diff --git a/cpp/src/arrow/compute/exec/hash_join.cc b/cpp/src/arrow/compute/exec/hash_join.cc
index 5a9afaa5bdf..56d02dd3b11 100644
--- a/cpp/src/arrow/compute/exec/hash_join.cc
+++ b/cpp/src/arrow/compute/exec/hash_join.cc
@@ -103,12 +103,14 @@ class HashJoinBasicImpl : public HashJoinImpl {
     filter_ = std::move(filter);
     output_batch_callback_ = std::move(output_batch_callback);
     finished_callback_ = std::move(finished_callback);
-    local_states_.resize(num_threads + 1);  // +1 for calling thread + worker threads
+    // TODO(ARROW-15732)
+    // Each side of join might have an IO thread being called from.
+    local_states_.resize(GetCpuThreadPoolCapacity() + io::GetIOThreadPoolCapacity() + 1);
     for (size_t i = 0; i < local_states_.size(); ++i) {
       local_states_[i].is_initialized = false;
       local_states_[i].is_has_match_initialized = false;
     }
-    dict_probe_.Init(num_threads);
+    dict_probe_.Init(GetCpuThreadPoolCapacity() + io::GetIOThreadPoolCapacity() + 1);
 
     has_hash_table_ = false;
     num_batches_produced_.store(0);
diff --git a/cpp/src/arrow/compute/exec/hash_join_dict.cc b/cpp/src/arrow/compute/exec/hash_join_dict.cc
index ac1fbbaa3df..63d7d1143c9 100644
--- a/cpp/src/arrow/compute/exec/hash_join_dict.cc
+++ b/cpp/src/arrow/compute/exec/hash_join_dict.cc
@@ -566,7 +566,7 @@ Status HashJoinDictBuildMulti::PostDecode(
 }
 
 void HashJoinDictProbeMulti::Init(size_t num_threads) {
-  local_states_.resize(num_threads + 1);  // +1 for calling thread + worker threads
+  local_states_.resize(num_threads);
   for (size_t i = 0; i < local_states_.size(); ++i) {
     local_states_[i].is_initialized = false;
   }
@@ -576,6 +576,7 @@ bool HashJoinDictProbeMulti::BatchRemapNeeded(
     size_t thread_index, const SchemaProjectionMaps<HashJoinProjection>& proj_map_probe,
     const SchemaProjectionMaps<HashJoinProjection>& proj_map_build, ExecContext* ctx) {
   InitLocalStateIfNeeded(thread_index, proj_map_probe, proj_map_build, ctx);
+  DCHECK_LT(thread_index, local_states_.size());
   return local_states_[thread_index].any_needs_remap;
 }
 
diff --git a/cpp/src/arrow/compute/exec/util.cc b/cpp/src/arrow/compute/exec/util.cc
index 6e26927e40c..ef56e6128a3 100644
--- a/cpp/src/arrow/compute/exec/util.cc
+++ b/cpp/src/arrow/compute/exec/util.cc
@@ -321,7 +321,7 @@ size_t ThreadIndexer::operator()() {
 }
 
 size_t ThreadIndexer::Capacity() {
-  static size_t max_size = arrow::internal::ThreadPool::DefaultCapacity() + 1;
+  static size_t max_size = GetCpuThreadPoolCapacity() + io::GetIOThreadPoolCapacity() + 1;
   return max_size;
 }
 
diff --git a/r/tests/testthat/test-dplyr-join.R b/r/tests/testthat/test-dplyr-join.R
index 6f4adcb58ac..9d8e22596a6 100644
--- a/r/tests/testthat/test-dplyr-join.R
+++ b/r/tests/testthat/test-dplyr-join.R
@@ -310,27 +310,33 @@ test_that("summarize and join", {
   expect_equal(expected_col_names, res_col_names)
 })
 
-test_that("arrow dplyr query can join with tibble", {
-  # ARROW-14908
+test_that("arrow dplyr query can join two datasets", {
+  # ARROW-14908 and ARROW-15718
+  skip_if_not_available("dataset")
 
   # By default, snappy encoding will be used, and
   # Snappy has a UBSan issue: https://github.com/google/snappy/pull/148
   skip_on_linux_devel()
-  skip_if_not_available("dataset")
 
   dir_out <- tempdir()
-  write_dataset(iris, file.path(dir_out, "iris"))
-  species_codes <- data.frame(
-    Species = c("setosa", "versicolor", "virginica"),
-    code = c("SET", "VER", "VIR")
-  )
+
+  quakes %>%
+    select(stations, lat, long) %>%
+    group_by(stations) %>%
+    write_dataset(file.path(dir_out, "ds1"))
+
+  quakes %>%
+    select(stations, mag, depth) %>%
+    group_by(stations) %>%
+    write_dataset(file.path(dir_out, "ds2"))
 
   withr::with_options(
     list(arrow.use_threads = FALSE),
     {
-      iris <- open_dataset(file.path(dir_out, "iris"))
-      res <- left_join(iris, species_codes) %>% collect() # We should not segfault here.
-      expect_equal(nrow(res), 150)
+      res <- open_dataset(file.path(dir_out, "ds1")) %>%
+        left_join(open_dataset(file.path(dir_out, "ds2")), by = "stations") %>%
+        collect() # We should not segfault here.
+      expect_equal(nrow(res), 21872)
     }
   )
 })