[ORC-435] Ability to read stripes that are greater than 2GB

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 1.3.4, 1.4.4, 1.5.3, 1.6.0
• Fix Version/s: 1.5.4, 1.6.0
• Component/s: Reader
• Assignee: Prasanth Jayachandran
• Reporter: Prasanth Jayachandran
• Created: 09/Nov/18 21:35
• Updated: 05/May/21 16:44
• Resolved: 27/Nov/18 23:14

Description

ORC reader fails with NegativeArraySizeException if the stripe size is >2GB. Even though default stripe size is 64MB there are cases where stripe size will reach >2GB even before memory manager can kick in to check memory size. Say if we are inserting 500KB strings (mostly unique) by the time we reach 5000 rows stripe size is already over 2GB. Reader will have to chunk the disk range reads for such cases instead of reading the stripe as whole blob. 

Exception thrown when reading such files

2018-10-12 21:43:58,833 WARN [main] org.apache.hadoop.mapred.YarnChild: Exception running child : java.lang.NegativeArraySizeException
        at org.apache.hadoop.hive.ql.io.orc.RecordReaderUtils.readDiskRanges(RecordReaderUtils.java:272)
        at org.apache.hadoop.hive.ql.io.orc.RecordReaderImpl.readPartialDataStreams(RecordReaderImpl.java:1007)
        at org.apache.hadoop.hive.ql.io.orc.RecordReaderImpl.readStripe(RecordReaderImpl.java:835)
        at org.apache.hadoop.hive.ql.io.orc.RecordReaderImpl.advanceStripe(RecordReaderImpl.java:1029)
        at org.apache.hadoop.hive.ql.io.orc.RecordReaderImpl.advanceToNextRow(RecordReaderImpl.java:1062)
        at org.apache.hadoop.hive.ql.io.orc.RecordReaderImpl.next(RecordReaderImpl.java:1085)

Issue Links

• causes: ORC-793 Deprecate the incorrect getInt method and add a new setInt within OrcConf
• is blocked by: ORC-438 NPE in StringStatisticsImpl.merge()
• links to: GitHub Pull Request #338