Добавление обратного отсчета в data.table

Решение data.table не только предпочтительно, но и красиво.

library(data.table)

dt = structure(list(date = structure(19309:19318, class = c("IDate", 
"Date")), flag = c(0, 0, 0, 0, 0, 1, 0, 0, 0, 1)), class = c( 
"data.frame"), row.names = c(NA, -10L))

setDT(dt)

dt[, countdown := rev(1:.N), by=rleid(flag)][flag==1, countdown:=0 ]
dt
#>           date flag countdown
#>  1: 2022-11-13    0         5
#>  2: 2022-11-14    0         4
#>  3: 2022-11-15    0         3
#>  4: 2022-11-16    0         2
#>  5: 2022-11-17    0         1
#>  6: 2022-11-18    1         0
#>  7: 2022-11-19    0         3
#>  8: 2022-11-20    0         2
#>  9: 2022-11-21    0         1
#> 10: 2022-11-22    1         0

^{Created on 2022-11-07 with reprex v2.0.2}

редактировать

dt[, countdown := .N:1 * !flag, by=rleid(flag)]

для краткости.

Еще data.table вариант

> cbind(dt, dt[, .(countdonwn = .N - seq(.N)), rev(cumsum(rev(flag)))][, 2])
          date flag countdonwn
 1: 2022-11-13    0          5
 2: 2022-11-14    0          4
 3: 2022-11-15    0          3
 4: 2022-11-16    0          2
 5: 2022-11-17    0          1
 6: 2022-11-18    1          0
 7: 2022-11-19    0          3
 8: 2022-11-20    0          2
 9: 2022-11-21    0          1
10: 2022-11-22    1          0

Использование sequence:

dt[ , countdown := {
  ix = which(flag == 1)
  v = ix - c(1, head(ix, -1) + 1)
  sequence(v + 1, v, -1L)
}]

          date flag countdown
 1: 2022-11-13    0         5
 2: 2022-11-14    0         4
 3: 2022-11-15    0         3
 4: 2022-11-16    0         2
 5: 2022-11-17    0         1
 6: 2022-11-18    1         0
 7: 2022-11-19    0         3
 8: 2022-11-20    0         2
 9: 2022-11-21    0         1
10: 2022-11-22    1         0

Используя rleid и rowid:

dt[ , countdown := {rf = rev(flag); rev(rowid(rleid(rf)) * !rf)}]

Обе альтернативы значительно быстрее на больших данных:

n = 1e7
set.seed(1)
d = data.table(flag = c(rbinom(n, size = 1, prob = 0.2), 1))

d1 = copy(d)
d2 = copy(d)
d3 = copy(d)
d4 = copy(d)

microbenchmark(
  ric = d1[, r := .N:1 * !flag, by=rleid(flag)],
  thomas = {d2_2 = cbind(d2, d2[, .(r = .N - seq(.N)), rev(cumsum(rev(flag)))][, 2])},
  h1 = d4[ , r := {
    ix = which(flag == 1)
    v = ix - c(1, head(ix, -1) + 1)
    sequence(v + 1, v, -1L)
  }],
  h2 = d3[ , r := {rf = rev(flag); rev(rowid(rleid(rf)) * !rf)}],
  times = 10L)

Unit: milliseconds
   expr        min         lq       mean     median         uq        max neval
    ric  3349.0538  3693.8509  3707.6691  3720.3821  3777.8165  3923.4453    10
 thomas 12914.1821 12942.3472 13049.8039 12985.2630 13162.3061 13359.6016    10
     h1   107.5380   109.0322   140.6891   110.0880   120.4707   261.9755    10
     h2   261.1989   283.2164   310.9347   293.2629   337.0358   431.7687    10

all.equal(d1, d2_2)
# [1] TRUE
all.equal(d1, d3)
# [1] TRUE
all.equal(d1, d4)
# [1] TRUE