Details
-
Bug
-
Status: Resolved
-
Normal
-
Resolution: Fixed
-
None
-
Code - Bug - Unclear Impact
-
Normal
-
Normal
-
Shadow Traffic
-
All
-
None
-
Description
tl;dr The internode TCP user timeout that provides more responsive detection of dead nodes for internode message will cause streaming to fail if system calls to fsync/fdatasync exceed the timeout (default 30s).
To workaround, explicitly set internode_tcp_user_timeout to longer than fsync/fdatasync, or to zero to revert to the operating system default.
Details:
While bootstrapping a replacement 4.0beta3 node in an existing cluster, bootstrap streaming repeatedly failed with the streaming follower logging
ERROR 2020-09-10T14:29:34,711 [NettyStreaming-Outbound-1.1.1.1.7000:1] org.apache.cassandra.streaming.StreamSession:693 - [Stream #7cb67c00-f3ac-11ea-b940-f7836f164528] Streaming error occurred on session with peer 1.1.1.1:7000 org.apache.cassandra.net.AsyncChannelOutputPlus$FlushException: The channel this output stream was writing to has been closed at org.apache.cassandra.net.AsyncChannelOutputPlus.propagateFailedFlush(AsyncChannelOutputPlus.java:200) at org.apache.cassandra.net.AsyncChannelOutputPlus.waitUntilFlushed(AsyncChannelOutputPlus.java:158) at org.apache.cassandra.net.AsyncChannelOutputPlus.waitForSpace(AsyncChannelOutputPlus.java:140) at org.apache.cassandra.net.AsyncChannelOutputPlus.beginFlush(AsyncChannelOutputPlus.java:97) at org.apache.cassandra.net.AsyncStreamingOutputPlus.lambda$writeToChannel$0(AsyncStreamingOutputPlus.java:142) at org.apache.cassandra.db.streaming.CassandraCompressedStreamWriter.lambda$write$0(CassandraCompressedStreamWriter.java:90) at org.apache.cassandra.net.AsyncStreamingOutputPlus.writeToChannel(AsyncStreamingOutputPlus.java:138) at org.apache.cassandra.db.streaming.CassandraCompressedStreamWriter.write(CassandraCompressedStreamWriter.java:89) at org.apache.cassandra.db.streaming.CassandraOutgoingFile.write(CassandraOutgoingFile.java:180) at org.apache.cassandra.streaming.messages.OutgoingStreamMessage.serialize(OutgoingStreamMessage.java:87) at org.apache.cassandra.streaming.messages.OutgoingStreamMessage$1.serialize(OutgoingStreamMessage.java:45) at org.apache.cassandra.streaming.messages.OutgoingStreamMessage$1.serialize(OutgoingStreamMessage.java:34) at org.apache.cassandra.streaming.messages.StreamMessage.serialize(StreamMessage.java:40) at org.apache.cassandra.streaming.async.NettyStreamingMessageSender$FileStreamTask.run(NettyStreamingMessageSender.java:347) at java.util.concurrent.Executors$RunnableAdapter.call(Executors.java:515) [?:?] at java.util.concurrent.FutureTask.run(FutureTask.java:264) [?:?] at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1128) [?:?] at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:628) [?:?] at io.netty.util.concurrent.FastThreadLocalRunnable.run(FastThreadLocalRunnable.java:30) [netty-all-4.1.50.Final.jar:4.1.50.Final] at java.lang.Thread.run(Thread.java:834) [?:?] Suppressed: java.nio.channels.ClosedChannelException at org.apache.cassandra.net.AsyncStreamingOutputPlus.doFlush(AsyncStreamingOutputPlus.java:78) at org.apache.cassandra.net.AsyncChannelOutputPlus.flush(AsyncChannelOutputPlus.java:229) at org.apache.cassandra.net.AsyncChannelOutputPlus.close(AsyncChannelOutputPlus.java:248) at org.apache.cassandra.streaming.async.NettyStreamingMessageSender$FileStreamTask.run(NettyStreamingMessageSender.java:348) at java.util.concurrent.Executors$RunnableAdapter.call(Executors.java:515) [?:?] at java.util.concurrent.FutureTask.run(FutureTask.java:264) [?:?] at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1128) [?:?] at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:628) [?:?] at io.netty.util.concurrent.FastThreadLocalRunnable.run(FastThreadLocalRunnable.java:30) [netty-all-4.1.50.Final.jar:4.1.50.Final] at java.lang.Thread.run(Thread.java:834) [?:?] Caused by: io.netty.channel.unix.Errors$NativeIoException: writeAddress(..) failed: Connection timed out
and the boostrapping node (the streaming initiator) logging (times are from two separate attempts, pattern was very similar each time, IP addresses doctored to protect the innocent)
ERROR 2020-09-11T09:45:54,720 [Stream-Deserializer-2.2.2.2:7000-93837c82] org.apache.cassandra.streaming.StreamSession:917 - [Stream #f921ea30-f44d-11ea-b7f1-356e3edd6247] Remote peer 2.2.2.2:7000 failed stream session.
DEBUG 2020-09-11T09:45:54,721 [Stream-Deserializer-2.2.2.2:7000-93837c82] org.apache.cassandra.streaming.StreamSession:510 - [Stream #f921ea30-f44d-11ea-b7f1-356e3edd6247] Will close attached channels {93837c82=[id: 0x93837c82, L:/1.1.1.1:56859 ! R:2.2.2.2/2.2.2.2:7000], a608109b=[id: 0xa608109b, L:/1.1.1.1:7000 - R:/2.2.2.2:49426]}
INFO 2020-09-11T09:45:54,722 [Stream-Deserializer-2.2.2.2:7000-93837c82] org.apache.cassandra.streaming.StreamResultFuture:192 - [Stream #f921ea30-f44d-11ea-b7f1-356e3edd6247] Session with 2.2.2.2:7000 is complete
DEBUG 2020-09-11T09:45:54,723 [Messaging-EventLoop-3-15] org.apache.cassandra.streaming.async.NettyStreamingMessageSender:553 - [Stream #f921ea30-f44d-11ea-b7f1-356e3edd6247] Closing stream connection channels on 2.2.2.2:7000
ERROR 2020-09-11T09:45:55,121 [Stream-Deserializer-2.2.2.2:7000-a608109b] org.apache.cassandra.streaming.StreamSession:693 - [Stream #f921ea30-f44d-11ea-b7f1-356e3edd6247] Streaming error occurred on session with peer 2.2.2.2:7000
org.apache.cassandra.streaming.StreamReceiveException: java.io.EOFException
at org.apache.cassandra.streaming.messages.IncomingStreamMessage$1.deserialize(IncomingStreamMessage.java:60)
at org.apache.cassandra.streaming.messages.IncomingStreamMessage$1.deserialize(IncomingStreamMessage.java:38)
at org.apache.cassandra.streaming.messages.StreamMessage.deserialize(StreamMessage.java:51)
at org.apache.cassandra.streaming.async.StreamingInboundHandler$StreamDeserializingTask.run(StreamingInboundHandler.java:172)
at io.netty.util.concurrent.FastThreadLocalRunnable.run(FastThreadLocalRunnable.java:30) [netty-all-4.1.50.Final.jar:4.1.50.Final]
at java.lang.Thread.run(Thread.java:834) [?:?]
Caused by: java.io.EOFException
at org.apache.cassandra.net.AsyncStreamingInputPlus.reBuffer(AsyncStreamingInputPlus.java:133)
at org.apache.cassandra.io.util.RebufferingInputStream.readByte(RebufferingInputStream.java:178)
at org.apache.cassandra.io.util.RebufferingInputStream.readPrimitiveSlowly(RebufferingInputStream.java:142)
at org.apache.cassandra.io.util.RebufferingInputStream.readLong(RebufferingInputStream.java:231)
at org.apache.cassandra.io.compress.CompressionMetadata$ChunkSerializer.deserialize(CompressionMetadata.java:538)
at org.apache.cassandra.io.compress.CompressionMetadata$ChunkSerializer.deserialize(CompressionMetadata.java:528)
at org.apache.cassandra.db.streaming.CompressionInfo$CompressionInfoSerializer.deserialize(CompressionInfo.java:88)
at org.apache.cassandra.db.streaming.CompressionInfo$CompressionInfoSerializer.deserialize(CompressionInfo.java:61)
at org.apache.cassandra.db.streaming.CassandraStreamHeader$CassandraStreamHeaderSerializer.deserialize(CassandraStreamHeader.java:230)
at org.apache.cassandra.db.streaming.CassandraStreamHeader$CassandraStreamHeaderSerializer.deserialize(CassandraStreamHeader.java:210)
at org.apache.cassandra.db.streaming.CassandraStreamHeader$CassandraStreamHeaderSerializer.deserialize(CassandraStreamHeader.java:178)
at org.apache.cassandra.db.streaming.CassandraIncomingFile.read(CassandraIncomingFile.java:69)
at org.apache.cassandra.streaming.messages.IncomingStreamMessage$1.deserialize(IncomingStreamMessage.java:53)
... 5 more
DEBUG 2020-09-11T09:45:55,123 [Stream-Deserializer-2.2.2.2:7000-93837c82] org.apache.cassandra.streaming.StreamSession:657 - [Stream #f921ea30-f44d-11ea-b7f1-356e3edd6247] Socket closed after session completed with state FAILED
The issue was repeatable across many attempts to bootstrap. Cluster was deployed on Linux hosts running a 4.1.x kernel.
Investigating the cause of the NativeIoException error message showed that a write(2) call return must have been ETIMEDOUT, which occurs when a TCP user timeout is set on the socket. In 4.0 connections now do that when EPOLL is available, with a default of 30s to detect failures faster.
Investigating the socket buffers around the time of crash showed the send buffer filling on the streaming follower and the receive buffer filling on the streaming initiator.
Capturing a heap dump and analyzing with MAT showed a thread in this state until the stream failed.
Object / Stack Frame |Name | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- io.netty.util.concurrent.FastThreadLocalThread @ 0x6916ea170 |Stream-Deserializer-2.2.2.2:7000-a608109b| | - at sun.nio.ch.FileDispatcherImpl.force0(Ljava/io/FileDescriptor;Z)I (Native Method)| | | - at sun.nio.ch.FileDispatcherImpl.force(Ljava/io/FileDescriptor;Z)I (FileDispatcherImpl.java:82)| | | - at sun.nio.ch.FileChannelImpl.force(Z)V (FileChannelImpl.java:461)| | | - at org.apache.cassandra.utils.SyncUtil.force(Ljava/nio/channels/FileChannel;Z)V (SyncUtil.java:172)| | | - at org.apache.cassandra.io.util.SequentialWriter.syncDataOnlyInternal()V (SequentialWriter.java:190)| | | - at org.apache.cassandra.io.util.SequentialWriter.syncInternal()V (SequentialWriter.java:206)| | | - at org.apache.cassandra.io.compress.CompressedSequentialWriter.access$100(Lorg/apache/cassandra/io/compress/CompressedSequentialWriter;)V (CompressedSequentialWriter.java:39)| | | - at org.apache.cassandra.io.compress.CompressedSequentialWriter$TransactionalProxy.doPrepare()V (CompressedSequentialWriter.java:353)| | | - at org.apache.cassandra.utils.concurrent.Transactional$AbstractTransactional.prepareToCommit()V (Transactional.java:168)| | | - at org.apache.cassandra.io.util.SequentialWriter.prepareToCommit()V (SequentialWriter.java:379)| | | - at org.apache.cassandra.io.sstable.format.big.BigTableWriter$TransactionalProxy.doPrepare()V (BigTableWriter.java:409)| | | - at org.apache.cassandra.utils.concurrent.Transactional$AbstractTransactional.prepareToCommit()V (Transactional.java:168)| | | - at org.apache.cassandra.utils.concurrent.Transactional$AbstractTransactional.finish()Ljava/lang/Object; (Transactional.java:179)| | | - at org.apache.cassandra.io.sstable.format.SSTableWriter.finish(Z)Lorg/apache/cassandra/io/sstable/format/SSTableReader; (SSTableWriter.java:266)| | | - at org.apache.cassandra.io.sstable.SimpleSSTableMultiWriter.finish(Z)Ljava/util/Collection; (SimpleSSTableMultiWriter.java:59)| | | - at org.apache.cassandra.io.sstable.format.RangeAwareSSTableWriter.finish(Z)Ljava/util/Collection; (RangeAwareSSTableWriter.java:135)| | | - at org.apache.cassandra.db.streaming.CassandraStreamReceiver.received(Lorg/apache/cassandra/streaming/IncomingStream;)V (CassandraStreamReceiver.java:108)| | | - at org.apache.cassandra.streaming.StreamReceiveTask.received(Lorg/apache/cassandra/streaming/IncomingStream;)V (StreamReceiveTask.java:91)| | | - at org.apache.cassandra.streaming.StreamSession.receive(Lorg/apache/cassandra/streaming/messages/IncomingStreamMessage;)V (StreamSession.java:848)| | | - at org.apache.cassandra.streaming.StreamSession.messageReceived(Lorg/apache/cassandra/streaming/messages/StreamMessage;)V (StreamSession.java:597)| | | - at org.apache.cassandra.streaming.async.StreamingInboundHandler$StreamDeserializingTask.run()V (StreamingInboundHandler.java:189)| | | - at io.netty.util.concurrent.FastThreadLocalRunnable.run()V (FastThreadLocalRunnable.java:30)| | | - at java.lang.Thread.run()V (Thread.java:834)| | '- Total: 23 entries | | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Looking at the Java sources, the force0 function calls fdatasync under Linux. Using perf trace record showed that some fdatasyncs were exceeding the socket user timeout threshold, explaining the failures.
This issue required significant effort to determine the root cause, so at least it would be good to improve the ability to diagnose this for an operator that may not have access to system call profiling.
Obviously fdatasync going away for 30s+ is a system tuning issue but I do not know of a way to guarantee an upper bound on fdatasync time, just make it more probable it completes.
One workaround is to increase the TCP user timeout from 30s to whatever the host needs to complete, or to reset to the system default with 0, however this impacts regular internode messaging as well as streaming, and it's desirable to discover dead nodes as early as possible.
Possible improvements:
- Separate the user timeout thresholds for internode messaging and streaming.
- Create a metric for FileChannel.force times by unique filesystem to surface that as a possible timeout.
- Introduce a log warning when FileChannel.force times exceed a threshold.
- Run FileChannel.force on a separate (unbounded) excecutor service and either continue to stream while force completes in the background with at least some backpressure on the stream, or throw on the initiator to improve diagnosis.
- Periodically call force by time and/or volume written on a background thread.
