[SPARK-11715][SPARKR] Add R support corr for Column Aggregration

[felixcheung]

Need to match existing method signature

[SparkQA]

Test build #45789 has finished for PR 9680 at commit 940164c.

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

[shivaram]

cc @NarineK

I think #9366 is a more general form of this ? If so we could probably just wait for #9366 ?

[felixcheung]

I think 9366 is about computing corr or cov matrix whereas this is computing corr between two columns. They seem to be useful in their own ways. Also this is already supported in Scala and Python.

[NarineK]

in R the general formula for correlation is the following:
cor(x, y = NULL, use = "everything", method = c("pearson", "kendall", "spearman"))
in #9366 is the special case when x = y, but in general x and y are 2 different dataframes with the same number of rows.
I did the pull request specifically for x = y, but the jira is more general.

[felixcheung]

So #9366 is for all columns in DataFrame x=y or different x, y DataFrames
And this #9680 is for 2 columns in one DataFrame.

[sun-rui]

these are two different issues.

[sun-rui]

this signature looks confusing. Maybe change the generic function definition of "corr" is better:

setGeneric("corr", function(x, ...) {standardGeneric("corr") })
setMethod("corr",
signature(x = "DataFrame"),
function(x, col1, col2, method = "pearson") {
...
}
setMethod("corr", signature(x = "characterOrColumn"),
function(x, col2) {
...
}

[sun-rui]

One concern is how documentation for these "corr" methods are generated?

[felixcheung]

right, I like the approach of changing the existing generic definition.

perhaps we should align the method signature with the stats::cor

cor(x, y = NULL, use = "everything",
method = c("pearson", "kendall", "spearman"))

do you know why we decide to name it corr (vs. cor) in other places?

[felixcheung]

as for doc, the DataFrame corr in stats.R has @rdname statfunctions
this one has @rdname corr

so they go to different HTML page generated by roxygen2

[sun-rui]

Maybe we can add cor() as alias for corr(), as you did in #9489

[felixcheung]

Thinking more about this, I think what's being added in #9366 matches https://stat.ethz.ch/R-manual/R-devel/library/stats/html/cor.html better. When we are adding that support in R we could add it as cor matching stats::cor.

Meanwhile I'll change corr to what you suggested with function(x, ...)

[sun-rui]

#9366 only supports inter-column cov and cor of a DataFrame, not between columns of two DataFrames. I think actually it is better to add alias in this PR. corr() operating on two columns is similar to R cor() on two vectors.

[felixcheung]

@sun-rui I updated it. I think it's a bit not as strongly typed as I'd like but if I add col2 = "Column" to signature I get this error:

Error in match.call(definition, call, expand.dots, envir) :
unused argument (col2 = c("Column", ""))

[SparkQA]

Test build #46055 has finished for PR 9680 at commit c5f5524.

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

[felixcheung]

any more comment?

[sun-rui]

There are two versions of corr():

def corr(column1: Column, column2: Column): Column
def corr(columnName1: String, columnName2: String): Column

We'd better support both. Something like:

setMethod("corr", signature(x = "characterOrColumn"),

[felixcheung]

That's the same for count, max, mean and so on, so change we would need to change every function here - should we do that?

[felixcheung]

@shivaram Can we go ahead with this? I think we could consider adding all character overload for DataFrame functions in a different JIRA.

[sun-rui]

adding all character overload for DataFrame functions in a different JIRA is OK.

But for alias of corr(), #9366 only supports inter-column cov and cor of a DataFrame, not between columns of two DataFrames. I think actually it is better to add alias in this PR. corr() operating on two columns is similar to R cor() on two vectors.

[felixcheung]

As per this https://stat.ethz.ch/R-manual/R-devel/library/stats/html/cor.html
It support "vector, matrix or data frame", and it doesn't say a subset of a data frame.
I'm not sure a column in a data frame is like a vector?
Instead, I think #9366 should support 2 DataFrames. And we should come up with a signature similar to stats::cor

[felixcheung]

Also, since here we are working with 2 columns, by adding a alias cor we will need to create a generic with a different signature that again masks stats::cor, like what's happening with stats::cov, and that seems problematic.

[felixcheung]

@sun-rui ? I"m fine with adding cor - are we ok with masking stats::cor?

[sun-rui]

@felixcheung, sorry for late response.

Since there is no agreement now, I am fine that we don't add "cor" alias in this PR. Let's get this PR merged.

Could you submit a new JIRA addressing the issue of adding alias of "cor" and also the issue of existing "cov" which masks stats::cov?

[sun-rui]

LGTM

[felixcheung]

stats::cov name conflict: https://issues.apache.org/jira/browse/SPARK-11886
will open a new JIRA on cor alias once this is merged

[felixcheung]

thanks, rebased.

[SparkQA]

Test build #46911 has finished for PR 9680 at commit b8090f3.

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

[felixcheung]

second is a git error (seems like having a lot these days?)
hudson.plugins.git.GitException: Failed to fetch from https://github.com/apache/spark.git at hudson.plugins.git.GitSCM.fetchFrom(GitSCM.java:763)

[shivaram]

Jenkins, retest this please

[SparkQA]

Test build #46922 has finished for PR 9680 at commit 659f805.

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

[shivaram]

LGTM. @felixcheung I think the current resolution of not adding cor until we can support it is fine by me. Its better to not mask existing functions unless we can match existing functionality.

Merging this to master, branch-1.6