rollapply for large data using sparklyr

Question

I want to estimate rolling value-at-risk for a dataset of about 22.5 million observations, thus I want to use sparklyr for fast computation. Here is what I did (using a sample database):

library(PerformanceAnalytics)
library(reshape2)
library(dplyr)

data(managers)
data <- zerofill(managers)
data<-as.data.frame(data)
class(data)
data$date=row.names(data)
lmanagers<-melt(data, id.vars=c('date'))

Now I estimate VaR using dplyr and PerformanceAnalytics packages:

library(zoo) # for rollapply()
var <- lmanagers %>% group_by(variable) %>% arrange(variable,date) %>% 
  mutate(var=rollapply(value, 10,FUN=function(x) VaR(x, p=.95, method="modified",align = "right"), partial=T))

This works fine. Now I do this to make use of sparklyr:

library(sparklyr)
sc <- spark_connect(master = "local")
lmanagers_sp <- copy_to(sc,lmanagers)
src_tbls(sc)

var_sp <- lmanagers_sp %>% group_by(variable) %>% arrange(variable,date) %>% 
  mutate(var=rollapply(value, 10,FUN=function(x) VaR(x, p=.95, method="modified",align = "right"), partial=T)) %>% 
  collect

But this gives the following error:

Error: Unknown input type: pairlist

Can anyone please tell me where is the error and what is the correct code? Or any other solution to estimate rolling VaR faster is also appreciates.

Answer 1

For custom dplyr backends like sparklyr, mutate does not currently support arbitrary R functions defined in other packages; therefore, rollapply() is currently unsupported.

In order to calculate value-at-risk in sparklyr, one approach is to extend sparklyr using Scala and R and follow an approach similar to: Estimating Financial Risk with Apache Spark.