[SPARK-19909][SS] Batches will fail in case that temporary checkpoint dir is on local file system while metadata dir is on HDFS

[sarutak]

What changes were proposed in this pull request?

When we try to run Structured Streaming in local mode but use HDFS for the storage, batches will be fail because of error like as follows.

val handle = stream.writeStream.format("console").start()
17/03/09 16:54:45 ERROR StreamMetadata: Error writing stream metadata StreamMetadata(fc07a0b1-5423-483e-a59d-b2206a49491e) to /private/var/folders/4y/tmspvv353y59p3w4lknrf7cc0000gn/T/temporary-79d4fe05-4301-4b6d-a902-dff642d0ddca/metadata
org.apache.hadoop.security.AccessControlException: Permission denied: user=kou, access=WRITE, inode="/private/var/folders/4y/tmspvv353y59p3w4lknrf7cc0000gn/T/temporary-79d4fe05-4301-4b6d-a902-dff642d0ddca/metadata":hdfs:supergroup:drwxr-xr-x

It's because that a temporary checkpoint directory is created on local file system but metadata whose path is based on the checkpoint directory will be created on HDFS.

This PR will fixe this issue.

How was this patch tested?

I tested manually in local mode with HDFS.

[SparkQA]

Test build #74334 has finished for PR 17248 at commit cc7a3f8.

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

[mgaido91]

I don't think this PR resolve properly the issue. Indeed, it somewhat forces the metadata to be written in a local dir instead of the configured default filesystem.
Of course, this fixes the exception, but we loose all the benefits of a distributed file system, as fault tolerance.
Thus in my opinion, it would be better to let the metadata be written on the default file system, but changing the default location.

[gatorsmile]

We are closing it due to inactivity. please do reopen if you want to push it forward. Thanks!

[sarutak]

O.K. I'll close this PR. Thanks!