[KYLIN-4153] Failed to read big resource /dict/xxxx at "Build Dimension Dictionary" Step

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: v2.6.0
• Fix Version/s: v3.0.0-beta, v2.6.4
• Component/s: Metadata
• Assignee: Xiaoxiang Yu
• Reporter: Xiaoxiang Yu
• Created: 30/Aug/19 11:25
• Updated: 13/Oct/19 00:37
• Resolved: 13/Oct/19 00:34

Description

At the version of Kylin 2.6.0, kylin team has introduce an important refactor of Kylin's Metadata Store, which add a lot of enhancement such as upload/download metadata concurrently, store metadata with JDBC etc. Please refer to https://issues.apache.org/jira/browse/KYLIN-3671 for detail.

When kylin want to save a big resource(such as dict or snapshot) into metadata store, it won't store it into metadata store(HBase or RDBMS) directly. Instead, kylin will first save it into HDFS(Step 1), and then write a empty byte array as marker into metadata store(Step 2) . If first action succeed and second action failed, a rollback method will be called to revert modification for HDFS files. We could regard it as a complete and atomic transaction.

Here is part of the source code added in KYLIN-3671. Check it at https://github.com/apache/kylin/blob/8737bc1f555a2789a67462c8f8420b6ab3be97ce/core-common/src/main/java/org/apache/kylin/common/persistence/PushdownResourceStore.java#L58 .

final void putBigResource(String resPath, ContentWriter content, long newTS) throws IOException {
    // pushdown the big resource to DFS file
    RollbackablePushdown pushdown = writePushdown(resPath, content); // Step 1: write big resource into HDFS

    try {
        // write a marker in resource store, to indicate the resource is now available
        logger.debug("Writing marker for big resource {}", resPath);
        putResourceWithRetry(resPath, ContentWriter.create(BytesUtil.EMPTY_BYTE_ARRAY), newTS); // Step 2: write marker into HBase/RDBMS

    } catch (Throwable ex) {
        pushdown.rollback();
        throw ex;
    } finally {
        pushdown.close();
    }
}

But in some case, both step 1 and step 2 succeed but an exception still throwed in step 2, the rollback won't clear marker written in Step 2, which break the atomicity of this put action, thus cause the FileNotFoundException when Kylin want to read that dict later.

Here is part of reporter's kylin.log of incomplete rollback action.

 2019-08-29 05:13:51,237 INFO  [Scheduler 169045403 Job ca4a4a08-54e2-b922-70bb-2aa2bf58709f-492] dict.DictionaryManager:388 : Saving dictionary at /dict/KYLIN_VIEW.USER_SECRET_TABLE/COUNTRY/66292068-e8eb-975a-3e44-b56c933c14cc.dict
2019-08-29 05:13:51,238 DEBUG [Scheduler 169045403 Job ca4a4a08-54e2-b922-70bb-2aa2bf58709f-492] persistence.HDFSResourceStore:98 : Writing pushdown file /kylin/kylin_metadata/resources/dict/KYLIN_VIEW.USER_SECRET_TABLE/COUNTRY/66292068-e8eb-975a-3e44-b56c933c14cc.dict.temp.-1798610090
2019-08-29 05:13:51,256 DEBUG [Scheduler 169045403 Job ca4a4a08-54e2-b922-70bb-2aa2bf58709f-492] persistence.HDFSResourceStore:117 : Move /kylin/kylin_metadata/resources/dict/KYLIN_VIEW.USER_SECRET_TABLE/COUNTRY/66292068-e8eb-975a-3e44-b56c933c14cc.dict.temp.-1798610090 to /kylin/kylin_metadata/resources/dict/KYLIN_VIEW.USER_SECRET_TABLE/COUNTRY/66292068-e8eb-975a-3e44-b56c933c14cc.dict
2019-08-29 05:13:51,258 DEBUG [Scheduler 169045403 Job ca4a4a08-54e2-b922-70bb-2aa2bf58709f-492] persistence.HDFSResourceStore:65 : Writing marker for big resource /dict/KYLIN_VIEW.USER_SECRET_TABLE/COUNTRY/66292068-e8eb-975a-3e44-b56c933c14cc.dict

2019-08-29 05:13:56,263 WARN  [hconnection-0x56f3258e-shared--pool10944-t54867] client.AsyncProcess:1263 : #10545, table=kylin_metadata, attempt=1/1 failed=1ops, last exception: java.io.IOException: Call to tx-dn41.data/10.14.243.51:60020 failed on local exception: org.apache.hadoop.hbase.ipc.CallTimeoutException: Call id=2662317, waitTime=5001, operationTimeout=5000 expired. on tx-dn41.data,60020,1565943919204, tracking started Thu Aug 29 05:13:51 GMT+08:00 2019; not retrying 1 - final failure
2019-08-29 05:13:56,266 ERROR [Scheduler 169045403 Job ca4a4a08-54e2-b922-70bb-2aa2bf58709f-492] persistence.HDFSResourceStore:134 : Rollback /kylin/kylin_metadata/resources/dict/KYLIN_VIEW.USER_SECRET_TABLE/COUNTRY/66292068-e8eb-975a-3e44-b56c933c14cc.dict from <empty>
2019-08-29 05:13:56,274 ERROR [Scheduler 169045403 Job ca4a4a08-54e2-b922-70bb-2aa2bf58709f-492] common.HadoopShellExecutable:65 : error execute HadoopShellExecutable{id=ca4a4a08-54e2-b922-70bb-2aa2bf58709f-03, name=Build Dimension Dictionary, state=RUNNING}
2019-08-29 05:13:56,274 INFO  [Scheduler 169045403 Job ca4a4a08-54e2-b922-70bb-2aa2bf58709f-492] execution.AbstractExecutable:162 : Retry 1

Here is part of reporter's kylin.log of reading a non-exist dict in HDFS in "Build Dimension Dictionary" Step.

2019-08-29 14:54:59,602 INFO  [Scheduler 343338459 Job af4b847d-afa6-3729-4c19-03a5db08447b-498] steps.CreateDictionaryJob:110 : DictionaryProvider read dict from file: hdfs://CDH-cluster-main/kylin/kylin_metadata/kylin-af4b847d-afa6-3729-4c19-03a5db08447b/209_new_device/fact_distinct_columns/USER_SECRET_TABLE.COUNTRY/COUNTRY.rldict-r-00004
2019-08-29 14:54:59,602 DEBUG [Scheduler 343338459 Job af4b847d-afa6-3729-4c19-03a5db08447b-498] cli.DictionaryGeneratorCLI:73 : Dict for 'COUNTRY' has already been built, save it
2019-08-29 14:54:59,720 ERROR [Scheduler 343338459 Job af4b847d-afa6-3729-4c19-03a5db08447b-498] persistence.ResourceStore:233 : Error reading resource /dict/KYLIN_VIEW.USER_SECRET_TABLE/COUNTRY/66292068-e8eb-975a-3e44-b56c933c14cc.dict
java.io.IOException: Failed to read big resource /dict/KYLIN_VIEW.USER_SECRET_TABLE/COUNTRY/66292068-e8eb-975a-3e44-b56c933c14cc.dict
       at org.apache.kylin.common.persistence.PushdownResourceStore.openPushdown(PushdownResourceStore.java:176)
       at org.apache.kylin.storage.hbase.HBaseResourceStore.getInputStream(HBaseResourceStore.java:256)
       at org.apache.kylin.storage.hbase.HBaseResourceStore.rawResource(HBaseResourceStore.java:226)
       at org.apache.kylin.storage.hbase.HBaseResourceStore.access$000(HBaseResourceStore.java:64)
       at org.apache.kylin.storage.hbase.HBaseResourceStore$1.visit(HBaseResourceStore.java:159)
       at org.apache.kylin.storage.hbase.HBaseResourceStore.visitFolder(HBaseResourceStore.java:204)
       at org.apache.kylin.storage.hbase.HBaseResourceStore.visitFolderImpl(HBaseResourceStore.java:152)
       at org.apache.kylin.common.persistence.ResourceStore.visitFolderInner(ResourceStore.java:689)
       at org.apache.kylin.common.persistence.ResourceStore.visitFolderAndContent(ResourceStore.java:675)
       at org.apache.kylin.common.persistence.ResourceStore$2.call(ResourceStore.java:224)
       at org.apache.kylin.common.persistence.ResourceStore$2.call(ResourceStore.java:220)
       at org.apache.kylin.common.persistence.ExponentialBackoffRetry.doWithRetry(ExponentialBackoffRetry.java:52)
       at org.apache.kylin.common.persistence.ResourceStore.getAllResources(ResourceStore.java:220)
       at org.apache.kylin.common.persistence.ResourceStore.getAllResources(ResourceStore.java:209)
       at org.apache.kylin.dict.DictionaryManager.checkDupByInfo(DictionaryManager.java:334)
       at org.apache.kylin.dict.DictionaryManager.saveDictionary(DictionaryManager.java:314)
       at org.apache.kylin.cube.CubeManager$DictionaryAssist.saveDictionary(CubeManager.java:1127)
       at org.apache.kylin.cube.CubeManager.saveDictionary(CubeManager.java:1089)
       at org.apache.kylin.cube.cli.DictionaryGeneratorCLI.processSegment(DictionaryGeneratorCLI.java:74)
       at org.apache.kylin.cube.cli.DictionaryGeneratorCLI.processSegment(DictionaryGeneratorCLI.java:55)
       at org.apache.kylin.engine.mr.steps.CreateDictionaryJob.run(CreateDictionaryJob.java:73)
       at org.apache.kylin.engine.mr.MRUtil.runMRJob(MRUtil.java:93)
       at org.apache.kylin.engine.mr.common.HadoopShellExecutable.doWork(HadoopShellExecutable.java:63)
       at org.apache.kylin.job.execution.AbstractExecutable.execute(AbstractExecutable.java:167)
       at org.apache.kylin.job.execution.DefaultChainedExecutable.doWork(DefaultChainedExecutable.java:71)
       at org.apache.kylin.job.execution.AbstractExecutable.execute(AbstractExecutable.java:167)
       at org.apache.kylin.job.impl.threadpool.DefaultScheduler$JobRunner.run(DefaultScheduler.java:114)
       at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1142)
       at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:617)
       at java.lang.Thread.run(Thread.java:745)
Caused by: java.io.FileNotFoundException: /kylin/kylin_metadata/resources/dict/KYLIN_VIEW.USER_SECRET_TABLE/COUNTRY/66292068-e8eb-975a-3e44-b56c933c14cc.dict  (FS: DFS[DFSClient[clientName=DFSClient_NONMAPREDUCE_-784809092_27, ugi=kylin (auth:SIMPLE)]])
       at org.apache.kylin.common.persistence.PushdownResourceStore.openPushdown(PushdownResourceStore.java:173)
       ... 29 more

This often happen in Build Step 4: Build Dimension Dictionary. And this incomplete metadata entry will cause same failure(FileNotFoundException) of ALL following cube rebuild job.

As far as I can see, my workaround should be delete that marker. Since this is a broken metadata entry, deletion won't make damage. After the deletion, following rebuilt job will succeed.

This is some related report mail :
1. http://apache-kylin.74782.x6.nabble.com/How-to-repair-the-cube-that-it-lost-someone-dictionary-td12989.html
2. http://mail-archives.apache.org/mod_mbox/kylin-user/201908.mbox/[e-mail]%3e

Issue Links

• links to: GitHub Pull Request #818