Attempt to rewrite stratified for `data.table`. The `data.frame` version can be found at https://gist.github.com/mrdwab/6424112

stratifiedDT.R

stratifiedDT <- function(indt, group, size, select = NULL, 
                         replace = FALSE, keep.rownames = FALSE,
                         bothSets = FALSE) {
  if (is.numeric(group)) group <- names(indt)[group]
  if (!is.data.table(indt)) indt <- as.data.table(
    indt, keep.rownames = keep.rownames)
  if (is.null(select)) {
    indt <- indt
  } else {
    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 <- indt[rowSums(temp) == length(select), ]
  }
  df.table <- indt[, .N, by = group]
  df.table
  if (length(size) > 1) {
    if (length(size) != nrow(df.table))
      stop("Number of groups is ", nrow(df.table),
           " but number of sizes supplied is ", length(size))
    if (is.null(names(size))) {
      stop("size should be entered as a named vector")
    } else {
      ifelse(all(names(size) %in% do.call(
        paste, df.table[, group, with = FALSE])),
        n <- merge(
          df.table, 
          setnames(data.table(names(size), ss = size), 
                   c(group, "ss")), by = group),
        stop("Named vector supplied with names ",
             paste(names(size), collapse = ", "),
             "\n but the names for the group levels are ",
             do.call(paste, c(unique(
               df.table[, group, with = FALSE]), collapse = ", "))))
    }
  } else if (size < 1) {
    n <- df.table[, ss := round(N * size, digits = 0)]
  } else if (size >= 1) {
    if (all(df.table$N >= size) || isTRUE(replace)) {
      n <- cbind(df.table, ss = size)
    } else {
      message(
        "Some groups\n---",
        do.call(paste, c(df.table[df.table$N < size][, group, with = FALSE], 
                         sep = ".", collapse = ", ")),
        "---\ncontain fewer observations",
        " than desired number of samples.\n",
        "All observations have been returned from those groups.")
      n <- cbind(df.table, ss = pmin(df.table$N, size))
    }
  }
  setkeyv(indt, group)
  setkeyv(n, group)
  indt[, .RNID := sequence(nrow(indt))]
  out1 <- indt[indt[n, list(
    .RNID = sample(.RNID, ss, replace)), by = .EACHI]$`.RNID`]

  if (isTRUE(bothSets)) {
    out2 <- indt[!.RNID %in% out1$`.RNID`]
    indt[, .RNID := NULL]
    out1[, .RNID := NULL]
    out2[, .RNID := NULL]
    list(SAMP1 = out1, SAMP2 = out2)
  } else {
    indt[, .RNID := NULL]
    out1[, .RNID := NULL][]
  }
}

Changing the last few rows to use .EACHI, like in the following, makes a big difference:

setkey(indt, GrpKey)
indt[, .RNID := sequence(nrow(indt))]
matchdt <- data.table(GrpKey = names(n), 
                    .ss = n, key = "GrpKey")
out <- indt[indt[matchdt, list(
.RNID = sample(.RNID, .ss, FALSE)), by = .EACHI]$`.RNID`]
out[, c("GrpKey", ".RNID") := NULL][]

Here are some new benchmarks:

microbenchmark(df1(), dt1(), times = 20)
# Unit: milliseconds
#   expr      min       lq   median       uq       max neval
#  df1() 86.21819 87.62596 89.66106 93.42402 115.06335    20
#  dt1() 38.72521 38.90182 39.16075 39.77761  56.36423    20
microbenchmark(df2(), dt2(), times = 10)
# Unit: milliseconds
#   expr      min        lq    median       uq       max neval
#  df2() 702.7874 715.86875 735.96753 757.0229 954.10381    10
#  dt2()  62.0462  62.74495  63.08022  64.7172  69.79193    10
microbenchmark(df3(), dt3(), times = 3)
# Unit: milliseconds
#   expr        min         lq     median         uq       max neval
#  df3() 1529.44492 1553.13860 1576.83227 1581.50246 1586.1726     3
#  dt3()   84.64663   86.77603   88.90543   89.87317   90.8409     3