Applying sd() to a single value in a spark data frame (via sparklyr package in R) results in a missing value that cannot be filtered out based on it being a missing value.
Can someone explain this / provide a good solution?
Example below.
library(sparklyr)
library(dplyr)
sc <- spark_connect(master = "local")
#> * Using Spark: 2.1.0
x <- data.frame(grp = c("a", "a", "c"), x = c(1, 2, 3))
copy_to(sc, x, "tmp", overwrite = TRUE)
#> # Source: table<tmp> [?? x 2]
#> # Database: spark_connection
#> grp x
#> <chr> <dbl>
#> 1 a 1
#> 2 a 2
#> 3 c 3
x_tbl <- tbl(sc, "tmp") %>% group_by(grp) %>% mutate(x_sd = sd(x))
x_tbl
#> # Source: lazy query [?? x 3]
#> # Database: spark_connection
#> # Groups: grp
#> grp x x_sd
#> <chr> <dbl> <dbl>
#> 1 a 1 0.7071068
#> 2 a 2 0.7071068
#> 3 c 3 NaN
x_tbl %>% filter(!is.na(x_sd)) %>% collect()
#> # A tibble: 3 x 3
#> # Groups: grp [2]
#> grp x x_sd
#> <chr> <dbl> <dbl>
#> 1 a 1 0.7071068
#> 2 a 2 0.7071068
#> 3 c 3 NaN
This is a matter of incompatibility between sparklyr and Spark. In Spark there are both NULLS (somewhat equivalent to R NA) and NaNs, each with different processing rules, but both values are fetched as NaN in sparklyr.
To filter out NaN you have to use isnan (don't confuse it with R is.nan):
x_tbl %>% filter(!isnan(x_sd)) %>% collect()
# A tibble: 2 x 3
# Groups: grp [1]
grp x x_sd
<chr> <dbl> <dbl>
1 a 1 0.7071068
2 a 2 0.7071068
To better illustrate the problem:
df <- copy_to(sc,
data.frame(x = c("1", "NaN", "")), "df", overwrite = TRUE
) %>% mutate(x = as.double(x))
df %>% mutate_all(funs(isnull, isnan))
# Source: lazy query [?? x 3]
# Database: spark_connection
x isnull isnan
<dbl> <lgl> <lgl>
1 1 FALSE FALSE
2 NaN FALSE TRUE
3 NaN TRUE FALSE