Details
-
Bug
-
Status: Resolved
-
Blocker
-
Resolution: Fixed
-
2.3.0, 2.3.1
Description
After Spark 2.3, we observed lots of errors like the following in some of our production job
18/06/15 20:59:42 ERROR TransportRequestHandler: Error sending result ChunkFetchSuccess{streamChunkId=StreamChunkId{streamId=91672904003, chunkIndex=0}, buffer=org.apache.spark.storage.BlockManagerManagedBuffer@783a9324} to /172.22.18.7:60865; closing connection
java.io.IOException: Broken pipe
at sun.nio.ch.FileDispatcherImpl.write0(Native Method)
at sun.nio.ch.SocketDispatcher.write(SocketDispatcher.java:47)
at sun.nio.ch.IOUtil.writeFromNativeBuffer(IOUtil.java:93)
at sun.nio.ch.IOUtil.write(IOUtil.java:65)
at sun.nio.ch.SocketChannelImpl.write(SocketChannelImpl.java:471)
at org.apache.spark.network.protocol.MessageWithHeader.writeNioBuffer(MessageWithHeader.java:156)
at org.apache.spark.network.protocol.MessageWithHeader.copyByteBuf(MessageWithHeader.java:142)
at org.apache.spark.network.protocol.MessageWithHeader.transferTo(MessageWithHeader.java:123)
at io.netty.channel.socket.nio.NioSocketChannel.doWriteFileRegion(NioSocketChannel.java:355)
at io.netty.channel.nio.AbstractNioByteChannel.doWrite(AbstractNioByteChannel.java:224)
at io.netty.channel.socket.nio.NioSocketChannel.doWrite(NioSocketChannel.java:382)
at io.netty.channel.AbstractChannel$AbstractUnsafe.flush0(AbstractChannel.java:934)
at io.netty.channel.nio.AbstractNioChannel$AbstractNioUnsafe.flush0(AbstractNioChannel.java:362)
at io.netty.channel.AbstractChannel$AbstractUnsafe.flush(AbstractChannel.java:901)
at io.netty.channel.DefaultChannelPipeline$HeadContext.flush(DefaultChannelPipeline.java:1321)
at io.netty.channel.AbstractChannelHandlerContext.invokeFlush0(AbstractChannelHandlerContext.java:776)
at io.netty.channel.AbstractChannelHandlerContext.invokeFlush(AbstractChannelHandlerContext.java:768)
at io.netty.channel.AbstractChannelHandlerContext.flush(AbstractChannelHandlerContext.java:749)
at io.netty.channel.ChannelOutboundHandlerAdapter.flush(ChannelOutboundHandlerAdapter.java:115)
at io.netty.channel.AbstractChannelHandlerContext.invokeFlush0(AbstractChannelHandlerContext.java:776)
at io.netty.channel.AbstractChannelHandlerContext.invokeFlush(AbstractChannelHandlerContext.java:768)
at io.netty.channel.AbstractChannelHandlerContext.flush(AbstractChannelHandlerContext.java:749)
at io.netty.channel.ChannelDuplexHandler.flush(ChannelDuplexHandler.java:117)
at io.netty.channel.AbstractChannelHandlerContext.invokeFlush0(AbstractChannelHandlerContext.java:776)
at io.netty.channel.AbstractChannelHandlerContext.invokeFlush(AbstractChannelHandlerContext.java:768)
at io.netty.channel.AbstractChannelHandlerContext.flush(AbstractChannelHandlerContext.java:749)
at io.netty.channel.DefaultChannelPipeline.flush(DefaultChannelPipeline.java:983)
at io.netty.channel.AbstractChannel.flush(AbstractChannel.java:248)
at io.netty.channel.nio.AbstractNioByteChannel$1.run(AbstractNioByteChannel.java:284)
at io.netty.util.concurrent.AbstractEventExecutor.safeExecute(AbstractEventExecutor.java:163)
at io.netty.util.concurrent.SingleThreadEventExecutor.runAllTasks(SingleThreadEventExecutor.java:403)
at io.netty.channel.nio.NioEventLoop.run(NioEventLoop.java:463)
at io.netty.util.concurrent.SingleThreadEventExecutor$5.run(SingleThreadEventExecutor.java:858)
at io.netty.util.concurrent.DefaultThreadFactory$DefaultRunnableDecorator.run(DefaultThreadFactory.java:138)
Here is a small reproducible for a small cluster of 2 executors (say host-1 and host-2) each with 8 cores. Here, the memory of driver and executors are not an import factor here as long as it is big enough, say 20G.
val n = 100000000 val df0 = sc.parallelize(1 to n).toDF val df = df0.withColumn("x0", rand()).withColumn("x0", rand() ).withColumn("x1", rand() ).withColumn("x2", rand() ).withColumn("x3", rand() ).withColumn("x4", rand() ).withColumn("x5", rand() ).withColumn("x6", rand() ).withColumn("x7", rand() ).withColumn("x8", rand() ).withColumn("x9", rand()) df.cache; df.count (1 to 10).toArray.par.map { i => println(i); df.groupBy("x1").agg(count("value")).show() }
In the above example, we generate a random DataFrame of size around 7G; cache it and then perform a parallel DataFrame operations by using `array.par.map`. Because of the parallel computation, with high possibility, some task could be scheduled to a host-2 where it needs to read the cache block data from host-1. This follows the code path of https://github.com/apache/spark/blob/master/core/src/main/scala/org/apache/spark/storage/BlockManager.scala#L691 and then tries to transfer a big block (~ 500MB) of cache block from host-1 to host-2. Often, this big transfer makes the cluster suffer time out issue (it will retry 3 times, each with 120s timeout, and then do recompute to put the cache block into the local MemoryStore).
We couldn't to reproduce the same issue in Spark 2.2.1. From the log of Spark 2.2.1, we found that
18/06/16 17:23:47 DEBUG BlockManager: Getting local block rdd_3_0 18/06/16 17:23:47 TRACE BlockInfoManager: Task 0 trying to acquire read lock for rdd_3_0 18/06/16 17:23:47 DEBUG BlockManager: Block rdd_3_0 was not found 18/06/16 17:23:47 DEBUG BlockManager: Getting remote block rdd_3_0 18/06/16 17:23:47 DEBUG BlockManager: Block rdd_3_0 not found 18/06/16 17:23:47 TRACE BlockInfoManager: Task 0 trying to put rdd_3_0 18/06/16 17:23:47 TRACE BlockInfoManager: Task 0 trying to acquire read lock for rdd_3_0 18/06/16 17:23:47 TRACE BlockInfoManager: Task 0 trying to acquire write lock for rdd_3_0 18/06/16 17:23:47 TRACE BlockInfoManager: Task 0 acquired write lock for rdd_3_0 18/06/16 17:23:58 INFO MemoryStore: Block rdd_3_0 stored as values in memory (estimated size 538.2 MB, free 11.1 GB)
That is, when a task is scheduled to a host-2 where it needs to read the cache block rdd_3_0 data from host-1, the endpoint of `master.getLocations(..)` ( see https://github.com/apache/spark/blob/v2.2.1/core/src/main/scala/org/apache/spark/storage/BlockManager.scala#L622) reports a remote cache block is not found and triggered the recompute.
I believe this behavior change is introduced by this change set https://github.com/apache/spark/commit/e1960c3d6f380b0dfbba6ee5d8ac6da4bc29a698#diff-2b643ea78c1add0381754b1f47eec132
We have two questions here
- what is the right behavior, should we re-compute or should we transfer block from remote?
- if we should transfer from remote, why the performance is so bad for cache block?
