Merge pull request #964 from mlr-org/integer_overflow

fix: integer overflow in remove constants

Author: mb706

NEWS.md

@@ -6,6 +6,7 @@
 * Added support for internal validation tasks to `PipeOpFeatureUnion`.
 * feat: `PipeOpLearnerCV` can reuse the cross-validation models during prediction by averaging their outputs (`resampling.predict_method = "cv_ensemble"`).
 * feat: `PipeOpRegrAvg` gets new `se_aggr` and `se_aggr_rho` hyperparameters and now allows various forms of SE aggregation.
+* Fix: `PipeOpRemoveConstants` now avoids integer overflow when evaluating relative tolerances for near-`integer.max` data.
 * Compatibility with new testthat version 3.3.0
 
 # mlr3pipelines 0.9.0

R/PipeOpRemoveConstants.R

@@ -109,7 +109,7 @@ is_constant_enough = function(x, ratio, rel_tol, abs_tol, na_ignore) {
 
     # now consider finite values: sort them and see if items that are
     # 'required_size - 1' steps away from each other differ by at most 'abs_tol' or 'rel_tol'
-    x = sort(x[is.finite(x)])
+    x = as.numeric(sort(x[is.finite(x)]))
     if (length(x) < required_size) {
       return(FALSE)
     }

tests/testthat/test_pipeop_removeconstants.R

@@ -70,3 +70,19 @@ test_that("PipeOpRemoveConstants removes expected cols", {
   test_dropping(minus.iris , minus.iris, list())
 
 })
+
+test_that("PipeOpRemoveConstants handles integer overflow constants", {
+  n = 5L
+  big = .Machine$integer.max
+  dt = data.table(
+    const = c(rep.int(big, n - 1L), big - 1L),
+    vary = c(big, big - 10L, big - 20L, big - 30L, big - 40L),
+    target = factor(c("a", "a", "b", "b", "b"))
+  )
+
+  task = TaskClassif$new("overflow", dt, target = "target")
+  po = PipeOpRemoveConstants$new(param_vals = list(abs_tol = 0))
+
+  result = po$train(list(task))[[1]]
+  expect_false("const" %in% result$feature_names)
+})