Details
-
Bug
-
Status: Resolved
-
Minor
-
Resolution: Workaround
-
None
-
None
-
None
Description
I'm seeing issues where the Hadoop FileSystem instance is closed out from under other objects. From what I understand, the Hadoop FileSystem object (org.apache.hadoop.fs.FileSystem) retrieved with FileSystem.get as is done in HdfsDirectory's ctor is a shared object that if closed, can affect other code using that same shared instance. You can see this is a cached, shared object here ->
https://github.com/apache/hadoop/blob/trunk/hadoop-common-project/hadoop-common/src/main/java/org/apache/hadoop/fs/FileSystem.java#L455
Thus, I suspect Solr should not be closing any FileSystem instance retrieved with get. It's important to mention that if I set the fs.$SCHEME.impl.disable.cache to true, then my problems go away, which seems to confirm that Solr is using the API incorrectly. That being said, I'm surprised this hasn't been raised before, so maybe I've missed something basic in Solr's use of HDFS?
Attachments
Issue Links
- is related to
-
SOLR-9958 The FileSystem used by HdfsBackupRepository gets closed before the backup completes.
- Resolved
- relates to
-
SOLR-13587 Close BackupRepository after every usage
- Patch Available
-
SOLR-11473 Make HDFSDirectoryFactory support other prefixes (besides hdfs:/)
- Closed