SparkR window function

问题:

I found from JIRA that 1.6 release of SparkR has implemented window functions including lag and rank, but over function is not implemented yet. How can I use window function like lag function without over in SparkR(not the SparkSQL way)? Can someone provide an example?

回答1:

Spark 2.0.0+

SparkR provides DSL wrappers with over, window.partitionBy / partitionBy, window.orderBy / orderBy and rowsBetween / rangeBeteen functions.

Spark <= 1.6

Unfortunately it is not possible in 1.6.0. While some window functions, including lag, have been implemented SparkR doesn't support window definitions yet which renders these completely useless.

As long as SPARK-11395 is not resolved the only option is to use raw SQL:

set.seed(1)

hc <- sparkRHive.init(sc)
sdf <- createDataFrame(hc, data.frame(x=1:12, y=1:3, z=rnorm(12)))
registerTempTable(sdf, "sdf")

sql(hc, "SELECT x, y, z, LAG(z) OVER (PARTITION BY y ORDER BY x) FROM sdf") %>% 
  head()

##    x y          z        _c3
## 1  1 1 -0.6264538         NA
## 2  4 1  1.5952808 -0.6264538
## 3  7 1  0.4874291  1.5952808
## 4 10 1 -0.3053884  0.4874291
## 5  2 2  0.1836433         NA
## 6  5 2  0.3295078  0.1836433

Assuming that the corresponding PR will be merged without significant changes window definition and example query should look as follows:

w <- Window.partitionBy("y") %>% orderBy("x")
select(sdf, over(lag(sdf$z), w))