Apache Hudi初探(五)(与flink的结合)--Flink 中hudi clean操作

背景

本文主要是具体说说Flink中的clean操作的实现

杂说闲谈

在flink中主要是CleanFunction函数：

  @Override
  public void open(Configuration parameters) throws Exception {
    super.open(parameters);
    this.writeClient = FlinkWriteClients.createWriteClient(conf, getRuntimeContext());
    this.executor = NonThrownExecutor.builder(LOG).waitForTasksFinish(true).build();
    String instantTime = HoodieActiveTimeline.createNewInstantTime();
    LOG.info(String.format("exec clean with instant time %s...", instantTime));
    executor.execute(() -> writeClient.clean(instantTime), "wait for cleaning finish");
  }

  @Override
  public void notifyCheckpointComplete(long l) throws Exception {
    if (conf.getBoolean(FlinkOptions.CLEAN_ASYNC_ENABLED) && isCleaning) {
      executor.execute(() -> {
        try {
          this.writeClient.waitForCleaningFinish();
        } finally {
          // ensure to switch the isCleaning flag
          this.isCleaning = false;
        }
      }, "wait for cleaning finish");
    }
  }

  @Override
  public void snapshotState(FunctionSnapshotContext context) throws Exception {
    if (conf.getBoolean(FlinkOptions.CLEAN_ASYNC_ENABLED) && !isCleaning) {
      try {
        this.writeClient.startAsyncCleaning();
        this.isCleaning = true;
      } catch (Throwable throwable) {
        // catch the exception to not affect the normal checkpointing
        LOG.warn("Error while start async cleaning", throwable);
      }
    }
  }

• open函数

  • writeClient =FlinkWriteClients.createWriteClient(conf, getRuntimeContext())

    创建FlinkWriteClient,用于写hudi数据

  • this.executor = NonThrownExecutor.builder(LOG).waitForTasksFinish(true).build();

    创建一个只有一个线程的线程池，改线程池的主要作用来异步执行hudi写操作

  • executor.execute(() -> writeClient.clean(instantTime)

    异步执行hudi的清理操作,该clean函数的主要代码如下：

       if (!tableServicesEnabled(config)) {
         return null;
       }
       final Timer.Context timerContext = metrics.getCleanCtx();
       CleanerUtils.rollbackFailedWrites(config.getFailedWritesCleanPolicy(),
           HoodieTimeline.CLEAN_ACTION, () -> rollbackFailedWrites(skipLocking));
    
       HoodieTable table = createTable(config, hadoopConf);
       if (config.allowMultipleCleans() || !table.getActiveTimeline().getCleanerTimeline().filterInflightsAndRequested().firstInstant().isPresent()) {
         LOG.info("Cleaner started");
         // proceed only if multiple clean schedules are enabled or if there are no pending cleans.
         if (scheduleInline) {
           scheduleTableServiceInternal(cleanInstantTime, Option.empty(), TableServiceType.CLEAN);
           table.getMetaClient().reloadActiveTimeline();
         }
       }
    
       // Proceeds to execute any requested or inflight clean instances in the timeline
       HoodieCleanMetadata metadata = table.clean(context, cleanInstantTime, skipLocking);
       if (timerContext != null && metadata != null) {
         long durationMs = metrics.getDurationInMs(timerContext.stop());
         metrics.updateCleanMetrics(durationMs, metadata.getTotalFilesDeleted());
         LOG.info("Cleaned " + metadata.getTotalFilesDeleted() + " files"
             + " Earliest Retained Instant :" + metadata.getEarliestCommitToRetain()
             + " cleanerElapsedMs" + durationMs);
       }
       return metadata;
    

    • CleanerUtils.rollbackFailedWrites(config.getFailedWritesCleanPolicy(),HoodieTimeline.CLEAN_ACTION,() -> rollbackFailedWrites *
      根据配置 hoodie.cleaner.policy.failed.writes* 默认是EAGER ,也就是在写数据失败的时候，会立即进行这次写失败的数据的清理，在这种情况下，

      就不会执行rollbackFailedWrites操作，也就是回滚写失败文件的操作

    • HoodieTable table = createTable *
      创建HoodieFlinkMergeOnReadTable*类型的hudi表,用来做clean等操作

    • scheduleTableServiceInternal

      如果hoodie.clean.allow.multiple 为true(默认为true)或者没有正在运行中clean操作，则会生成Clean计划

      这里最终调用的是FlinkWriteClient.scheduleCleaning 方法,即CleanPlanActionExecutor.execute方法

      这里最重要的就是requestClean方法：

      CleanPlanner<T, I, K, O> planner = new CleanPlanner<>(context, table, config);
      Option<HoodieInstant> earliestInstant = planner.getEarliestCommitToRetain();
      List<String> partitionsToClean = planner.getPartitionPathsToClean(earliestInstant)    
      int cleanerParallelism = Math.min(partitionsToClean.size(), config.getCleanerParallelism());
      Map<String, Pair<Boolean, List<CleanFileInfo>>> cleanOpsWithPartitionMeta = context
          .map(partitionsToClean, partitionPathToClean -> Pair.of(partitionPathToClean, planner.getDeletePaths(partitionPathToClean)), cleanerParallelism)
          .stream()
          .collect(Collectors.toMap(Pair::getKey, Pair::getValue))    
      Map<String, List<HoodieCleanFileInfo>> cleanOps = cleanOpsWithPartitionMeta.entrySet().stream()
          .collect(Collectors.toMap(Map.Entry::getKey,
              e -> CleanerUtils.convertToHoodieCleanFileInfoList(e.getValue().getValue())))    
      List<String> partitionsToDelete = cleanOpsWithPartitionMeta.entrySet().stream().filter(entry -> entry.getValue().getKey()).map(Map.Entry::getKey)
          .collect(Collectors.toList())    
      return new HoodieCleanerPlan(earliestInstant
          .map(x -> new HoodieActionInstant(x.getTimestamp(), x.getAction(), x.getState().name())).orElse(null),
          planner.getLastCompletedCommitTimestamp(),
          config.getCleanerPolicy().name(), CollectionUtils.createImmutableMap(),
          CleanPlanner.LATEST_CLEAN_PLAN_VERSION, cleanOps, partitionsToDelete)    
      

      • planner.getEarliestCommitToRetain();
        根据保留策略，获取到最早需要保留的commit的HoodieInstant,在这里会兼顾考虑到hoodie.cleaner.commits.retained (默认是10)以及hoodie.cleaner.hours.retained 默认是24小时以及hoodie.cleaner.policy策略（默认是KEEP_LATEST_COMMITS）
      • planner.getPartitionPathsToClean(earliestInstant);
        根据保留的最新commit的HoodieInstant,得到要删除的分区,这里会根据配置hoodie.cleaner.incremental.mode (默认是true)来进行增量清理,
        这个时候就会根据上一次已经clean的信息，只需要删除差量的分区数据就行
      • cleanOpsWithPartitionMeta = context
        根据上面得到的需要删除的分区信息，获取需要删除的文件信息,具体的实现可以参考CleanPlanner.getFilesToCleanKeepingLatestCommits
        这里的操作主要是先通过fileSystemView 获取分区下所有的FileGroup ,之后再获取每个FileGroup 下的所有的FileSlice(这里的FileSlice就有版本的概念，也就是commit的版本),之后再与最新保留的commit的时间戳进行比较得到需要删除的文件信息
      • new HoodieCleanerPlan
        最后组装成HoodieCleanPlan的计划,并且在外层调用table.getActiveTimeline().saveToCleanRequested(cleanInstant, TimelineMetadataUtils.serializeCleanerPlan(cleanerPlan)); 方法把clean request的状态存储到对应的.hoodie目录下，并建立一个xxxx.clean.requested的元数据文件

    • table.getMetaClient().reloadActiveTimeline()

      重新加载timeline,便于过滤出来刚才scheduleTableServiceInternal 操作生成的xxxxxxxxxxxxxx.clean.requested的元数据文件

    • table.clean(context, cleanInstantTime, skipLocking)

      真正执行clean的部分，主要是调用CleanActionExecutor.execute的方法,最终调用的是*runPendingClean(table, hoodieInstant)*方法:

       HoodieCleanerPlan cleanerPlan = CleanerUtils.getCleanerPlan(table.getMetaClient(), cleanInstant);
       return runClean(table, cleanInstant, cleanerPlan);
      

      首先是反序列化CleanPlan,然后在进行清理，主要是删除1. 如果没有满足的分区，直接删除该分区，2. 否则删除该分区下的满足条件的文件,最后返回HoodieCleanStat包含删除的文件信息等。

• snapshotState方法

  • 如果clean.async.enabled 是true(默认是true)，并且不是正在进行clean动作，则会进行异步清理
    this.writeClient.startAsyncCleaning(); 这里最终也是调用的writeClient.clean方法。
  • this.isCleaning = true;
    设置标志位，用来保证clean操作的有序性

• notifyCheckpointComplete方法

  • 如果clean.async.enabled 是true(默认是true)，并且正在进行clean动作，则等待clean操作完成，
    并且设置清理标识位，用来和snapshotState方法进行呼应以保证clean操作的有序性