[SPARK-13260][SQL] count(*) does not work with CSV data source

[HyukjinKwon]

https://issues.apache.org/jira/browse/SPARK-13260
This is a quicky fix for count(*).

When the requiredColumns is empty, currently it returns sqlContext.sparkContext.emptyRDD[Row] which does not have the count.

Just like JSON datasource, this PR lets the CSV datasource count the rows but do not parse each set of tokens.

[SparkQA]

Test build #51091 has finished for PR 11169 at commit b52e156.

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

[HyukjinKwon]

cc @rxin @falaki

[falaki]

@HyukjinKwon Thanks for submitting this. I think it is better to parse everything when required columns is empty. The reason is depending on parsing mode, number of rows might turn out different.

[rxin]

hm it seems like an important optimization that we wouldn't be able to do if we cannot skip parsing columns.

[HyukjinKwon]

@rxin @falaki I think I should have described this in more details. This works identical with the original CSV datasource.

When the parsing mode is drop-malformed, then it will try to parse all and in other modes, it would not.

The similar issue was found here https://github.com/databricks/spark-csv/issues/219 and it was fixed here databricks/spark-csv#220.

[HyukjinKwon]

This CSVRelation.scala#L193-L199 will make sure it parses everything when drop-malformed mode but it does not in other modes.

[HyukjinKwon]

@rxin Can we maybe merge this for now and then take the optimisation into account in another PR?

That optimisation would apply to all the pruned scan (when it's drop-malformed mode) as well and I think we should deal with this in another PR.

[falaki]

Thanks @HyukjinKwon this looks good to me. Github diff makes it look much bigger change than it is.

[rxin]

I'm going to merge this. Thanks.