tests/testthat/test.transformations.R

context("Transformations")

test_that("Safe lookup works", {
  dt <- data.table(id = c(1L, 2L, 3L, 4L), ref = c(2L, 3L, 2L, 4L))
  dt.lookup <- data.table(
    ref = c(4L, 1L, 2L, 3L),
    value = c("c", "a", "b", NA_character_)
  )
  res <- safeLookup(dt, dt.lookup, by = "ref")
  expect <- c("b", NA_character_, "b", "c")
  expect_equal(res$value, expect)

  dt <- data.table(id = c(1L, 2L, 3L), ref = c(2L, 3L, 2L))
  dt.lookup <- data.table(
    ref = c(1L, 2L, 3L, 2L),
    value = c("a", "b", NA_character_, "c")
  )
  res <- expect_error(
    safeLookup(dt, dt.lookup, by = "ref"),
    regexp = "unique"
  )

  dt <- data.table(id = c(1L, 2L, 3L), ref = c(2L, 3L, 2L))
  dt <- dt[id == 0]
  dt.lookup <- data.table(ref = c(0), value = c(0))
  expect_warning(
    res <- safeLookup(dt, dt.lookup, by = "ref"),
    regexp = "records"
  )
  expect_equal(colnames(res), c("id", "ref", "value"))

  dt <- data.frame(id = c(1L, 2L, 3L), ref = c(2L, 3L, 2L))
  dt.lookup <- data.frame(
    ref = c(1L, 2L, 3L),
    value = c("a", "b", NA_character_)
  )
  res <- expect_error(safeLookup(dt, dt.lookup, by = "ref"), "data.frame")

  # can join by two columns
  dt <- data.table(
    id = c(1L, 2L, 3L, 4L, 5L),
    ref = c(2L, 3L, 2L, 4L, 3L),
    ref.b = c(12L, 13L, 12L, 14L, 14L)
  )
  dt.lookup <- data.table(
    ref = c(4L, 1L, 2L, 3L, 3L),
    ref.b = c(14L, 11L, 12L, 13L, 14L),
    value = c("c", "a", "b", NA_character_, "x")
  )
  res <- safeLookup(dt, dt.lookup, by = c("ref", "ref.b"))
  expect <- c("b", NA_character_, "b", "c", "x")
  expect_equal(res$value, expect)

  # Merging to create a new object doesn't alter the original tables.
  dt <- data.table(
    id = c(1L, 2L, 3L, 4L, 5L),
    ref = c(2L, 3L, 2L, 4L, 3L),
    ref.b = c(12L, 13L, 12L, 14L, 14L)
  )
  dt.lookup <- data.table(
    ref = c(4L, 1L, 2L, 3L, 3L),
    ref.b = c(14L, 11L, 12L, 13L, 14L),
    value = c("c", "a", "b", NA_character_, "x")
  )
  res <- safeLookup(dt, dt.lookup, by = c("ref", "ref.b"))
  expect_true(!"000000" %in% colnames(dt))
})

test_that("Header transformation works", {
  dt <- data.table(
    "_under_SCORE_" = "underscore",
    "-da-Sh-" = "dash",
    " sP ace " = "space",
    "C-om bin.ed" = "combined"
  )
  res <- conformHeader(names(dt))
  expect <- c(".under.score.", ".da.sh.", ".sp.ace.", "c.om.bin.ed")
  expect_identical(res, expect)

  res <- conformHeader(names(dt), separator = "_")
  expect <- c("_under_score_", "_da_sh_", "_sp_ace_", "c_om_bin_ed")
  expect_identical(res, expect)
})

test_that("disaggregate", {
  # Correct number of rows.
  dt <- fread(
    "
    sku weight
    CUSISLA03BLU-UK 3
    RUGISIS04BEI-UK 4
    TBLBRM001WHI-UK 0",
    header = T,
    stringsAsFactors = F
  )

  calculated <- disaggregate(dt = dt, by = "weight")
  expected <- fread(
    "
    sku weight
    CUSISLA03BLU-UK 1
    CUSISLA03BLU-UK 1
    CUSISLA03BLU-UK 1
    RUGISIS04BEI-UK 1
    RUGISIS04BEI-UK 1
    RUGISIS04BEI-UK 1
    RUGISIS04BEI-UK 1",
    header = T,
    stringsAsFactors = F
  )
  expect_equal(
    calculated, expected,
    label = "Each line is split into multiple matching the weight."
  )

  # Data table retains its class.
  dt <- fread(
    "
    sku weight
    CUSISLA03BLU-UK 3
    RUGISIS04BEI-UK 4
    TBLBRM001WHI-UK 0",
    header = TRUE,
    stringsAsFactors = FALSE
  )
  calculated <- disaggregate(dt = dt, by = "weight")
  expect_equal(
    sum(calculated$weight), sum(dt$weight),
    label = "The total weight matches that of the original."
  )

  # Row numbers are kept correctly.
  dt <- fread(
    "
    sku weight
    CUSISLA03BLU-UK 3
    RUGISIS04BEI-UK 0
    TBLBRM001WHI-UK 2",
    header = TRUE,
    stringsAsFactors = FALSE
  )
  calculated <- disaggregate(dt = dt, by = "weight", keep.row.number = TRUE)
  expect_true(
    all(calculated$rn, as.integer(rep(rownames(dt), dt$weight))),
    label = "The total weight matches that of the original."
  )
})
madedotcom/retl documentation built on March 25, 2020, 4:28 p.m.