[jira] [Commented] (NUTCH-2756) Segment Part problem with HDFS on distibuted mode

classic Classic list List threaded Threaded
1 message Options
Reply | Threaded
Open this post in threaded view
|

[jira] [Commented] (NUTCH-2756) Segment Part problem with HDFS on distibuted mode

Sebastian Nagel (Jira)

    [ https://issues.apache.org/jira/browse/NUTCH-2756?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16991467#comment-16991467 ]

Sebastian Nagel commented on NUTCH-2756:
----------------------------------------

Hi [~lucasp], if always the same partition is affected, this makes a bug in Nutch more likely. Parser (and fetcher) output is partitioned using the default [HashPartitioner|https://hadoop.apache.org/docs/r3.2.1/api/org/apache/hadoop/mapreduce/lib/partition/HashPartitioner.html] on the URL which is used as key. This means the same URL and the data behind it is written to the same partition if the number of partitions hasn't changed.
{quote}Do you think that if I increase the value of the replication it will reduce the chance that this kind of bug happen?
{quote}
Only if it's a HDFS (configuration) or hardware issue. To get this confirmed or sorted out, the following information would be more important:
 - the segment directory listing just after the job has completed - could you add a {{hadoop fs -ls -R .../segments/<segment_name>}} before and after the parser job?
 - the task log of the fetcher/parser task writing partition 4
 - is the state reproducible by running the parser job on the same segment again? Remove the subdirectories crawl_parse, parse_data and parse_text and run the parser job again.
 - you could also try to use less or more partitions (--num-tasks if bin/crawl is used or property mapreduce.job.reduces)

The subdir crawl_parse contains all the data used to update CrawlDb: outlinks, scores, signatures, meta data. No glue what could go wrong with this file. If it's a bug in Nutch it's a very rare corner case, not seen before when processing/parsing billions of pages.


> Segment Part problem with HDFS on distibuted mode
> -------------------------------------------------
>
>                 Key: NUTCH-2756
>                 URL: https://issues.apache.org/jira/browse/NUTCH-2756
>             Project: Nutch
>          Issue Type: Bug
>          Components: parser
>    Affects Versions: 1.15
>            Reporter: Lucas Pauchard
>            Priority: Major
>         Attachments: 0_byte_file_screenshot.PNG, hadoop-env.sh, hdfs-site.xml, mapred-site.xml, yarn-env.sh, yarn-site.xml
>
>
> During the parsing, it happens sometimes that parts of the data on the HDFS is missing after the parsing.
> When I take a look at our HDFS, I've got this file with 0 bytes (see attachments).
> After that the CrawlDB complains about this specific (corrupted?) part:
> {panel:title=log_crawl}
> 2019-12-04 22:25:57,454 INFO mapreduce.Job: Task Id : attempt_1575479127636_0047_m_000017_2, Status : FAILED
> Error: java.io.EOFException: hdfs://jobmaster:9000/user/hadoop/crawlmultiokhttp/segment/20191204221308/crawl_parse/part-r-00004 not a SequenceFile
>         at org.apache.hadoop.io.SequenceFile$Reader.init(SequenceFile.java:1964)
>         at org.apache.hadoop.io.SequenceFile$Reader.initialize(SequenceFile.java:1923)
>         at org.apache.hadoop.io.SequenceFile$Reader.<init>(SequenceFile.java:1872)
>         at org.apache.hadoop.io.SequenceFile$Reader.<init>(SequenceFile.java:1886)
>         at org.apache.hadoop.mapreduce.lib.input.SequenceFileRecordReader.initialize(SequenceFileRecordReader.java:54)
>         at org.apache.hadoop.mapred.MapTask$NewTrackingRecordReader.initialize(MapTask.java:560)
>         at org.apache.hadoop.mapred.MapTask.runNewMapper(MapTask.java:798)
>         at org.apache.hadoop.mapred.MapTask.run(MapTask.java:347)
>         at org.apache.hadoop.mapred.YarnChild$2.run(YarnChild.java:174)
>         at java.security.AccessController.doPrivileged(Native Method)
>         at javax.security.auth.Subject.doAs(Subject.java:422)
>         at org.apache.hadoop.security.UserGroupInformation.doAs(UserGroupInformation.java:1729)
>         at org.apache.hadoop.mapred.YarnChild.main(YarnChild.java:168)~
> {panel}
> When I check the namenode logs, I don't see any error during the writing of the segment part but one hour later, I've got the following log:
> {panel:title=log_namenode}
> 2019-12-04 23:23:13,750 INFO org.apache.hadoop.hdfs.server.namenode.FSNamesystem: Recovering [Lease.  Holder: DFSClient_attempt_1575479127636_0046_r_000004_1_1307945884_1, pending creates: 2], src=/user/hadoop/crawlmultiokhttp/segment/20191204221308/parse_data/part-r-00004/index
> 2019-12-04 23:23:13,750 WARN org.apache.hadoop.hdfs.StateChange: BLOCK* internalReleaseLease: All existing blocks are COMPLETE, lease removed, file /user/hadoop/crawlmultiokhttp/segment/20191204221308/parse_data/part-r-00004/index closed.
> 2019-12-04 23:23:13,750 INFO org.apache.hadoop.hdfs.server.namenode.FSNamesystem: Recovering [Lease.  Holder: DFSClient_attempt_1575479127636_0046_r_000004_1_1307945884_1, pending creates: 1], src=/user/hadoop/crawlmultiokhttp/segment/20191204221308/crawl_parse/part-r-00004
> 2019-12-04 23:23:13,750 WARN org.apache.hadoop.hdfs.StateChange: BLOCK* internalReleaseLease: All existing blocks are COMPLETE, lease removed, file /user/hadoop/crawlmultiokhttp/segment/20191204221308/crawl_parse/part-r-00004 closed.
> {panel}
> This issue is hard to reproduce and I can't figure out what are the preconditions. It seems that it just happens randomly.
> Maybe the problem is coming from a bad management when we close the file.



--
This message was sent by Atlassian Jira
(v8.3.4#803005)