Flink任务物理内存溢出问题定位

问题现象

一个使用10秒滚动窗口的任务在平稳运行一段时间之后出现了频繁的重启。在TaskManager日志中能看到以下文本:

2019-03-17 16:05:28,854 INFO  org.apache.flink.yarn.YarnTaskExecutorRunner                  - RECEIVED SIGNAL 15: SIGTERM. Shutting down as requested.

原因定位

  • 首先可以看到是YarnTaskExecutorRunner收到了SIGTERM信号, 因为是部署在Yarn上,所以基本可以定位到是Yarn因为什么原因从OS的层面将这个进程给Kill掉的。
    • 代码上也可以根据这个日志可以定位到Flink的SignalHandler,下图可以看到Handler的构造调用过程。不管是YarnSessionClusterEntrypoint还是YarnTaskExecutorRunner的主函数都会注册,并且会在接收到OS的"TERM", "HUP", "INT"信号是打出日志。
      private static class Handler implements sun.misc.SignalHandler {
          private final Logger LOG;
          private final sun.misc.SignalHandler prevHandler;
          Handler(String name, Logger LOG) {
              this.LOG = LOG;
              prevHandler = Signal.handle(new Signal(name), this);
          }
          /**
           * Handle an incoming signal.
           *
           * @param signal    The incoming signal
           */
          @Override
          public void handle(Signal signal) {
              LOG.info("RECEIVED SIGNAL {}: SIG{}. Shutting down as requested.",
                  signal.getNumber(),
                  signal.getName());
              prevHandler.handle(signal);
          }
      }
    
    Handler构造器的调用
  • 接着可以观察这个TaskManager所在机器的Yarn的NodeManager的日志,grep出和这个容器相关的日志,可以看到最后如下。TaskManager内存超出了物理内存的限制。但是从GC日志来看,连Full GC都很少发生。
    2019-03-19 16:48:10,647 INFO   org.apache.hadoop.yarn.server.nodemanager.containermanager.monitor.ContainersMonitorImpl: Memory usage of ProcessTree 10265 for container-id container_1541225469893_4985_01_000005: 3.4 GB of 4 GB physical memory used; 5.4 GB of 8.4 GB virtual memory used
    
  • 因为开了taskmanager.memory.off-heap=true选项,所以Flink内部也会使用一些堆外的内存。还有就是RocksDB也会直接通过malloc分配内存。

堆外内存排查(大型绕弯路现场,想知道直接原因可以直接跳到最后)

  • 在开启堆外内存优化时,Flink的MemoryManager和NetworkBufferPool会使用ByteBuffer.allocateDirect方法来创建DirectByteBuffer,以此来使用堆外内存。但是通过heap dump,可以看到DirectByteBuffer数量及其有限,因为使用的是默认的taskmanager.memory.segment-size,也就是32KB,所以占用的堆外内存也只有几百兆,而我预留了两点几G的堆外内存,显然不是这个引起的。这时问题排查一度陷入了死胡同。
    Class Name               | Objects | Shallow Heap | Retained Heap
    ------------------------------------------------------------------
    java.nio.DirectByteBuffer|   9,470 |      606,080 |    >= 606,096
    ------------------------------------------------------------------
    
  • 还有一个可疑的点是RocksDB,但是很难去排查它到底占用了多少内存。
  • 在大佬的帮助下,在性能环境上安装了Jemalloc来代替原来的malloc,关于jemalloc的安装参考文档
  • 并且在flink-conf.yaml中配置如下参数,将其注入到container的系统环境变量中,使其生效。这样可以定期把memory profile dump出来,进行分析, 发现最后malloc最多的是rocksdb的rocksdb::UncompressBlockContentsForCompressionType方法,并且最终占到了2.15G内存的47%,TaskManager也被Yarn给kill掉。
    containerized.master.env.LD_PRELOAD: "/opt/jemalloc/lib/libjemalloc.so.2"
    containerized.master.env.MALLOC_CONF: "prof:true,lg_prof_interval:25,lg_prof_sample:17"
    
  • 使用/opt/jemalloc/bin/jeprof --show_bytes /opt/java/bin/java jeprof.xxx 来分析dump文件,使用top来显示对调用malloc最多的方法
  • 对比前后dump文件如下
    Using local file /opt/java/bin/java.
    Using local file jeprof.18091.9812.i9812.heap.
    Welcome to jeprof!  For help, type 'help'.
    (jeprof) top
    Total: 1107642087 B
    884271340  79.8%  79.8% 884271340  79.8% os::malloc@921040
    150994944  13.6%  93.5% 150994944  13.6% rocksdb::Arena::AllocateNewBlock
    51761789   4.7%  98.1% 51761789   4.7% rocksdb::UncompressBlockContentsForCompressionType
     5242880   0.5%  98.6%  5242880   0.5% init
     5184828   0.5%  99.1%  5184828   0.5% updatewindow
     4204536   0.4%  99.5%  4204536   0.4% readCEN
     1643018   0.1%  99.6%  1643018   0.1% std::basic_string::_Rep::_S_create
     1346886   0.1%  99.7%  1346886   0.1% inflateInit2_
      917840   0.1%  99.8%  1181009   0.1% rocksdb::LRUCacheShard::Insert
      393336   0.0%  99.8% 52155125   4.7% rocksdb::BlockBasedTable::GetTableProperties
    
    (jeprof) top
    Total: 2259309361 B
    1062208712  47.0%  47.0% 1062208712  47.0% rocksdb::UncompressBlockContentsForCompressionType
    884120659  39.1%  86.1% 884120659  39.1% os::malloc@921040
    285348930  12.6%  98.8% 285348930  12.6% rocksdb::Arena::AllocateNewBlock
     5451379   0.2%  99.0%  5451379   0.2% std::basic_string::_Rep::_S_create
     5242880   0.2%  99.3%  5242880   0.2% init
     5036690   0.2%  99.5%  5036690   0.2% updatewindow
     4204536   0.2%  99.7%  4204536   0.2% readCEN
     2621559   0.1%  99.8%  2621559   0.1% rocksdb::WritableFileWriter::Append
     1346886   0.1%  99.8%  1346886   0.1% inflateInit2_
      524472   0.0%  99.9%   788155   0.0% rocksdb::LRUCacheShard::Insert
    
jeprof.pdf
  • 在搜索这个方法后,发现有这个issue的github issue链接,实际上也不是Memory Leak,在默认配置下,rocksdb会为所有flush的sst文件在内存中保留索引,索引会随着文件数越来越多而占用更多的内存空间,如果限制内存中索引的消耗,会导致经常需要去sst文件中获取元信息来搜索,大量增加io消耗(这块不是特别熟悉,有可能说的有点问题),那为什么RocksDB文件会不停膨胀?

最终问题定位(走完弯路)

  • RocksDB文件不断膨胀,可以从checkpoint的大小来看出来,将incremental checkpoint关闭后,发现每次Checkpoint大小都在递增,但是用户代码的逻辑实际是使用一个10s的滚动窗口,不应该会出现这样的情况。
  • 之后在flink窗口算子中加了几行日志,如下所示,以ClarkTest开头
    @Override
    public void onProcessingTime(InternalTimer<K, W> timer) throws Exception {
        triggerContext.key = timer.getKey();
        triggerContext.window = timer.getNamespace();

        MergingWindowSet<W> mergingWindows;

        if (windowAssigner instanceof MergingWindowAssigner) {
            mergingWindows = getMergingWindowSet();
            W stateWindow = mergingWindows.getStateWindow(triggerContext.window);
            if (stateWindow == null) {
                // Timer firing for non-existent window, this can only happen if a
                // trigger did not clean up timers. We have already cleared the merging
                // window and therefore the Trigger state, however, so nothing to do.
                return;
            } else {
                windowState.setCurrentNamespace(stateWindow);
            }
        } else {
            windowState.setCurrentNamespace(triggerContext.window);
            mergingWindows = null;
        }

        TriggerResult triggerResult = triggerContext.onProcessingTime(timer.getTimestamp());

        int randomInt = random.nextInt(1000);
        if (triggerResult.isFire()) {
            ACC contents = windowState.get();
            if (randomInt == 1) {
                LOG.info("ClarkTest: Window state namespace: " + triggerContext.window + " and key " + triggerContext.key);
                LOG.info("ClarkTest: Window state value is going to fire is null ? " + (windowState.get() == null));
            }
            if (contents != null) {
                emitWindowContents(triggerContext.window, contents);
            }
        }

        if (triggerResult.isPurge()) {
            if (randomInt == 1) {
                LOG.info("ClarkTest: Window state get purged. ");
            }
            windowState.clear();
        }

        if (!windowAssigner.isEventTime() && isCleanupTime(triggerContext.window, timer.getTimestamp())) {
            windowState.setCurrentNamespace(triggerContext.window);
            if (randomInt == 1) {
                LOG.info("ClarkTest: Window State namespace before cleaning: " + triggerContext.window + " and key " + triggerContext.key);
                LOG.info("ClarkTest: Window state value before clear is null ? " + (windowState.get() == null));
            }
            clearAllState(triggerContext.window, windowState, mergingWindows);
            if (randomInt == 1) {
                LOG.info("ClarkTest: Window state value after clear is null ? " + (windowState.get() == null));
            }

        }

        if (mergingWindows != null) {
            // need to make sure to update the merging state in state
            mergingWindows.persist();
        }
    }
  • 发现每次在emitWindowContents之前window state的结果都不为null,但是在clean up之前,window state的结果已经变成了null。说明在这两段逻辑之间出了什么问题。通过将key打印出来,发现前后key有所变化,所以最后确定是用户代码的process function中改变了keyby的key的值导致窗口状态无法清理
  • 最后总结就是在keyby的时候key一定要是不变量,不然有可能导致状态无法清理。还有就是在分布式系统中,大量使用不变量是规避风险的最佳途径之一。

推荐阅读更多精彩内容