[SPARK-18155] CLONE - HDFSMetadataLog should not leak CRC files

• Type: Sub-task
• Status: Resolved
• Priority: Major
• Resolution: Invalid
• Component/s: DStreams
• Assignee: Unassigned
• Reporter: Sunil Kumar
• Created: 28/Oct/16 07:06
• Updated: 28/Oct/16 09:58
• Resolved: 28/Oct/16 09:58

Description

When HDFSMetadataLog uses a log directory on a filesystem other than HDFS (i.e. NFS or the driver node's local filesystem), the class leaves orphan checksum (CRC) files in the log directory. The files have names that follow the pattern "..[long UUID hex string].tmp.crc". These files exist because HDFSMetaDataLog renames other temporary files without renaming the corresponding checksum files. There is one CRC file per batch, so the directory fills up quite quickly.

I'm not certain, but this problem might also occur on certain versions of the HDFS APIs.