Question

How do I select all the rows that have a missing value in the primary key in a data table.

DT = data.table(x=rep(c("a","b",NA),each=3), y=c(1,3,6), v=1:9)
setkey(DT,x)   

Selecting for a particular value is easy

DT["a",]  

Selecting for the missing values seems to require a vector search. One cannot use binary search. Am I correct?

DT[NA,]# does not work
DT[is.na(x),] #does work

Answer 1

Fortunately, DT[is.na(x),] is nearly as fast as (e.g.) DT["a",], so in practice, this may not really matter much:

library(data.table)
library(rbenchmark)

DT = data.table(x=rep(c("a","b",NA),each=3e6), y=c(1,3,6), v=1:9)
setkey(DT,x)  

benchmark(DT["a",],
          DT[is.na(x),],
          replications=20)
#             test replications elapsed relative user.self sys.self user.child
# 1      DT["a", ]           20    9.18    1.000      7.31     1.83         NA
# 2 DT[is.na(x), ]           20   10.55    1.149      8.69     1.85         NA

===

Addition from Matthew (won’t fit in comment) :

The data above has 3 very large groups, though. So the speed advantage of binary search is dominated here by the time to create the large subset (1/3 of the data is copied).

benchmark(DT["a",],  # repeat select of large subset on my netbook
    DT[is.na(x),],
    replications=3)
          test replications elapsed relative user.self sys.self
     DT["a", ]            3   2.406    1.000     2.357    0.044
DT[is.na(x), ]            3   3.876    1.611     3.812    0.056

benchmark(DT["a",which=TRUE],   # isolate search time
    DT[is.na(x),which=TRUE],
    replications=3)
                      test replications elapsed relative user.self sys.self
     DT["a", which = TRUE]            3   0.492    1.000     0.492    0.000
DT[is.na(x), which = TRUE]            3   2.941    5.978     2.932    0.004

As the size of the subset returned decreases (e.g. adding more groups), the difference becomes apparent. Vector scans on a single column aren’t too bad, but on 2 or more columns it quickly degrades.

Maybe NAs should be joinable to. I seem to remember a gotcha with that, though. Here’s some history linked from FR#1043 Allow or disallow NA in keys?. It mentions there that NA_integer_ is internally a negative integer. That trips up radix/counting sort (iirc) resulting in setkey going slower. But it’s on the list to revisit.