GH-55605: [R] Don't abort when comparing R extension types off the R thread

r/src/extension-impl.cpp

@@ -37,7 +37,8 @@ bool RExtensionType::ExtensionEquals(const arrow::ExtensionType& other) const {
   }
 
   // With any ambiguity, we need to materialize the R6 instance and call its
-  // ExtensionEquals method. We can't do this on the non-R thread.
+  // ExtensionEquals method. This is only possible from the R thread or from
+  // inside RunWithCapturedR().
   arrow::Result<bool> result = SafeCallIntoR<bool>(
       [&]() {
         cpp11::environment instance = r6_instance();
@@ -52,8 +53,14 @@ bool RExtensionType::ExtensionEquals(const arrow::ExtensionType& other) const {
       },
       "RExtensionType$ExtensionEquals()");
 
+  // An error here means R could not be called from this thread (e.g., when a
+  // Dataset is scanned on worker threads). Throwing would terminate the process,
+  // because this is not the R thread (GH-55605). Fall back to the default
+  // ExtensionEquals() behaviour instead: the serialized metadata differs, so the
+  // types are not equal. If the R method itself errored, RunWithCapturedR()
+  // re-raises that error.
   if (!result.ok()) {
-    throw std::runtime_error(result.status().message());
+    return false;
   }
 
   return result.ValueUnsafe();

r/tests/testthat/test-extension.R

@@ -172,6 +172,32 @@ test_that("extension subclasses can override the ExtensionEquals method", {
   unregister_extension_type("some_extension_subclass")
 })
 
+test_that("comparing extension types on a non-R thread does not abort (GH-55605)", {
+  skip_if_not_available("dataset")
+
+  unit_type <- function(unit) {
+    new_extension_type(int32(), "arrow.test.unit", charToRaw(unit))
+  }
+  register_extension_type(unit_type("m"))
+  on.exit(unregister_extension_type("arrow.test.unit"))
+
+  tf <- tempfile()
+  dir.create(tf)
+  on.exit(unlink(tf, recursive = TRUE), add = TRUE)
+
+  for (unit in c("m", "km")) {
+    write_ipc_file(
+      arrow_table(x = unit_type(unit)$WrapArray(Array$create(1:3))),
+      file.path(tf, paste0(unit, ".arrow"))
+    )
+  }
+  ds <- open_dataset(tf, format = "arrow")
+
+  # The parts have different extension metadata, so they can't be combined
+  expect_error(dplyr::collect(ds), "not permitted")
+  expect_error(Scanner$create(ds)$ToTable(), "not permitted")
+})
+
 test_that("vctrs extension type works", {
   custom_vctr <- vctrs::new_vctr(
     1:4,