throws IOException;
  // 用来读取commit对应的索引数据
  StandardDirectoryReader getReader() {
    return null;
  }
}

IndexCommit有三个实现类：

这个三个实现类都有对应的使用场景，在用到的时候我会再详细介绍，本文中会涉及到SnapshotCommitPoint，后面会详细介绍它。

IndexDeletionPolicy

在索引的生命周期中，可以有多次的commit操作，因此也会生成多个segments_N文件，对于这些文件是否要保留还是删除，lucene中是通过IndexDeletionPolicy来管理的。我们先来看下IndexDeletionPolicy的接口定义：

public abstract class IndexDeletionPolicy {
  protected IndexDeletionPolicy() {}
  // 重新打开索引的时候，对所有commit的处理
  public abstract void onInit(List<? extends IndexCommit> commits) throws IOException;
  // 有新提交时对所有commit的处理
  public abstract void onCommit(List<? extends IndexCommit> commits) throws IOException;
}

从上面我可以看到，索引的删除策略其实只在两个地方进行应用，一个是加载索引的时候，打开一个旧索引时，根据当前设置的IndexDeletionPolicy进行处理。另一个是有新的commit产生时，借这个机会处理所有的commit。Lucene中提供的索引删除策略一共有四种，不过可以分为三类：

NoDeletionPolicy

NoDeletionPolicy索引删除策略就是保留所有的commit信息，效果就是你有多少次commit就多少个segments_N文件，看个例子：

public class DeletionPolicyTest {
    private static final Random RANDOM = new Random();
    public static void main(String[] args) throws IOException {
        Directory directory = FSDirectory.open(new File("D:\\code\\lucene-9.1.0-learning\\data").toPath());
        WhitespaceAnalyzer analyzer = new WhitespaceAnalyzer();
        IndexWriterConfig indexWriterConfig = new IndexWriterConfig(analyzer);
        indexWriterConfig.setUseCompoundFile(true);
        indexWriterConfig.setIndexDeletionPolicy(NoDeletionPolicy.INSTANCE);
        IndexWriter indexWriter = new IndexWriter(directory, indexWriterConfig);
        indexWriter.addDocument(getDoc(RANDOM.nextInt(10000),RANDOM.nextInt(10000)));
        // 第一次commit，生成segments_1
        indexWriter.commit();
        indexWriter.addDocument(getDoc(RANDOM.nextInt(10000),RANDOM.nextInt(10000)));
        // 第二次commit，生成segments_2
        indexWriter.commit();
        indexWriter.close();
    }
    private static Document getDoc(int... point) {
        Document doc = new Document();
        IntPoint intPoint = new IntPoint("point", point);
        doc.add(intPoint);
        return doc;
    }
}

上面的例子中有两次commit，下图是NoDeletionPolicy策略进行了两次commit的索引目录结构，可以看到生成了两个segments_N文件：

NoDeletionPolicy的代码实现非常简单，单例实现，并且在onCommit和onInit的时候都是空操作：

public final class NoDeletionPolicy extends IndexDeletionPolicy {
  public static final IndexDeletionPolicy INSTANCE = new NoDeletionPolicy();
  private NoDeletionPolicy() {
  }
  public void onCommit(List<? extends IndexCommit> commits) {}
  public void onInit(List<? extends IndexCommit> commits) {}
}

KeepOnlyLastCommitDeletionPolicy

KeepOnlyLastCommitDeletionPolicy是Lucene默认的索引删除策略，只保留最新的一次commit，从索引目录看不管执行多少次commit只保留了N最大的segments_N文件，下图是KeepOnlyLastCommitDeletionPolicy策略进行了两次commit的结果，KeepOnlyLastCommitDeletionPolicy删除策略只保留了segments_2。把上面示例代码中的删除策略替换成KeepOnlyLastCommitDeletionPolicy，即可得到，注意需要先清空索引目录：

KeepOnlyLastCommitDeletionPolicy代码实现也比较简单，除了最后一个commit之外，其他的commit都删除：

public final class KeepOnlyLastCommitDeletionPolicy extends IndexDeletionPolicy {
  public KeepOnlyLastCommitDeletionPolicy() {}
  public void onInit(List<? extends IndexCommit> commits) {
    onCommit(commits);
  }
  // commits是从旧到新排序的
  public void onCommit(List<? extends IndexCommit> commits) {
    // 只保留最新的一个
    int size = commits.size();
    for (int i = 0; i < size - 1; i++) {
      commits.get(i).delete();
    }
  }
}

两个快照相关的删除策略

快照相关的删除策略有两个，SnapshotDeletionPolicy和PersistentSnapshotDeletionPolicy，分别对应了不可持久化和可持久化的模式。不管是SnapshotDeletionPolicy还是PersistentSnapshotDeletionPolicy，他们都封装了其他的IndexDeletionPolicy来执行删除策略，他们只是提供了为当前最新的commit生成快照的能力。只要快照存在，则跟快照相关的所有索引文件都会被无条件保留。

SnapshotDeletionPolicy

例子

public class SnapshotDeletionPolicyTest {
    private static final Random RANDOM = new Random();
    public static void main(String[] args) throws IOException, InterruptedException {
        Directory directory = FSDirectory.open(new File("D:\\code\\lucene-9.1.0-learning\\data").toPath());
        WhitespaceAnalyzer analyzer = new WhitespaceAnalyzer();
        IndexWriterConfig indexWriterConfig = new IndexWriterConfig(analyzer);
        indexWriterConfig.setUseCompoundFile(true);
        SnapshotDeletionPolicy snapshotDeletionPolicy = new SnapshotDeletionPolicy(new KeepOnlyLastCommitDeletionPolicy());
        indexWriterConfig.setIndexDeletionPolicy(snapshotDeletionPolicy);
        IndexWriter indexWriter = new IndexWriter(directory, indexWriterConfig);
        indexWriter.addDocument(getDoc(RANDOM.nextInt(10000),RANDOM.nextInt(10000)));
        // 第一次commit，生成segments_1
        indexWriter.commit();
        indexWriter.addDocument(getDoc(RANDOM.nextInt(10000),RANDOM.nextInt(10000)));
        // 第二次commit，生成segments_2
        indexWriter.commit();
        // segments_2当做快照，无条件保留
        snapshotDeletionPolicy.snapshot();
        indexWriter.addDocument(getDoc(RANDOM.nextInt(10000),RANDOM.nextInt(10000)));
        // 第三次commit，生成segments_3
        indexWriter.commit();
        indexWriter.close();
    }
    private static Document getDoc(int... point) {
        Document doc = new Document();
        IntPoint intPoint = new IntPoint("point", point);
        doc.add(intPoint);
        return doc;
    }
}

在上面的例子中，我们使用SnapshotDeletionPolicy，SnapshotDeletionPolicy底层封装的是KeepOnlyLastCommitDeletionPolicy，我们进行了三次commit，理论上KeepOnlyLastCommitDeletionPolicy只会保留最后一次，但是因为我们对第一次的commit进行了快照，所以第一次commit也被保留了：

接下来我们看看SnapshotDeletionPolicy是怎么实现。SnapshotDeletionPolicy保证生成快照的commit不会被删除的原理就是引用计数，SnapshotDeletionPolicy会记录每个commit生成快照的次数，在删除的时候，只会删除引用计数为0的commit。

成员变量

  // key是IndexCommit的generation，value是对应的IndexCommit有多少个快照
  // 需要注意的是，有被快照引用的才会记录在refCounts中，也就是只要被记录在refCounts中，引用次数至少是1
  protected final Map<Long, Integer> refCounts = new HashMap<>();
  // key是IndexCommit的generation，value是对应的IndexCommit
  protected final Map<Long, IndexCommit> indexCommits = new HashMap<>();
  // SnapshotDeletionPolicy只是增加了支持快照的功能，删除的逻辑是由primary参数对应的删除策略提供的
  private final IndexDeletionPolicy primary;
  // 最近一次提交的commit，只会对这个IndexCommit生成快照
  protected IndexCommit lastCommit;
  // 是否初始化的标记，实例化后，必须先调用onInit方法
  private boolean initCalled;

生成快照

生成快照只会对当前最新的一个commit进行快照：

  public synchronized IndexCommit snapshot() throws IOException {
    if (!initCalled) {
      throw new IllegalStateException(
          "this instance is not being used by IndexWriter; be sure to use the instance returned from writer.getConfig().getIndexDeletionPolicy()");
    }
    if (lastCommit == null) {
      throw new IllegalStateException("No index commit to snapshot");
    }
    // 新增lastCommit的引用计数
    incRef(lastCommit);
    return lastCommit;
  }
  protected synchronized void incRef(IndexCommit ic) {
    long gen = ic.getGeneration();
    Integer refCount = refCounts.get(gen);
    int refCountInt;
    if (refCount == null) { // 第一次被引用
      indexCommits.put(gen, lastCommit);
      refCountInt = 0;
    } else {
      refCountInt = refCount.intValue();
    }
    // 引用计数加+1  
    refCounts.put(gen, refCountInt + 1);
  }

释放指定的快照

public synchronized void release(IndexCommit commit) throws IOException {
  long gen = commit.getGeneration();
  releaseGen(gen);
}
protected void releaseGen(long gen) throws IOException {
  if (!initCalled) {
    throw new IllegalStateException(
        "this instance is not being used by IndexWriter; be sure to use the instance returned from writer.getConfig().getIndexDeletionPolicy()");
  }
  Integer refCount = refCounts.get(gen);
  if (refCount == null) {
    throw new IllegalArgumentException("commit gen=" + gen + " is not currently snapshotted");
  }
  int refCountInt = refCount.intValue();
  assert refCountInt > 0;
  refCountInt--;
  if (refCountInt == 0) { // 引用计数为0，直接从refCounts中移除
    refCounts.remove(gen);
    indexCommits.remove(gen);
  } else {
    refCounts.put(gen, refCountInt);
  }
}

删除commit

  public synchronized void onCommit(List<? extends IndexCommit> commits) throws IOException {
    // 把commits中的所有IndexCommit都封装成SnapshotCommitPoint，再使用primary执行onCommit方法  
    primary.onCommit(wrapCommits(commits));
    // 更新最新的commit  
    lastCommit = commits.get(commits.size() - 1);
  }
  @Override
  public synchronized void onInit(List<? extends IndexCommit> commits) throws IOException {
    // 设置初始化的标记  
    initCalled = true;
    primary.onInit(wrapCommits(commits));
    for (IndexCommit commit : commits) { 
      if (refCounts.containsKey(commit.getGeneration())) {
        indexCommits.put(commit.getGeneration(), commit);
      }
    }
    if (!commits.isEmpty()) {
      lastCommit = commits.get(commits.size() - 1);
    }
  }
  private List<IndexCommit> wrapCommits(List<? extends IndexCommit> commits) {
    List<IndexCommit> wrappedCommits = new ArrayList<>(commits.size());
    for (IndexCommit ic : commits) {
      // 把IndexCommit都封装成 SnapshotCommitPoint
      wrappedCommits.add(new SnapshotCommitPoint(ic));
    }
    return wrappedCommits;
  }

前面我们列出了SnapshotCommitPoint是IndexCommit的一个实现类，但是没有详细介绍，SnapshotCommitPoint除了能够提供IndexCommit接口所提供的信息之外，最核心的是在删除的时候，会先判断IndexCommit是否被快照引用，只有没有任何快照引用的IndexCommit才能删除：

public void delete() {
    synchronized (SnapshotDeletionPolicy.this) {
        if (!refCounts.containsKey(cp.getGeneration())) {
            cp.delete();
        }
    }
}

存在的问题

需要注意的是SnapshotDeletionPolicy的快照信息是没有持久化，我们重新打开SnapshotDeletionPolicyTest例子中生成的索引：

public class SnapshotDeletionPolicyTest2 {
    public static void main(String[] args) throws IOException, InterruptedException {
        Directory directory = FSDirectory.open(new File("D:\\code\\lucene-9.1.0-learning\\data").toPath());
        WhitespaceAnalyzer analyzer = new WhitespaceAnalyzer();
        IndexWriterConfig indexWriterConfig = new IndexWriterConfig(analyzer);
        indexWriterConfig.setUseCompoundFile(true);
        SnapshotDeletionPolicy snapshotDeletionPolicy = new SnapshotDeletionPolicy(new KeepOnlyLastCommitDeletionPolicy());
        indexWriterConfig.setIndexDeletionPolicy(snapshotDeletionPolicy);
        // 重新打开索引
        IndexWriter indexWriter = new IndexWriter(directory, indexWriterConfig);
        indexWriter.close();
    }
}

可以发现segments_1被删除了，因为没有持久化快照信息，所以根据KeepOnlyLastCommitDeletionPolicy的删除策略，只保留了最新的一个commit：

PersistentSnapshotDeletionPolicy

例子

PersistentSnapshotDeletionPolicy主要是为了解决SnapshotDeletionPolicy无法持久化的问题。PersistentSnapshotDeletionPolicy持久化的时候会生成snapshots_N的索引文件，我们看个例子：

public class PersistentSnapshotDeletionPolicyTest {
    private static final Random RANDOM = new Random();
    public static void main(String[] args) throws IOException, InterruptedException {
        Directory directory = FSDirectory.open(new File("D:\\code\\lucene-9.1.0-learning\\data").toPath());
        WhitespaceAnalyzer analyzer = new WhitespaceAnalyzer();
        IndexWriterConfig indexWriterConfig = new IndexWriterConfig(analyzer);
        indexWriterConfig.setUseCompoundFile(true);
        PersistentSnapshotDeletionPolicy persistentSnapshotDeletionPolicy = new PersistentSnapshotDeletionPolicy(new KeepOnlyLastCommitDeletionPolicy(), directory);
        indexWriterConfig.setIndexDeletionPolicy(persistentSnapshotDeletionPolicy);
        IndexWriter indexWriter = new IndexWriter(directory, indexWriterConfig);
        indexWriter.addDocument(getDoc(RANDOM.nextInt(10000),RANDOM.nextInt(10000)));
        // 第一次commit，生成segments_1
        indexWriter.commit();
        // segments_1当做快照，无条件保留
        persistentSnapshotDeletionPolicy.snapshot();
        indexWriter.addDocument(getDoc(RANDOM.nextInt(10000),RANDOM.nextInt(10000)));
        // 第二次commit，生成segments_2
        indexWriter.commit();
        indexWriter.addDocument(getDoc(RANDOM.nextInt(10000),RANDOM.nextInt(10000)));
        // 第三次commit，生成segments_3
        indexWriter.commit();
        indexWriter.close();
    }
    private static Document getDoc(int... point) {
        Document doc = new Document();
        IntPoint intPoint = new IntPoint("point", point);
        doc.add(intPoint);
        return doc;
    }
}

上面的例子和我们在介绍SnapshotDeletionPolicy的时候逻辑一样，只是把SnapshotDeletionPolicy换成了PersistentSnapshotDeletionPolicy，我们看结果：

从上面结果图中可以看到，segments_1和segments_3同样被保留了，但是多了一个持久化的快照信息的文件snapshots_0，有了这个文件，索引重新打开的时候就可以恢复快照信息，segments_1还是会被保留，用下面的例子我们重新打开索引，可以发现segments_1还是被保留了：

public class PersistentSnapshotDeletionPolicyTest2 {
    public static void main(String[] args) throws IOException, InterruptedException {
        Directory directory = FSDirectory.open(new File("D:\\code\\lucene-9.1.0-learning\\data").toPath());
        WhitespaceAnalyzer analyzer = new WhitespaceAnalyzer();
        IndexWriterConfig indexWriterConfig = new IndexWriterConfig(analyzer);
        indexWriterConfig.setUseCompoundFile(true);
        PersistentSnapshotDeletionPolicy persistentSnapshotDeletionPolicy = new PersistentSnapshotDeletionPolicy(new KeepOnlyLastCommitDeletionPolicy(), directory);
        indexWriterConfig.setIndexDeletionPolicy(persistentSnapshotDeletionPolicy);
        IndexWriter indexWriter = new IndexWriter(directory, indexWriterConfig);
        indexWriter.close();
    }
}

接下来我们看看PersistentSnapshotDeletionPolicy的实现，主要就是持久化和恢复快照信息的逻辑。

成员变量

  // 持久化快照信息的文件名snapshots_N中的N，从0开始
  private long nextWriteGen;
  // 持久化的文件所在的目录
  private final Directory dir;

构造函数

  public PersistentSnapshotDeletionPolicy(IndexDeletionPolicy primary, Directory dir)
      throws IOException {
    this(primary, dir, OpenMode.CREATE_OR_APPEND);
  }
  public PersistentSnapshotDeletionPolicy(IndexDeletionPolicy primary, Directory dir, OpenMode mode)
      throws IOException {
    super(primary);
    this.dir = dir;
    if (mode == OpenMode.CREATE) { // 新建索引的模式，则需要清除所有的快照信息，索引模式以后再介绍
      clearPriorSnapshots();
    }
    // 加载快照信息
    loadPriorSnapshots();
    if (mode == OpenMode.APPEND && nextWriteGen == 0) {
      throw new IllegalStateException("no snapshots stored in this directory");
    }
  }

生成快照

public synchronized IndexCommit snapshot() throws IOException {
  // 使用SnapshotDeletionPolicy来生成快照  
  IndexCommit ic = super.snapshot();
  // 标记持久化是否成功，不成功的话需要删除快照  
  boolean success = false;
  try {
    // 持久化最新的快照信息
    persist();
    success = true;
  } finally {
    if (!success) { // 持久化失败，删除快照
      try {
        super.release(ic);
      } catch (
          @SuppressWarnings("unused")
          Exception e) {
        // Suppress so we keep throwing original exception
      }
    }
  }
  return ic;
}

释放快照

public synchronized void release(IndexCommit commit) throws IOException {
  // 使用SnapshotDeletionPolicy来释放快照  
  super.release(commit);
  // 持久化快照信息是否成功  
  boolean success = false;
  try {
    // 持久化最新的快照信息  
    persist();
    success = true;
  } finally {
    if (!success) { // 持久化失败，重新加回快照信息
      try {
        incRef(commit);
      } catch (
          @SuppressWarnings("unused")
          Exception e) {
        // Suppress so we keep throwing original exception
      }
    }
  }
}

持久化快照信息

private synchronized void persist() throws IOException {
  // 快照文件名  
  String fileName = SNAPSHOTS_PREFIX + nextWriteGen;
  boolean success = false;
  try (IndexOutput out = dir.createOutput(fileName, IOContext.DEFAULT)) {
    CodecUtil.writeHeader(out, CODEC_NAME, VERSION_CURRENT);
    out.writeVInt(refCounts.size());
    for (Entry<Long, Integer> ent : refCounts.entrySet()) { // 持久化所有的引用信息
      out.writeVLong(ent.getKey());
      out.writeVInt(ent.getValue());
    }
    success = true;
  } finally {
    if (!success) {
      IOUtils.deleteFilesIgnoringExceptions(dir, fileName);
    }
  }
  dir.sync(Collections.singletonList(fileName));
  if (nextWriteGen > 0) {
    String lastSaveFile = SNAPSHOTS_PREFIX + (nextWriteGen - 1);
    // 删除前一个快照文件，因为每次持久化都是把当前的快照信息全量持久化，所以只需要保留最新的一个就可以
    // 这里有可能删除失败，所以在启动加载的时候会再次尝试把旧版本的文件都删掉  
    IOUtils.deleteFilesIgnoringExceptions(dir, lastSaveFile);
  }
  nextWriteGen++;
}

加载快照信息

private synchronized void loadPriorSnapshots() throws IOException {
  long genLoaded = -1;
  IOException ioe = null;
  List<String> snapshotFiles = new ArrayList<>();
  for (String file : dir.listAll()) {
    if (file.startsWith(SNAPSHOTS_PREFIX)) { // 找到快照文件
      long gen = Long.parseLong(file.substring(SNAPSHOTS_PREFIX.length()));
      if (genLoaded == -1 || gen > genLoaded) { // 找到gen最大的快照文件
        snapshotFiles.add(file);
        Map<Long, Integer> m = new HashMap<>();
        IndexInput in = dir.openInput(file, IOContext.DEFAULT);
        try {
          CodecUtil.checkHeader(in, CODEC_NAME, VERSION_START, VERSION_START);
          int count = in.readVInt();
          for (int i = 0; i < count; i++) {
            long commitGen = in.readVLong();
            int refCount = in.readVInt();
            m.put(commitGen, refCount);
          }
        } catch (IOException ioe2) {
          // 保存第一个捕获到的异常
          if (ioe == null) {
            ioe = ioe2;
          }
        } finally {
          in.close();
        }
        genLoaded = gen;
        // 清除旧数据  
        refCounts.clear();
        // 保留最新的  
        refCounts.putAll(m);
      }
    }
  }
  if (genLoaded == -1) { // 没有加载快照文件
    if (ioe != null) { // 加载过程中捕获到异常了，直接抛出
      throw ioe;
    }
  } else { // 把旧版本的快照文件都删掉
    if (snapshotFiles.size() > 1) {
      String curFileName = SNAPSHOTS_PREFIX + genLoaded;
      for (String file : snapshotFiles) {
        if (!curFileName.equals(file)) {
          IOUtils.deleteFilesIgnoringExceptions(dir, file);
        }
      }
    }
    nextWriteGen = 1 + genLoaded;
  }
}

总结

本文介绍的索引删除策略是在IndexCommit粒度的控制，具体到每个索引文件是怎么控制的，我们下一篇文章介绍，更多关于Lucene 索引删除策略的资料请关注脚本之家其它相关文章！

您可能感兴趣的文章:
阅读全文