[HIVE-24322] In case of direct insert, the attempt ID has to be checked when reading the manifest files

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 4.0.0
• Fix Version/s: 4.0.0-alpha-1
• Labels: pull-request-available
• Assignee: Marta Kuczora
• Reporter: Marta Kuczora
• Created: 28/Oct/20 17:09
• Updated: 17/Nov/22 13:14
• Resolved: 15/Dec/20 13:01

Description

In IMPALA-10247 there was an exception from Hive when tyring to load the data:

2020-10-13T16:50:53,424 ERROR [HiveServer2-Background-Pool: Thread-23832] exec.Task: Job Commit failed with exception 'org.apache.hadoop.hive.ql.metadata.HiveException(java.io.EOFException)'
org.apache.hadoop.hive.ql.metadata.HiveException: java.io.EOFException
 at org.apache.hadoop.hive.ql.exec.FileSinkOperator.jobCloseOp(FileSinkOperator.java:1468)
 at org.apache.hadoop.hive.ql.exec.Operator.jobClose(Operator.java:798)
 at org.apache.hadoop.hive.ql.exec.Operator.jobClose(Operator.java:803)
 at org.apache.hadoop.hive.ql.exec.Operator.jobClose(Operator.java:803)
 at org.apache.hadoop.hive.ql.exec.tez.TezTask.close(TezTask.java:627)
 at org.apache.hadoop.hive.ql.exec.tez.TezTask.execute(TezTask.java:342)
 at org.apache.hadoop.hive.ql.exec.Task.executeTask(Task.java:213)
 at org.apache.hadoop.hive.ql.exec.TaskRunner.runSequential(TaskRunner.java:105)
 at org.apache.hadoop.hive.ql.Executor.launchTask(Executor.java:357)
 at org.apache.hadoop.hive.ql.Executor.launchTasks(Executor.java:330)
 at org.apache.hadoop.hive.ql.Executor.runTasks(Executor.java:246)
 at org.apache.hadoop.hive.ql.Executor.execute(Executor.java:109)
 at org.apache.hadoop.hive.ql.Driver.runInternal(Driver.java:721)
 at org.apache.hadoop.hive.ql.Driver.run(Driver.java:488)
 at org.apache.hadoop.hive.ql.Driver.run(Driver.java:482)
 at org.apache.hadoop.hive.ql.reexec.ReExecDriver.run(ReExecDriver.java:166)
 at org.apache.hive.service.cli.operation.SQLOperation.runQuery(SQLOperation.java:225)
 at org.apache.hive.service.cli.operation.SQLOperation.access$700(SQLOperation.java:87)
 at org.apache.hive.service.cli.operation.SQLOperation$BackgroundWork$1.run(SQLOperation.java:322)
 at java.security.AccessController.doPrivileged(Native Method)
 at javax.security.auth.Subject.doAs(Subject.java:422)
 at org.apache.hadoop.security.UserGroupInformation.doAs(UserGroupInformation.java:1876)
 at org.apache.hive.service.cli.operation.SQLOperation$BackgroundWork.run(SQLOperation.java:340)
 at java.util.concurrent.Executors$RunnableAdapter.call(Executors.java:511)
 at java.util.concurrent.FutureTask.run(FutureTask.java:266)
 at java.util.concurrent.Executors$RunnableAdapter.call(Executors.java:511)
 at java.util.concurrent.FutureTask.run(FutureTask.java:266)
 at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1149)
 at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:624)
 at java.lang.Thread.run(Thread.java:748)
Caused by: java.io.EOFException
 at java.io.DataInputStream.readInt(DataInputStream.java:392)
 at org.apache.hadoop.hive.ql.exec.Utilities.handleDirectInsertTableFinalPath(Utilities.java:4587)
 at org.apache.hadoop.hive.ql.exec.FileSinkOperator.jobCloseOp(FileSinkOperator.java:1462)
 ... 29 more

The reason of the exception was that Hive was trying to read an empty manifest file. Manifest files are used in case of direct insert to determine which files needs to be kept and which one needs to be cleaned up. They are created by the tasks and they use the task attempt Id as postfix. In this particular test what happened is that one of the container ran out of memory so Tez decided to kill it right after the manifest file got created but before the paths got written into the manifest file. This was the manifest file for the task attempt 0. Then Tez assigned a new container to the task, so a new attempt was made with attemptId=1. This one was successful, and wrote the manifest file correctly. But Hive didn't know about this, since this out of memory issue got handled by Tez under the hood, so there was no exception in Hive, therefore no clean-up in the manifest folder. And when Hive is reading the manifest files, it just reads every file from the defined folder, so it tried to read the manifest files for attempt 0 and 1 as well.
If there are multiple manifest files with the same name but different attemptId, Hive should only read the one with the biggest attempt Id.

Issue Links

• causes: IMPALA-10247 Data loading of functional-query ORC fails with EOFException
• Is contained by: HIVE-26751 Bug Fixes and Improvements for 3.2.0 release
• links to: GitHub Pull Request #1774