摘要：的選哪個首選斷點還原可以記錄偏移量可配置文件組，里面使用正則表達式配置多個要監控的文件就憑第一點其他的都被比下去了這么好的有一點不完美，不能支持遞歸監控文件夾。

Flume的source選哪個？
taildir source首選！
1.斷點還原 positionFile可以記錄偏移量
2.可配置文件組，里面使用正則表達式配置多個要監控的文件
就憑第一點其他的source都被比下去了！
這么好的taildir source有一點不完美，不能支持遞歸監控文件夾。
所以就只能修改源代碼了……好玩，我喜歡~

Flume的taildir source啟動會調用start()方法作初始化，里面創建一個ReliableTaildirEventReader,這里用到了建造者模式

@Override
public synchronized void start() {
    logger.info("{} TaildirSource source starting with directory: {}", getName(), filePaths);
    try {
        reader = new ReliableTaildirEventReader.Builder()
                .filePaths(filePaths)
                .headerTable(headerTable)
                .positionFilePath(positionFilePath)
                .skipToEnd(skipToEnd)
                .addByteOffset(byteOffsetHeader)
                .cachePatternMatching(cachePatternMatching)
                .recursive(isRecursive)
                .annotateFileName(fileHeader)
                .fileNameHeader(fileHeaderKey)
                .build();
    } catch (IOException e) {
        throw new FlumeException("Error instantiating ReliableTaildirEventReader", e);
    }
    idleFileChecker = Executors.newSingleThreadScheduledExecutor(
            new ThreadFactoryBuilder().setNameFormat("idleFileChecker").build());
    idleFileChecker.scheduleWithFixedDelay(new idleFileCheckerRunnable(),
            idleTimeout, checkIdleInterval, TimeUnit.MILLISECONDS);

    positionWriter = Executors.newSingleThreadScheduledExecutor(
            new ThreadFactoryBuilder().setNameFormat("positionWriter").build());
    positionWriter.scheduleWithFixedDelay(new PositionWriterRunnable(),
            writePosInitDelay, writePosInterval, TimeUnit.MILLISECONDS);

    super.start();
    logger.debug("TaildirSource started");
    sourceCounter.start();
}

taildir source屬于PollableSource，

/**
 * A {@link Source} that requires an external driver to poll to determine
 * whether there are {@linkplain Event events} that are available to ingest
 * from the source.
 *
 * @see org.apache.flume.source.EventDrivenSourceRunner
 */
public interface PollableSource extends Source {
    ...

這段注釋的意思是PollableSource是需要一個外部驅動去查看有沒有需要消費的事件，從而拉取事件，講白了就是定時拉取。所以flume也不一定是真正實時的，只是隔一會兒不停地來查看事件而已。(與之相應的是另一種EventDrivenSourceRunner)
那么taildir source在定時拉取事件的時候是調用的process方法

@Override
public Status process() {
    Status status = Status.READY;
    try {
        existingInodes.clear();
        existingInodes.addAll(reader.updateTailFiles());
        for (long inode : existingInodes) {
            TailFile tf = reader.getTailFiles().get(inode);
            if (tf.needTail()) {
                tailFileProcess(tf, true);
            }
        }
        closeTailFiles();
        try {
            TimeUnit.MILLISECONDS.sleep(retryInterval);
        } catch (InterruptedException e) {
            logger.info("Interrupted while sleeping");
        }
    } catch (Throwable t) {
        logger.error("Unable to tail files", t);
        status = Status.BACKOFF;
    }
    return status;
}

重點就是下面這幾行

existingInodes.addAll(reader.updateTailFiles());
for (long inode : existingInodes) {
   TailFile tf = reader.getTailFiles().get(inode);
   if (tf.needTail()) {
   tailFileProcess(tf, true);
 } }
從reader.updateTailFiles()獲取需要監控的文件，然后對每一個進行處理，查看最后修改時間，判定是否需要tail，需要tail就tail
那么進入reader.updateTailFiles()

for (TaildirMatcher taildir : taildirCache) {
      Map headers = headerTable.row(taildir.getFileGroup());

      for (File f : taildir.getMatchingFiles()) {
        long inode = getInode(f);
        TailFile tf = tailFiles.get(inode);
        if (tf == null || !tf.getPath().equals(f.getAbsolutePath())) {
          long startPos = skipToEnd ? f.length() : 0;
          tf = openFile(f, headers, inode, startPos);

遍歷每一個正則表達式匹配對應的匹配器，每個匹配器去獲取匹配的文件！taildir.getMatchingFiles()

List getMatchingFiles() {
    long now = TimeUnit.SECONDS.toMillis(
            TimeUnit.MILLISECONDS.toSeconds(System.currentTimeMillis()));
    long currentParentDirMTime = parentDir.lastModified();
    List result;

    // calculate matched files if
    // - we don"t want to use cache (recalculate every time) OR
    // - directory was clearly updated after the last check OR
    // - last mtime change wasn"t already checked for sure
    //   (system clock hasn"t passed that second yet)
    if (!cachePatternMatching ||
            lastSeenParentDirMTime < currentParentDirMTime ||
            !(currentParentDirMTime < lastCheckedTime)) {
        lastMatchedFiles = sortByLastModifiedTime(getMatchingFilesNoCache(isRecursive));
        lastSeenParentDirMTime = currentParentDirMTime;
        lastCheckedTime = now;
    }

    return lastMatchedFiles;
}

可以看到getMatchingFilesNoCache(isRecursive)就是獲取匹配的文件的方法，也就是需要修改的方法了！
ps：這里的isRecursive是我加的~
點進去：

private List getMatchingFilesNoCache() {
    List result = Lists.newArrayList();
    try (DirectoryStream stream = Files.newDirectoryStream(parentDir.toPath(), fileFilter)) {
        for (Path entry : stream) {
            result.add(entry.toFile());
        }
    } catch (IOException e) {
        logger.error("I/O exception occurred while listing parent directory. " +
                "Files already matched will be returned. " + parentDir.toPath(), e);
    }
    return result;
}

源碼是用了Files.newDirectoryStream(parentDir.toPath(), fileFilter))，將父目錄下符合正則表達式的文件都添加到一個迭代器里。（這里還用了try (...)的語法糖）

我在這個getMatchingFilesNoCache()方法下面下了一個重載的方法, 可增加擴展性：

private List getMatchingFilesNoCache(boolean recursion) {
    if (!recursion) {
        return getMatchingFilesNoCache();
    }
    List result = Lists.newArrayList();
    // 使用非遞歸的方式遍歷文件夾
    Queue dirs = new ArrayBlockingQueue<>(10);
    dirs.offer(parentDir);
    while (dirs.size() > 0) {
        File dir = dirs.poll();
        try {
            DirectoryStream stream = Files.newDirectoryStream(dir.toPath(), fileFilter);
            stream.forEach(path -> result.add(path.toFile()));
        } catch (IOException e) {
            logger.error("I/O exception occurred while listing parent directory. " +
                    "Files already matched will be returned. (recursion)" + parentDir.toPath(), e);
        }
        File[] dirList = dir.listFiles();
        assert dirList != null;
        for (File f : dirList) {
            if (f.isDirectory()) {
                dirs.add(f);
            }
        }
    }
    return result;
}

我使用了非遞歸的方式遍歷文件夾，就是樹到隊列的轉換。
到這里，核心部分就改完了。接下來要處理這個recursion的參數

一路改構造方法，添加這個參數，最終參數從哪來呢？
flume的source啟動時會調用configure方法，將Context中的內容配置進reader等對象中。
isRecursive = context.getBoolean(RECURSIVE, DEFAULT_RECURSIVE);
context從TaildirSourceConfigurationConstants中獲取配置名和默認值

  /**
   * Whether to support recursion. */
  public static final String RECURSIVE = "recursive";
  public static final boolean DEFAULT_RECURSIVE = false;

這里的recursive也就是flume配置文件里配置項了

# Whether to support recusion
a1.sources.r1.recursive = true

用maven只對這一個module打包。我把這個module的pom改了下artifactId，加上了自己名字作個紀念，哈哈
可惜pom里面不能寫中文……

org.apache.flume.flume-ng-sources
flume-taildir-source-recursive-by-Wish000
Flume Taildir Source

執行package將其放在flume的lib下，替換原來的flume-taildir-source***.jar
啟動，測試，成功！

具體代碼見GitHub地址：https://github.com/Wish000/me...