Mingliang Liu created HADOOP-14236:
--------------------------------------
Summary: S3Guard: S3AFileSystem::rename() should move non-listed
sub-directory entries in metadata store
Key: HADOOP-14236
URL: https://issues.apache.org/jira/browse/HADOOP-14236
Project: Hadoop Common
Issue Type: Sub-task
Components: fs/s3
Reporter: Mingliang Liu
Assignee: Mingliang Liu
After running integration test {{ITestS3AFileSystemContract}}, I found the
following items are not cleaned up in DynamoDB:
{code}
parent=/mliu-s3guard/user/mliu/s3afilesystemcontract/testRenameDirectoryAsExisting/dir,
child=subdir
parent=/mliu-s3guard/user/mliu/s3afilesystemcontract/testRenameDirectoryAsExistingNew/newdir/subdir,
child=file2
{code}
At first I thought it’s similar to [HADOOP-14226] or [HADOOP-14227], and we
need to be careful when cleaning up test data.
Then I found it’s a bug in the code of integrating S3Guard with S3AFileSystem:
for rename we miss sub-directory items to put (dest) and delete (src). The
reason is that in S3A, we delete those fake directory objects if they are not
necessary, e.g. non-empty. So when we list the objects to rename, the object
summaries will only return _file_ objects. This has two consequences after
rename:
# there will be left items for src path in metadata store - left-overs will
confuse {{get(Path)}} which should return null
# we are not persisting the whole subtree for dest path to metadata store -
this will break the DynamoDBMetadataStore invariant: _if a path exists, all its
ancestors will also exist in the table_.
Existing tests are not complaining about this though. If this is a real bug,
let’s address it here.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: common-dev-unsubscr...@hadoop.apache.org
For additional commands, e-mail: common-dev-h...@hadoop.apache.org
