I have a Spark data frame as below (dummy data) and would want to add percentiles as a column to it using SparkR:
UserId Values
U1 5
U2 50
U3 76
.
.
.
.
.
U1000 8
Example:
UserId Values Percentile
U1 5 0.011144
U2 50 0.212245
U3 76 0.811986
.
.
.
.
.
U1000 8 0.0134567
Well, there is a function, percent_rank, but it only operates on window partitions; adapting the documentation example:
library(SparkR)
library(magrittr)
sparkR.version()
# [1] "2.2.0"
df <- createDataFrame(mtcars)
ws <- orderBy(windowPartitionBy("am"), "hp")
out <- select(df, over(percent_rank(), ws), df$hp, df$am)
out <- out %>% withColumnRenamed(colnames(out)[1], 'percentile') # get rid of the ridiculously long column name
head(out)
# percentile hp am
# 1 0.00000000 62 0
# 2 0.05555556 95 0
# 3 0.11111111 97 0
# 4 0.16666667 105 0
# 5 0.22222222 110 0
# 6 0.27777778 123 0
If you don't want a window partition but a global percentile, as in your case, the easiest way I can think of is to add first a dummy constant column and group by this, which you can subsequently drop. So, here is the above example, but this time without partitioning by am:
df_new = df %>% withColumn("dummy", lit("N"))
ws <- orderBy(windowPartitionBy('dummy'), "hp")
out <- select(df_new, over(percent_rank(), ws), df_new$hp)
out <- out %>% withColumnRenamed(colnames(out)[1], 'percentile')
The result of collect(out) is
percentile hp
1 0.00000000 52
2 0.03225806 62
3 0.06451613 65
4 0.09677419 66
5 0.09677419 66
6 0.16129032 91
7 0.19354839 93
8 0.22580645 95
9 0.25806452 97
10 0.29032258 105
11 0.32258065 109
12 0.35483871 110
13 0.35483871 110
14 0.35483871 110
15 0.45161290 113
16 0.48387097 123
17 0.48387097 123
18 0.54838710 150
19 0.54838710 150
20 0.61290323 175
21 0.61290323 175
22 0.61290323 175
23 0.70967742 180
24 0.70967742 180
25 0.70967742 180
26 0.80645161 205
27 0.83870968 215
28 0.87096774 230
29 0.90322581 245
30 0.90322581 245
31 0.96774194 264
32 1.00000000 335