[WIP][SPARK-17046][SQL] prevent user using dataframe.select with empty param list

[WeichenXu123]

What changes were proposed in this pull request?

We can see the DataFrame API:
def select(col: String, cols: String*)
such definition can prevent user to call select in such way: df.select( )

but, currently we can still use df.select( ) and pass compiling,
because it match the API
def select(cols: Column*)

so, my modification is, add an API such as:
def select(col: Column, cols: Column*)
and change def select(cols: Column*) into private[spark] def select(cols: Column*)
so that the public select API can only be called with non-empty param list.

How was this patch tested?

Existing test.

[SparkQA]

Test build #63723 has finished for PR 14629 at commit 1179249.

• This patch fails to build.
• This patch merges cleanly.
• This patch adds no public classes.

[SparkQA]

Test build #63724 has finished for PR 14629 at commit 8060aa4.

• This patch fails to build.
• This patch merges cleanly.
• This patch adds no public classes.

[WeichenXu123]

@srowen
How do you think about this problem?
I found adding two method like
def select(cols: Column*)
def select(col: Column, cols: Column*)
causing ambiguous,
I prepare to change def select(cols: Column*) into def selectInternal(cols: Column*)
and mark selectInternal as private[spark]
do you think it reasonable?

[rxin]

Why do we want to enforce this? It is valid to have a DataFrame without any columns.

[srowen]

Yes that's a good question. A 0-column DataFrame is valid, though that's a little different from being able to select 0 columns from a DataFrame. I don't have a database handy, but can you select no columns in any SQL syntax? Maybe best to emulate that?

[WeichenXu123]

MySql do not allow select with 0 columns, and I think select() is useless, no one will do such operation, so, is it better to generate compiling error when detecting code use df.select() because it is usually a coding mistake?
Or, df.select() is useful in some particular scenario?

[hvanhovell]

A df.select() without any columns is not useless IMO: You can still get a valid count() from a data frame.

[srowen]

Interesting point, yeah, because normally in an RDBMS you have to COUNT(*) or COUNT(1) and the argument is useless anyway, so would be nice to not have to provide an argument to select in this context. But while comparing to DataFrames, of course you should already count() directly if this is desired.

[srowen]

@hvanhovell @rxin unless you've changed your stance a little bit on this, I think the conclusion is that this isn't worth changing this behavior and we can close this @WeichenXu123

[hvanhovell]

I haven't changed my mind of this. Lets close this one.