In rstudio/htmltools: Tools for HTML

Goal

Find for what values of key size and unique key size is split() faster.

Setup

Make a list of keys of size n and have the unique count of keys be nn. The values will be "1".

combineKeys <- function(x) {
  if (anyNA(x)) {
    na_idx <- is.na(x)
    if (all(na_idx)) {
      return(NA)
    }
    x <- x[!na_idx]
  }
  unlist(x, recursive = FALSE, use.names = FALSE)
}
flattenTagAttribs <- function(attribs) {
  # concatenate attributes
  # split() is very slow, so avoid it if possible
  if (anyDuplicated(names(attribs))) {
    attribs <- lapply(split(attribs, names(attribs)), combineKeys)
  }

  attribs
}

flattenTagAttribs2 <- function(attribs) {

  attribNames <- names(attribs)

  if (anyDuplicated(attribNames)) {
    uniqueAttribNames <- sort(unique(attribNames))
    attribs <- lapply(uniqueAttribNames, function(name) {
      obj <- attribs[attribNames == name]
      combineKeys(obj)
    })
    names(attribs) <- uniqueAttribNames
  }
  attribs
}

Test

dts <- list()
for (n in seq(from = 1, to = 25, by = 1)) {
  for (nn in seq(from = 1, to = n, by = 1)) {
    nn <- min(c(nn, n))
    cat("n: ", n, " nn: ", nn, "\n")
    k <- paste0("key", 1:nn)
    if (nn < n) k <- c(k, sample(k, n - nn, replace = TRUE))
    v <- as.list(sample(1L, n, replace = TRUE))
    attribs <- setNames(v, k)
    info <- bench::mark(
      check = TRUE,
      flattenTagAttribs(attribs),
      flattenTagAttribs2(attribs)
    )
    info$key_size <- n
    info$unique_size <- nn
    dts[[length(dts) + 1]] <- info
  }
}
dt <- dplyr::bind_rows(dts)

A negative diff shows that the new function (flattenTagAttribs2()) is faster than the old function (flattenTagAttribs()).

library(ggplot2)
dt_plot <-
  dt %>%
  dplyr::select(expression, median, key_size, unique_size) %>%
  tidyr::pivot_wider(names_from = expression, values_from = median) %>%
  dplyr::rename(old = `flattenTagAttribs(attribs)`, new = `flattenTagAttribs2(attribs)`) %>%
  dplyr::mutate(diff = (as.numeric(new) - as.numeric(old)) * 1000000)

dt_plot %>%
  {
    ggplot(., aes(x = key_size, y = unique_size, color = diff, shape = diff > 0)) +
    geom_point(size = 10) +
    scale_color_gradient2(low = "#006837", mid = "#ffffbf", high = "#a50026")
  }

When the key size is less than 20, the difference between the two methods is ~0 or negative. Negative is good for flattenTagAttribs2(), showing it is faster. When the key are all unique (the diagonal), then the difference is neglible. This would advocate for only using the new method as having a key count > 20 is not a typical case. The performance loss and is less in magnitude than the what is gained in the typical cases (< 5 key) situations.

rstudio/htmltools documentation built on March 29, 2024, 2:22 p.m.