Search code examples
rapache-sparkdataframeapache-spark-sqlsparkr

How can I convert groupedData into Dataframe in R

Consider I have the below dataframe

AccountId,CloseDate
1,2015-05-07
2,2015-05-09
3,2015-05-01
4,2015-05-07
1,2015-05-09
1,2015-05-12
2,2015-05-12
3,2015-05-01
3,2015-05-01
3,2015-05-02
4,2015-05-17
1,2015-05-12

I want to group it based on AccountId and then I want to add another column naming date_diff which will contain the difference in CloseDate between the current row and previous row. Please note that I want this date_diff to be calculated only for rows having same AccountId. So I need to group the data before adding another column

Below is the R code that I am using

  df <- read.df(sqlContext, "/home/ubuntu/work/csv/sample.csv", source = "com.databricks.spark.csv", inferSchema = "true", header="true")
  df$CloseDate <- to_date(df$CloseDate)
  groupedData <- SparkR::group_by(df, df$AccountId)
  SparkR::mutate(groupedData, DiffCloseDt = as.numeric(SparkR::datediff((CloseDate),(SparkR::lag(CloseDate,1)))))

To add another column I am using mutate. But as the group_by returns groupedData I am not able to use mutate here. I am getting the below error

 Error in (function (classes, fdef, mtable)  : 
  unable to find an inherited method for function ‘mutate’ for signature ‘"GroupedData"’

So how can I convert GroupedData into Dataframe so that I can add columns using mutate?

Solution

  • What you want is not possible to achieve using group_by. As already explained quite a few times on SO :

    group_by on a DataFrame doesn't physical group the data. Moreover order of operations after applying group_by is nondeterministic.

    To achieve desired output you'll have to use window functions and provide an explicit ordering:

    df <- structure(list(AccountId = c(1L, 2L, 3L, 4L, 1L, 1L, 2L, 3L, 
  3L, 3L, 4L, 1L), CloseDate = structure(c(3L, 4L, 1L, 3L, 4L, 
  5L, 5L, 1L, 1L, 2L, 6L, 5L), .Label = c("2015-05-01", "2015-05-02", 
  "2015-05-07", "2015-05-09", "2015-05-12", "2015-05-17"), class = "factor")), 
  .Names = c("AccountId", "CloseDate"),
  class = "data.frame", row.names = c(NA, -12L))

hiveContext <- sparkRHive.init(sc)
sdf <- createDataFrame(hiveContext, df)
registerTempTable(sdf, "df")

query <- "SELECT *, LAG(CloseDate, 1) OVER (
  PARTITION BY AccountId ORDER BY CloseDate
) AS DateLag FROM df"

dfWithLag <- sql(hiveContext, query)

withColumn(dfWithLag, "diff", datediff(dfWithLag$CloseDate, dfWithLag$DateLag)) %>%
  head()

##   AccountId  CloseDate    DateLag diff
## 1         1 2015-05-07       <NA>   NA
## 2         1 2015-05-09 2015-05-07    2
## 3         1 2015-05-12 2015-05-09    3
## 4         1 2015-05-12 2015-05-12    0
## 5         2 2015-05-09       <NA>   NA
## 6         2 2015-05-12 2015-05-09    3