hadoop - Apache Spark:挂起广播

Question

我很难在Yarn上调试我的Spark 1.6.2应用程序。它以客户端模式运行。本质上，它锁定时不会崩溃，并且锁定时控制台中的日志如下所示。

17/03/31 20:12:02 INFO BlockManagerInfo: Added broadcast_2_piece0 in memory on p3plcdsh007.prod.phx3.gdg:47579 (size: 26.7 KB, free: 511.1 MB)
    17/03/31 20:12:03 INFO BlockManagerInfo: Added broadcast_3_piece0 in memory on p3plcdsh011.prod.phx3.gdg:63228 (size: 5.4 KB, free: 511.1 MB)
    17/03/31 20:12:03 INFO BlockManagerInfo: Added broadcast_4_piece0 in memory on p3plcdsh015.prod.phx3.gdg:9377 (size: 5.4 KB, free: 511.1 MB)
    17/03/31 20:12:03 INFO BlockManagerInfo: Added broadcast_4_piece0 in memory on p3plcdsh015.prod.phx3.gdg:61897 (size: 5.4 KB, free: 511.1 MB)
    17/03/31 20:12:03 INFO BlockManagerInfo: Added broadcast_2_piece0 in memory on p3plcdsh002.prod.phx3.gdg:23170 (size: 26.7 KB, free: 511.1 MB)
    17/03/31 20:12:03 INFO BlockManagerInfo: Added broadcast_3_piece0 in memory on p3plcdsh016.prod.phx3.gdg:16649 (size: 5.4 KB, free: 511.1 MB)
    17/03/31 20:12:04 INFO BlockManagerInfo: Added broadcast_2_piece0 in memory on p3plcdsh003.prod.phx3.gdg:55147 (size: 26.7 KB, free: 511.1 MB)
    17/03/31 20:12:04 INFO BlockManagerInfo: Added broadcast_4_piece0 in memory on p3plcdsh008.prod.phx3.gdg:7619 (size: 5.4 KB, free: 511.1 MB)
    17/03/31 20:12:04 INFO BlockManagerInfo: Added broadcast_2_piece0 in memory on p3plcdsh003.prod.phx3.gdg:40830 (size: 26.7 KB, free: 511.1 MB)
    17/03/31 20:12:04 INFO BlockManagerInfo: Added broadcast_2_piece0 in memory on p3plcdsh011.prod.phx3.gdg:20056 (size: 26.7 KB, free: 511.1 MB)
    17/03/31 20:12:04 INFO BlockManagerInfo: Added broadcast_2_piece0 in memory on p3plcdsh008.prod.phx3.gdg:47385 (size: 26.7 KB, free: 511.1 MB)
    17/03/31 20:12:04 INFO BlockManagerInfo: Added broadcast_2_piece0 in memory on p3plcdsh003.prod.phx3.gdg:2063 (size: 26.7 KB, free: 511.1 MB)
    17/03/31 20:12:04 INFO BlockManagerInfo: Added broadcast_2_piece0 in memory on p3plcdsh011.prod.phx3.gdg:63228 (size: 26.7 KB, free: 511.1 MB)
    17/03/31 20:12:04 INFO BlockManagerInfo: Added broadcast_2_piece0 in memory on p3plcdsh008.prod.phx3.gdg:64036 (size: 26.7 KB, free: 511.1 MB)
    17/03/31 20:12:05 INFO BlockManagerInfo: Added broadcast_2_piece0 in memory on p3plcdsh016.prod.phx3.gdg:16649 (size: 26.7 KB, free: 511.1 MB)
    17/03/31 20:12:05 INFO BlockManagerInfo: Added broadcast_2_piece0 in memory on p3plcdsh013.prod.phx3.gdg:31979 (size: 26.7 KB, free: 511.1 MB)
    17/03/31 20:12:05 INFO BlockManagerInfo: Added broadcast_2_piece0 in memory on p3plcdsh013.prod.phx3.gdg:18407 (size: 26.7 KB, free: 511.1 MB)
    17/03/31 20:12:05 INFO BlockManagerInfo: Added broadcast_2_piece0 in memory on p3plcdsh004.prod.phx3.gdg:45536 (size: 26.7 KB, free: 511.1 MB)
    17/03/31 20:12:05 INFO BlockManagerInfo: Added broadcast_2_piece0 in memory on p3plcdsh008.prod.phx3.gdg:50826 (size: 26.7 KB, free: 511.1 MB)
    17/03/31 20:12:06 INFO BlockManagerInfo: Added broadcast_2_piece0 in memory on p3plcdsh015.prod.phx3.gdg:36247 (size: 26.7 KB, free: 511.1 MB)
    17/03/31 20:12:06 INFO BlockManagerInfo: Added broadcast_2_piece0 in memory on p3plcdsh015.prod.phx3.gdg:22848 (size: 26.7 KB, free: 511.1 MB)
    17/03/31 20:12:06 INFO BlockManagerInfo: Added broadcast_2_piece0 in memory on p3plcdsh015.prod.phx3.gdg:9377 (size: 26.7 KB, free: 511.1 MB)
    17/03/31 20:12:06 INFO BlockManagerInfo: Added broadcast_2_piece0 in memory on p3plcdsh015.prod.phx3.gdg:61897 (size: 26.7 KB, free: 511.1 MB)
    17/03/31 20:12:07 INFO BlockManagerInfo: Added broadcast_2_piece0 in memory on p3plcdsh008.prod.phx3.gdg:7619 (size: 26.7 KB, free: 511.1 MB)

在Spark UI中，锁定发生在 map 或过滤器功能处。

有没有人见过这种情况发生或知道如何调试情况？

看来这可能是由于内存问题或空间问题引起的，但没有明确的迹象表明是这种情况。我可以尝试增加内存，看看是否有帮助，但是有人可以调试吗？

谢谢

Answer 1

仅可序列化是不够的。问题可能出在很多方面:序列化机制(Java序列化不好； Kryo更好)，等等，机器内存，确保使用广播值而不是包装值，等等。

还有Spark配置spark.sql.autoBroadcastJoinThreshold:

“为执行联结时将广播到所有工作程序节点的表配置最大字节大小。通过将此值设置为-1，可以禁用广播。请注意，当前仅支持在Hive Metastore表中使用ANALYZE命令的统计信息表计算统计信息noscan已运行。”

默认值为10MB序列化。

最后，如果您删除了默认限制并且有足够的内存，那么您仍然希望大小小于最大的RDD / DataFrame，可以使用SizeEstimator进行检查:

import org.apache.spark.util.SizeEstimator._

logInfo(estimate(rdd))

最后，如果情况变得更糟，我将考虑在转换中从快如闪电的缓存数据存储区进行查找，而不是广播此文件。