[
https://issues.apache.org/jira/browse/HADOOP-13221?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16108924#comment-16108924
]
Steve Loughran commented on HADOOP-13221:
-----------------------------------------
good catch of the Adl problem; shows how the same assumption "who would create
a file under a file" reoccurs in independent codepaths. We all bring the same
assumption about what a sensible FS would do, and code for it.
For S3N, I propose a skip(). For S3A, well, that's the tough one, isn't it? I'd
go with a direct parent check, but the cost of checking an entire path is
potentially very expensive. And we have never, ever, seen this problem arise in
the field..it's a corner case which production code doesn't attempt. I think
that's partially because production code tends to call mkdirs() before creating
files, and mkdirs does contain the check
> s3a create() doesn't check for a parent path being a file
> ---------------------------------------------------------
>
> Key: HADOOP-13221
> URL: https://issues.apache.org/jira/browse/HADOOP-13221
> Project: Hadoop Common
> Issue Type: Sub-task
> Components: fs/s3
> Affects Versions: 2.7.2
> Reporter: Steve Loughran
> Assignee: Rajesh Balamohan
> Attachments: HADOOP-13321-test.patch
>
>
> Seen in a code review. Notable that if true, this got by all the FS contract
> tests —showing we missed a couple.
> {{S3AFilesystem.create()}} does not examine its parent paths to verify that
> there does not exist one which is a file. It looks for the destination path
> if overwrite=false (see HADOOP-13188 for issues there), but it doesn't check
> the parent for not being a file, or the parent of that path.
> It must go up the tree, verifying that either a path does not exist, or that
> the path is a directory. The scan can stop at the first entry which is is a
> directory, thus the operation is O(empty-directories) and not O(directories).
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
