diff --git a/lib/matplotlib/tests/test_backends_interactive.py b/lib/matplotlib/tests/test_backends_interactive.py
index dbc0f0928b74..6048b2647b17 100644
--- a/lib/matplotlib/tests/test_backends_interactive.py
+++ b/lib/matplotlib/tests/test_backends_interactive.py
@@ -660,7 +660,7 @@ def _test_figure_leak():
                     reason="appveyor tests fail; gh-22988 suggests reworking")
 @pytest.mark.parametrize("env", _get_testable_interactive_backends())
 @pytest.mark.parametrize("time_mem", [(0.0, 2_000_000), (0.1, 30_000_000)])
-def test_figure_leak_20490(env, time_mem):
+def test_figure_leak_20490(env, time_mem, request):
     pytest.importorskip("psutil", reason="psutil needed to run this test")
 
     # We haven't yet directly identified the leaks so test with a memory growth
@@ -669,9 +669,10 @@ def test_figure_leak_20490(env, time_mem):
     if env["MPLBACKEND"] == "wx":
         pytest.skip("wx backend is deprecated; tests failed on appveyor")
 
-    if env["MPLBACKEND"] == "macosx" or (
-            env["MPLBACKEND"] == "tkagg" and sys.platform == 'darwin'
-    ):
+    if env["MPLBACKEND"] == "macosx":
+        request.node.add_marker(pytest.mark.xfail(reason="macosx backend is leaky"))
+
+    if env["MPLBACKEND"] == "tkagg" and sys.platform == "darwin":
         acceptable_memory_leakage += 11_000_000
 
     result = _run_helper(
@@ -815,12 +816,12 @@ def custom_signal_handler(signum, frame):
     ('show', {'block': True}),
     ('pause', {'interval': 10})
 ])
-def test_other_signal_before_sigint(env, target, kwargs):
+def test_other_signal_before_sigint(env, target, kwargs, request):
     backend = env.get("MPLBACKEND")
     if not backend.startswith(("qt", "macosx")):
         pytest.skip("SIGINT currently only tested on qt and macosx")
-    if backend == "macosx" and target == "show":
-        pytest.xfail("test currently failing for macosx + show()")
+    if backend == "macosx":
+        request.node.add_marker(pytest.mark.xfail(reason="macosx backend is buggy"))
     proc = _WaitForStringPopen(
         [sys.executable, "-c",
          inspect.getsource(_test_other_signal_before_sigint_impl) +
diff --git a/lib/matplotlib/tests/test_pickle.py b/lib/matplotlib/tests/test_pickle.py
index 8d1e20b4cb93..e222a495e437 100644
--- a/lib/matplotlib/tests/test_pickle.py
+++ b/lib/matplotlib/tests/test_pickle.py
@@ -143,7 +143,7 @@ def test_pickle_load_from_subprocess(fig_test, fig_ref, tmp_path):
     proc = subprocess_run_helper(
         _pickle_load_subprocess,
         timeout=60,
-        extra_env={'PICKLE_FILE_PATH': str(fp)}
+        extra_env={'PICKLE_FILE_PATH': str(fp), 'MPLBACKEND': 'Agg'}
     )
 
     loaded_fig = pickle.loads(ast.literal_eval(proc.stdout))
diff --git a/lib/matplotlib/tests/test_ticker.py b/lib/matplotlib/tests/test_ticker.py
index c95526ecf987..9d08e335dbdd 100644
--- a/lib/matplotlib/tests/test_ticker.py
+++ b/lib/matplotlib/tests/test_ticker.py
@@ -1640,22 +1640,36 @@ def test_latex(self, is_latex, usetex, expected):
             assert fmt.format_pct(50, 100) == expected
 
 
-def test_locale_comma():
-    currentLocale = locale.getlocale()
+def _impl_locale_comma():
     try:
         locale.setlocale(locale.LC_ALL, 'de_DE.UTF-8')
-        ticks = mticker.ScalarFormatter(useMathText=True, useLocale=True)
-        fmt = '$\\mathdefault{%1.1f}$'
-        x = ticks._format_maybe_minus_and_locale(fmt, 0.5)
-        assert x == '$\\mathdefault{0{,}5}$'
-        # Do not change , in the format string
-        fmt = ',$\\mathdefault{,%1.1f},$'
-        x = ticks._format_maybe_minus_and_locale(fmt, 0.5)
-        assert x == ',$\\mathdefault{,0{,}5},$'
     except locale.Error:
-        pytest.skip("Locale de_DE.UTF-8 is not supported on this machine")
-    finally:
-        locale.setlocale(locale.LC_ALL, currentLocale)
+        print('SKIP: Locale de_DE.UTF-8 is not supported on this machine')
+        return
+    ticks = mticker.ScalarFormatter(useMathText=True, useLocale=True)
+    fmt = '$\\mathdefault{%1.1f}$'
+    x = ticks._format_maybe_minus_and_locale(fmt, 0.5)
+    assert x == '$\\mathdefault{0{,}5}$'
+    # Do not change , in the format string
+    fmt = ',$\\mathdefault{,%1.1f},$'
+    x = ticks._format_maybe_minus_and_locale(fmt, 0.5)
+    assert x == ',$\\mathdefault{,0{,}5},$'
+
+
+def test_locale_comma():
+    # On some systems/pytest versions, `pytest.skip` in an exception handler
+    # does not skip, but is treated as an exception, so directly running this
+    # test can incorrectly fail instead of skip.
+    # Instead, run this test in a subprocess, which avoids the problem, and the
+    # need to fix the locale after.
+    proc = mpl.testing.subprocess_run_helper(_impl_locale_comma, timeout=60,
+                                             extra_env={'MPLBACKEND': 'Agg'})
+    skip_msg = next((line[len('SKIP:'):].strip()
+                     for line in proc.stdout.splitlines()
+                     if line.startswith('SKIP:')),
+                    '')
+    if skip_msg:
+        pytest.skip(skip_msg)
 
 
 def test_majformatter_type():