reprex of very slow named vector search when searching with NA

name-value-NA-search-reprex.R

# NA key names using character vector
k1 <- c('a', 'b', NA_character_)
names(k1) <- k1
v1 <- sample(k1, 1e5, replace = TRUE, prob = c(.2, .2, .6))

# search with "null", but return NA value
k2 <- k1
names(k2) <- c('a', 'b', 'null') # new key name
# use "null" instead of NA;
v2 <- v1
v2[is.na(v2)] <- "null"

# same value in vector (after removing names)
identical(unname(k1[v1]), unname(k2[v2]))

# try with numbers
k3 <- c(1, 2, NA_real_)
names(k3) <- k3
v3 <- sample(k3, 1e5, replace = TRUE, prob = c(.2, .2, .6))

# try with factors
k4 <- as.factor(c('a', 'b', NA_character_))
names(k4) <- k4
v4 <- sample(k4, 1e5, replace = TRUE, prob = c(.2, .2, .6))

# but speed diff is enormous
kv_search <- function(k, v) k[v]
microbenchmark::microbenchmark(
    kv_search(k1,v1), # slow
    kv_search(k2,v2), # fast
    kv_search(k3,v3), # fast
    kv_search(k4,v4), # fast
  times = 5, unit = 's')

According to hadley:

There’s an internal heuristic used to decided whether to build up a hashmap/dictionary when doing character subsetting. I’d guess this is related.