背景

本文基于 Starrocks 3.1.7

结论

Starrocks 会启动一个线程周期性的去进行Compaction,该周期间隔为 200 MS, 该Compaction以table的partition为切入点,tablet(也就是bucket)为粒度进行task的创建。

分析

CompactionMgr start 方法会启动一个CompactionScheduler 用来启动一个 合并的周期性任务.
这里的周期会由 LOOP_INTERVAL_MS参数控制,默认是 200ms.
然后每个周期内会调用 runOneCycle 方法:

    protected void runOneCycle() {
        cleanPartition();

        // Schedule compaction tasks only when this is a leader FE and all edit logs have finished replay.
        // In order to ensure that the input rowsets of compaction still exists when doing publishing version, it is
        // necessary to ensure that the compaction task of the same partition is executed serially, that is, the next
        // compaction task can be executed only after the status of the previous compaction task changes to visible or
        // canceled.
        if (stateMgr.isLeader() && stateMgr.isReady() && allCommittedCompactionsBeforeRestartHaveFinished()) {
            schedule();
            history.changeMaxSize(Config.lake_compaction_history_size);
            failHistory.changeMaxSize(Config.lake_compaction_fail_history_size);
        }
    }

  • cleanPartition 这里会清除无效的分区,便于后续进行Compaction
  • 这里会有个 FE leader的判断(这里所涉及到的GlobalStateMgr只是单个FE的状态),只有是leader节点才可以进行Compaction,最主要的逻辑还是在schedule
    方法中:
      for (Iterator<Map.Entry<PartitionIdentifier, CompactionJob>> iterator = runningCompactions.entrySet().iterator();
      ...
      if (job.isCompleted()) {
          job.getPartition().setMinRetainVersion(0);
          try {
              commitCompaction(partition, job);
              assert job.transactionHasCommitted();
          } catch (Exception e) {
              ...
          }
      } else if (job.isFailed()) {
          job.getPartition().setMinRetainVersion(0);
          errorMsg = Objects.requireNonNull(job.getFailMessage(), "getFailMessage() is null");
          job.abort(); // Abort any executing task, if present.
      }
    
      if (errorMsg != null) {
          iterator.remove();
          job.finish();
          failHistory.offer(CompactionRecord.build(job, errorMsg));
          compactionManager.enableCompactionAfter(partition, MIN_COMPACTION_INTERVAL_MS_ON_FAILURE);
          abortTransactionIgnoreException(partition.getDbId(), job.getTxnId(), errorMsg);
          continue;
      }
      ...
      int index = 0;
      int compactionLimit = compactionTaskLimit();
      int numRunningTasks = runningCompactions.values().stream().mapToInt(CompactionJob::getNumTabletCompactionTasks).sum();
      if (numRunningTasks >= compactionLimit) {
          return;
      }
    
      List<PartitionIdentifier> partitions = compactionManager.choosePartitionsToCompact(runningCompactions.keySet());
      while (numRunningTasks < compactionLimit && index < partitions.size()) {
          PartitionIdentifier partition = partitions.get(index++);
          CompactionJob job = startCompaction(partition);
          if (job == null) {
              continue;
          }
          numRunningTasks += job.getNumTabletCompactionTasks();
          runningCompactions.put(partition, job);
          if (LOG.isDebugEnabled()) {
              LOG.debug("Created new compaction job. partition={} txnId={}", partition, job.getTxnId());
          }
      }
              
    
    • 选取正在进行的Compaction的job,如果该任务完成了compaction(每个tablets都完成了compaction) ,但是事务没有提交,则完成compaction事务的提交,
      否则如果任务失败了,则abort该job。最终会把该任务从runnning队列中移除掉。如果是失败任务的话,还会记录到failHistory中,并会重新进行Compaction的任务的延迟提交(延迟间隔为LOOP_INTERVAL_MS*10,其中LOOP_INTERVAL_MS 为200ms)

    • 如果Compaction事务已经提交了,则会记录到history中,并会重新进行Compaction的任务的延迟提交(延迟间隔为LOOP_INTERVAL_MS*2,其中LOOP_INTERVAL_MS 为200ms)

    • 处理完正在运行的Compaction任务后,会构建当前的Compaction任务

      • 首先会通过compactionTaskLimit方法获取本次Compaction任务的个数限制,如果lake_compaction_max_tasks大于等于0,则会根据lake_compaction_max_tasks配置来,否则会根据系统的BE数和CN数乘以16来计算。
      • 如果 运行的task(以Tablets为粒度计数的)大于了该compactionTaskLimit,则此次Compaction结束,否则继续下一步
      • compactionManager.choosePartitionsToCompact 从已有的分区中。并且排除掉 runningCompactions里正在运行的Compaction任务中涉及的partition。
        choosePartitionsToCompact 涉及到Sorter(默认ScoreSorter) 和selector(ScoreSelector),
        ScoreSelector 会选择 lake_compaction_score_selector_min_score(默认为10)并且到了合并的时间的分区
        ScoreSorter 会按照compactionScore 从高到低进行排序
      • 对于每一个被选出来的分区,会进行调用startCompaction方法进行compaction任务的构建
        这里会调用collectPartitionTablets方法,用来选择tablet以及对应的该tablet对应的backend
      • 调用createCompactionTasks创建CompactionTask,这里有多少个backend就有多少个task
        调用thrift rpc服务往对应的backend发送Compact请求,并组装成CompactionJob
          List<CompactionTask> tasks = new ArrayList<>();
          for (Map.Entry<Long, List<Long>> entry : beToTablets.entrySet()) {
              ComputeNode node = systemInfoService.getBackendOrComputeNode(entry.getKey());
              if (node == null) {
                  throw new UserException("Node " + entry.getKey() + " has been dropped");
              }
        
              LakeService service = BrpcProxy.getLakeService(node.getHost(), node.getBrpcPort());
        
              CompactRequest request = new CompactRequest();
              request.tabletIds = entry.getValue();
              request.txnId = txnId;
              request.version = currentVersion;
              request.timeoutMs = LakeService.TIMEOUT_COMPACT;
        
              CompactionTask task = new CompactionTask(node.getId(), service, request);
              tasks.add(task);
          }
          return tasks;
        
    • 累计numRunningTasks计数,便于控制Compaction的并发执行,并且回放到 runningCompactions

其他

前文提到的 一些 FE的配置 ,如lake_compaction_max_tasks 都是可以配置的,
可以通过 命令* admin set frontend config (“lake_compaction_max_tasks” = “0”);* ,具体的参考ADMIN_SET_CONFIG,
注意: 这个命令只是修改了当前内存中的变量的值,如果需要永久的修改,需要配置到fe.conf

点赞(0) 打赏

评论列表 共有 0 条评论

暂无评论

微信公众账号

微信扫一扫加关注

发表
评论
返回
顶部