[SPARK-6144]When in cluster mode using ADD JAR with a hdfs:// sourced ja...

[trystanleftwich]

...r will fail

[AmplabJenkins]

Can one of the admins verify this patch?

[andrewor14]

Jenkins, this is ok to test.

By the way in the future it would be good to open this against the master branch. In this case it's fine because the 1.3 branch hasn't diverged that much yet.

[SparkQA]

Test build #625 has started for PR 4881 at commit 6e9f069.

• This patch merges cleanly.

[vanzin]

To revive the discussion of the previows PR...

Both the problems @andrewor14 and I mentioned exist because of this line. listStatus behaves differently for files and directories, so when path is a directory here you'll get both unwanted behaviors:

• children will be copied to targetDir instead of $targetDir/${filename.getOrElse(path.getName())}
• the code will try to copy all children to the target directory with the same filename in the first call from doFetchFile

So this code needs to behave differently depending on whether path is a directory or not. The code Andrew posted here is pretty close to what needs to be done.

[SparkQA]

Test build #625 has finished for PR 4881 at commit 6e9f069.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds no public classes.

[trystanleftwich]

So to confirm, i think this function needs to be able to handle 5 states:

Path is a dir which has subdirs
(structure is hdfs://foo/foo1/foo2.jar)
path = hdfs://foo -> local_dir://foo
(Where local_dir://foo looks like local_dir://foo/foo1/foo2.jar)

Path is a file you want to copy with the same name
path = hdfs://foo.jar -> local_dir/foo.jar

Path is a file which you want to copy with a different name
path = hdfs://foo.jar -> local_dir/bar.jar

Path is a dir which contains multiple files that you want to copy with the same names
(Structure is hdfs://foo/foo1.jar, hdfs://foo/foo2.jar)
path = hdfs://foo -> local_dir/foo/
(where local_dir/foo looks like local_dir/foo/foo1.jar and local_dir/foo/foo2.jar)

Path is a dir which contains multiple files and subdirs you want to copy
(similar to above)

Anything I have missed? I have some code in testing now that achieves this ill post it when i've run it on my cluster.

[vanzin]

I tested with this version of fetchHcfsFile and my tests pass:

/**
* Fetch a file or directory from a Hadoop-compatible filesystem.
*
* Visible for testing
*/
private[spark] def fetchHcfsFile(
path: Path,
targetDir: File,
fs: FileSystem,
conf: SparkConf,
hadoopConf: Configuration,
fileOverwrite: Boolean,
filename: Option[String] = None): Unit = {
if (!targetDir.exists() && !targetDir.mkdir()) {
throw new IOException(s"Failed to create directory ${targetDir.getPath}")
}
val dest = new File(targetDir, filename.getOrElse(path.getName))
if (fs.isFile(path)) {
val in = fs.open(path)
try {
downloadFile(path.toString, in, dest, fileOverwrite)
} finally {
in.close()
}
} else {
fs.listStatus(path).foreach { fileStatus =>
fetchHcfsFile(fileStatus.getPath(), dest, fs, conf, hadoopConf, fileOverwrite)
}
}
}

[andrewor14]

@trystanleftwich I believe that's correct. To summarize:

• Before this patch, adding hdfs://single/file.jar doesn't work (a regression from Spark 1.2)
• In this patch in its current state, adding hdfs://some/directory/with/multiple/files will not work (not technically a regression from Spark 1.2, but breaks a new feature introduced in SPARK-4687. Add a recursive option to the addFile API #3670)
• The goal is to make both of these work

The code snippet @vanzin posted should fix both cases.

[trystanleftwich]

Ok, i've pushed my changes, I've added tests that should cover all the states, I was getting errors with @vanzin code snippet, if you pass in dir i.e path = hdfs://foo and it contains a foo.jar i.e hdfs://foo/foo.jar, The code will create a directory local_dir/foo/foo.jar and not a file local_dir/foo/foo.jar unless im mistaken.

[vanzin]

The code will create a directory local_dir/foo/foo.jar and not a file

Hmm. Let me check that.

[vanzin]

Hi @trystanleftwich , just tested my code again with more strict checks, and files show up as files, directories show up as directories.

[andrewor14]

Leaving a link to an alternate fix in #4894

[vanzin]

I think this is where our patches diverge a bit.

The behavior I expect when I upload a directory "foo", is that there will be a directory called "foo" inside the target directory, so that when I do SparkFiles.get("foo") I get the location of that directory.

Your patch seems to place the contents of "foo" under the target dir, which is not what I'd expect.

[trystanleftwich]

Ok, fair enough misunderstanding on my behalf, I've tested your patch and it works in my case, which I expected. I'm happy to close this as likewise I just want this solved for the next release.

[andrewor14]

Let's close this PR in favor of #4894, which I just merged. Thanks for reporting this blocker.