Merge pull request #1564 from exploratory-io/enable/chaid-perf-always-on

test(chaid): always run perf guardrails with 2x budgets

Author: hidekoji

tests/testthat/test_chaid_perf.R

@@ -1,7 +1,6 @@
-# CHAID performance guardrails. Skipped by default; run with
-#   EXPLORATORY_RUN_PERF=1 Rscript -e '...test_file("tests/testthat/test_chaid_perf.R")'
-# Budgets are generous relative to observed times (10k x 20 ~0.4s,
-# 100k x 50 ~3s on an M1 Mac) so they catch regressions, not noise.
+# CHAID performance guardrails. Always run in the normal testthat suite.
+# Budgets are ~2x measured Daily-machine times so parallel/loaded runs still
+# pass: Mac M5 ~0.6/4.4/1.1s, Win Azure 4-core ~2.5/17/6.4s for the three cases.
 
 make_chaid_data <- function(n, p, seed = 1) {
   set.seed(seed)
@@ -19,29 +18,26 @@ make_chaid_data <- function(n, p, seed = 1) {
 }
 
 test_that('CHAID fits 10k x 20 within budget', {
-  skip_if(Sys.getenv('EXPLORATORY_RUN_PERF') == '', 'perf tests are opt-in')
   df <- make_chaid_data(10000, 20)
   elapsed <- system.time(suppressWarnings(
     chaid_fit(df, target = 'target', predictors = paste0('v', 1:20),
               max_depth = 4, min_split = 100, min_bucket = 30)
   ))[['elapsed']]
-  expect_lt(elapsed, 10)
+  expect_lt(elapsed, 20)
 })
 
 test_that('CHAID fits 100k x 50 within budget', {
-  skip_if(Sys.getenv('EXPLORATORY_RUN_PERF') == '', 'perf tests are opt-in')
   df <- make_chaid_data(100000, 50)
   elapsed <- system.time(suppressWarnings(
     chaid_fit(df, target = 'target', predictors = paste0('v', 1:50),
               max_depth = 4, min_split = 500, min_bucket = 100)
   ))[['elapsed']]
-  expect_lt(elapsed, 90)
+  expect_lt(elapsed, 180)
 })
 
 test_that('permutation importance on 20k x 20 stays fast', {
-  skip_if(Sys.getenv('EXPLORATORY_RUN_PERF') == '', 'perf tests are opt-in')
   # Before vectorizing prediction this took ~10 minutes (201 predictions, each a
-  # per-row tree walk). Budget is generous relative to the observed ~0.4s.
+  # per-row tree walk). Budget is 2x measured Win Daily (~6.4s) for parallel load.
   df <- make_chaid_data(20000, 20)
   predictors <- paste0('v', 1:20)
   model <- suppressWarnings(
@@ -51,5 +47,5 @@ test_that('permutation importance on 20k x 20 stays fast', {
   elapsed <- system.time(
     chaid_permutation_importance(model, df, 'target', predictors)
   )[['elapsed']]
-  expect_lt(elapsed, 30)
+  expect_lt(elapsed, 60)
 })