Possible rewrite of `stratified`. Not considerably faster, but this seems to be easier to follow by refactoring the code.

stratified.R

## Helper functions. Won't bother exporting.
dt_check <- function(indt) {
  if (!is.data.table(indt)) as.data.table(indt) else indt
}
g_s <- function(indt, group)       indt[, .N, by = group]
g_n <- function(indt, group, size) indt[, list(ss = size), by = group]
g_f <- function(indt, group, size) indt[, list(ss = ceiling(.N * size)), by = group]
g_l <- function(indt, group, size) setnames(data.table(names(size), unname(size)), c(group, "ss"))[]
g_sel <- function(indt, select) {
  if (is.null(names(select))) {
    stop("'select' must be a named list")
  } 
  if (!all(names(select) %in% names(indt))) {
    stop("Please verify your 'select' argument")
  } 
  temp <- vapply(names(select), 
                 function(x) indt[[x]] %in% select[[x]], 
                 logical(nrow(indt)))
  indt[rowSums(temp) == length(select)]
}
g_comp <- function(indt, comparedt, group, replace) {
  checker <- indt[comparedt, on = group]
  if (checker[, any(ss > N)]) {
    if (!isTRUE(replace)) {
      checker_meta <- checker[which(ss > N)]
      checker[, ss := pmin(ss, N)]
    }
  }
  if (exists("checker_meta")) {
    message("Some groups have fewew values than requested:\n")
    message(paste(capture.output(checker_meta), collapse = "\n"))
    message("\n")
  }
  checker[, N := NULL][]
}

stratified <- function(indt, group, size, select = NULL, replace = FALSE, bothSets = FALSE) {
  indt <- dt_check(indt)
  group <- splitstackshape:::Names(indt, group)
  if (!is.null(select)) indt <- g_sel(indt, select)
  compare_me <- g_s(indt, group)
  n <- {
    if (length(size) > 1) {
      if (length(size) != nrow(compare_me)) {
        stop("Incorrect number of groups specified in 'size'")
      }
      g_comp(g_l(indt, group, size), compare_me, group, replace)
    } else if (size < 1) {
      g_comp(g_f(indt, group, size), compare_me, group, replace)
    } else {
      g_comp(g_n(indt, group, size), compare_me, group, replace)
    } 
  }
  out1 <- indt[n, sample(.I, unlist(ss), replace), on = group, by = .EACHI]$V1
  if (isTRUE(bothSets)) {
    out2 <- indt[!sequence(nrow(indt)) %in% out1]
    list(SAMP1 = indt[out1], SAMP2 = out2)
  } else {
    indt[out1]
  }
}

hi @mrdwab. stratified is a great function, generally, for cross validation. I am wondering whether you could make size = 1 available for sampling all observations in each stratum with replacement? I think that would be very useful for non-parametric bootstrapping. Thanks!