[
https://issues.apache.org/jira/browse/HDFS-15240?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17158991#comment-17158991
]
Hongbing Wang commented on HDFS-15240:
--------------------------------------
[~marvelrock] We have the same problem. At the same time as this problem, there
are frequent FullGCs (every few seconds). Dump and MAT found that there are
lots of ecWorker objects that almost fill the entire heap.
!image-2020-07-16-15-56-38-608.png|width=722,height=591!
look forward this patch into trunk.
> Erasure Coding: dirty buffer causes reconstruction block error
> --------------------------------------------------------------
>
> Key: HDFS-15240
> URL: https://issues.apache.org/jira/browse/HDFS-15240
> Project: Hadoop HDFS
> Issue Type: Bug
> Components: datanode, erasure-coding
> Reporter: HuangTao
> Assignee: HuangTao
> Priority: Major
> Fix For: 3.4.0
>
> Attachments: HDFS-15240.001.patch, HDFS-15240.002.patch,
> HDFS-15240.003.patch, HDFS-15240.004.patch, HDFS-15240.005.patch,
> image-2020-07-16-15-56-38-608.png
>
>
> When read some lzo files we found some blocks were broken.
> I read back all internal blocks(b0-b8) of the block group(RS-6-3-1024k) from
> DN directly, and choose 6(b0-b5) blocks to decode other 3(b6', b7', b8')
> blocks. And find the longest common sequenece(LCS) between b6'(decoded) and
> b6(read from DN)(b7'/b7 and b8'/b8).
> After selecting 6 blocks of the block group in combinations one time and
> iterating through all cases, I find one case that the length of LCS is the
> block length - 64KB, 64KB is just the length of ByteBuffer used by
> StripedBlockReader. So the corrupt reconstruction block is made by a dirty
> buffer.
> The following log snippet(only show 2 of 28 cases) is my check program
> output. In my case, I known the 3th block is corrupt, so need other 5 blocks
> to decode another 3 blocks, then find the 1th block's LCS substring is block
> length - 64kb.
> It means (0,1,2,4,5,6)th blocks were used to reconstruct 3th block, and the
> dirty buffer was used before read the 1th block.
> Must be noted that StripedBlockReader read from the offset 0 of the 1th block
> after used the dirty buffer.
> {code:java}
> decode from [0, 2, 3, 4, 5, 7] -> [1, 6, 8]
> Check Block(1) first 131072 bytes longest common substring length 4
> Check Block(6) first 131072 bytes longest common substring length 4
> Check Block(8) first 131072 bytes longest common substring length 4
> decode from [0, 2, 3, 4, 5, 6] -> [1, 7, 8]
> Check Block(1) first 131072 bytes longest common substring length 65536
> CHECK AGAIN: Block(1) all 27262976 bytes longest common substring length
> 27197440 # this one
> Check Block(7) first 131072 bytes longest common substring length 4
> Check Block(8) first 131072 bytes longest common substring length 4{code}
> Now I know the dirty buffer causes reconstruction block error, but how does
> the dirty buffer come about?
> After digging into the code and DN log, I found this following DN log is the
> root reason.
> {code:java}
> [INFO] [stripedRead-1017] : Interrupted while waiting for IO on channel
> java.nio.channels.SocketChannel[connected local=/xxxxxxxx:52586
> remote=/xxxxxxxx:50010]. 180000 millis timeout left.
> [WARN] [StripedBlockReconstruction-199] : Failed to reconstruct striped
> block: BP-714356632-xxxxxxxx-1519726836856:blk_-YYYYYYYYYYYYYY_3472979393
> java.lang.NullPointerException
> at
> org.apache.hadoop.hdfs.util.StripedBlockUtil.getNextCompletedStripedRead(StripedBlockUtil.java:314)
> at
> org.apache.hadoop.hdfs.server.datanode.erasurecode.StripedReader.doReadMinimumSources(StripedReader.java:308)
> at
> org.apache.hadoop.hdfs.server.datanode.erasurecode.StripedReader.readMinimumSources(StripedReader.java:269)
> at
> org.apache.hadoop.hdfs.server.datanode.erasurecode.StripedBlockReconstructor.reconstruct(StripedBlockReconstructor.java:94)
> at
> org.apache.hadoop.hdfs.server.datanode.erasurecode.StripedBlockReconstructor.run(StripedBlockReconstructor.java:60)
> at
> java.base/java.util.concurrent.Executors$RunnableAdapter.call(Executors.java:515)
> at java.base/java.util.concurrent.FutureTask.run(FutureTask.java:264)
> at
> java.base/java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1128)
> at
> java.base/java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:628)
> at java.base/java.lang.Thread.run(Thread.java:834) {code}
> Reading from DN may timeout(hold by a future(F)) and output the INFO log, but
> the futures that contains the future(F) is cleared,
> {code:java}
> return new StripingChunkReadResult(futures.remove(future),
> StripingChunkReadResult.CANCELLED); {code}
> futures.remove(future) cause NPE. So the EC reconstruction is failed. In the
> finally phase, the code snippet in *getStripedReader().close()*
> {code:java}
> reconstructor.freeBuffer(reader.getReadBuffer());
> reader.freeReadBuffer();
> reader.closeBlockReader(); {code}
> free buffer firstly, but the StripedBlockReader still holds the buffer and
> write it.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: hdfs-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: hdfs-issues-h...@hadoop.apache.org
