fix model fit for spark tbls (#1047)

NEWS.md

@@ -1,5 +1,7 @@
 # parsnip (development version)
 
+* Fixed bug in fitting some model types with the `"spark"` engine (#1045).
+
 * Fixed issue in `mlp()` metadata where the `stop_iter` engine argument had been mistakenly protected for the `"brulee"` engine. (#1050)
 
 * `.filter_eval_time()` was moved to the survival standalone file. 

R/arguments.R

@@ -328,7 +328,11 @@ min_cols <- function(num_cols, source) {
 #' @export
 #' @rdname min_cols
 min_rows <- function(num_rows, source, offset = 0) {
-  n <- nrow(source)
+  if (inherits(source, "tbl_spark")) {
+    n <- nrow_spark(source)
+  } else {
+    n <- nrow(source)
+  }
 
   if (num_rows > n - offset) {
     msg <- paste0(num_rows, " samples were requested but there were ", n,
@@ -340,3 +344,7 @@ min_rows <- function(num_rows, source, offset = 0) {
   as.integer(num_rows)
 }
 
+nrow_spark <- function(source) {
+  rlang::check_installed("sparklyr")
+  sparklyr::sdf_nrow(source)
+}

tests/testthat/test_boost_tree.R

@@ -28,6 +28,9 @@ test_that('bad input', {
 ## -----------------------------------------------------------------------------
 
 test_that('argument checks for data dimensions', {
+  skip_if_not_installed("sparklyr")
+  library(sparklyr)
+  skip_if(nrow(spark_installed_versions()) == 0)
 
   spec <-
     boost_tree(mtry = 1000, min_n = 1000, trees = 5) %>%
@@ -36,6 +39,10 @@ test_that('argument checks for data dimensions', {
 
   args <- translate(spec)$method$fit$args
   expect_equal(args$min_instances_per_node, expr(min_rows(1000, x)))
+
+  sc = spark_connect(master = "local")
+  cars = copy_to(sc, mtcars, overwrite = TRUE)
+  expect_equal(min_rows(10, cars), 10)
 })
 
 test_that('boost_tree can be fit with 1 predictor if validation is used', {