[SPARK-12934][SQL] Count-min sketch serialization

[liancheng]

This PR adds serialization support for CountMinSketch.

A version number is added to version the serialized binary format.

[SparkQA]

Test build #49967 has finished for PR 10893 at commit e97d7f9.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds the following public classes (experimental):
  • throw new CountMinSketchMergeException(\"Cannot merge estimator of class \" + other.getClass().getName());
  • public class CountMinSketchMergeException extends Exception

[rxin]

let's put the implementation in CountMinSketchImpl, and simply delegate to a static method there

[rxin]

also somewhere in CountMinSketchImpl we should document the serialization format.

[cloud-fan]

Can we move this out so that we can also use it in bloom filter?

[rxin]

This should be count min sketch specific, because the two binary protocols can and should evolve separately.

[SparkQA]

Test build #50023 has finished for PR 10893 at commit 9acfe33.

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

[SparkQA]

Test build #50025 has finished for PR 10893 at commit 4636af1.

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

[SparkQA]

Test build #50020 has finished for PR 10893 at commit 4abc4e0.

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

[cloud-fan]

should we move this comment near writeTo?

[liancheng]

Actually I couldn't decide whether to put it before writeTo or readFrom, and then ended up here...

[cloud-fan]

LGTM overall

[rxin]

should add some check here to throw an exception if the version number is not 1.

[rxin]

Going to merge this. Thanks. Please address my (one) feedback in your next pr.

[tedyu]

This should be closed before returning from the method, right ?

[liancheng]

But DataInputStream.close() also closes the underlying input stream, which isn't our expected behavior, is it?