test: mark PTY-dependent hooks tests as xfail on macOS (#821)

python/packages/jumpstarter/jumpstarter/exporter/hooks_test.py

@@ -1,4 +1,5 @@
 import os
+import sys
 from contextlib import nullcontext
 from unittest.mock import AsyncMock, MagicMock, patch
 
@@ -18,6 +19,16 @@
 
 pytestmark = pytest.mark.anyio
 
+# Tests that spawn real subprocesses via PTY and assert on captured logger
+# output are flaky on macOS due to a PTY kernel buffer timing race condition.
+# See https://github.com/jumpstarter-dev/jumpstarter/issues/821
+# Targeted for proper fix in 0.10.0.
+macos_pty_xfail = pytest.mark.xfail(
+    condition=sys.platform == "darwin",
+    reason="PTY output race condition on macOS (#821)",
+    strict=False,
+)
+
 
 class _PtyTracker:
     """Tracks PTY fd and EOF state for drain tests that need to intercept
@@ -196,6 +207,7 @@ async def test_hook_timeout(self, lease_scope) -> None:
         assert "timed out after 1 seconds" in str(exc_info.value)
         assert exc_info.value.on_failure == "exit"
 
+    @macos_pty_xfail
     async def test_hook_environment_variables(self, lease_scope) -> None:
         hook_config = HookConfigV1Alpha1(
             before_lease=HookInstanceConfigV1Alpha1(
@@ -210,6 +222,7 @@ async def test_hook_environment_variables(self, lease_scope) -> None:
             assert any("LEASE_NAME=test-lease-123" in call for call in info_calls)
             assert any("CLIENT_NAME=test-client" in call for call in info_calls)
 
+    @macos_pty_xfail
     async def test_real_time_output_logging(self, lease_scope) -> None:
         """Test that hook output is logged in real-time at INFO level."""
         hook_config = HookConfigV1Alpha1(
@@ -227,6 +240,7 @@ async def test_real_time_output_logging(self, lease_scope) -> None:
             assert any("Line 2" in call for call in info_calls)
             assert any("Line 3" in call for call in info_calls)
 
+    @macos_pty_xfail
     async def test_post_lease_hook_execution_on_completion(self, lease_scope) -> None:
         """Test that post-lease hook executes when called directly."""
         hook_config = HookConfigV1Alpha1(
@@ -337,6 +351,7 @@ async def test_successful_hook_returns_none(self, lease_scope) -> None:
         result = await executor.execute_before_lease_hook(lease_scope)
         assert result is None
 
+    @macos_pty_xfail
     async def test_exec_bash(self, lease_scope) -> None:
         """Test that exec=/bin/bash allows bash-specific syntax.
 
@@ -358,6 +373,7 @@ async def test_exec_bash(self, lease_scope) -> None:
             info_calls = [str(call) for call in mock_logger.info.call_args_list]
             assert any("BASH_OK: world" in call for call in info_calls)
 
+    @macos_pty_xfail
     async def test_exec_python3(self, lease_scope) -> None:
         """Test that exec=python3 runs inline Python.
 
@@ -380,6 +396,7 @@ async def test_exec_python3(self, lease_scope) -> None:
             # Expected total: 0 + 1 + 4 + 9 == 14
             assert any("PYTHON_OK: 14" in call for call in info_calls)
 
+    @macos_pty_xfail
     async def test_script_file_sh(self, lease_scope, tmp_path) -> None:
         """Test that a .sh file auto-detects /bin/sh as interpreter."""
         script_file = tmp_path / "hook_script.sh"
@@ -402,6 +419,7 @@ async def test_script_file_sh(self, lease_scope, tmp_path) -> None:
             debug_calls = [str(call) for call in mock_logger.debug.call_args_list]
             assert any("Executing script file" in call for call in debug_calls)
 
+    @macos_pty_xfail
     async def test_script_file_py_autodetects_python(self, lease_scope, tmp_path) -> None:
         """Test that a .py file auto-detects the exporter's Python as interpreter."""
         import sys
@@ -428,6 +446,7 @@ async def test_script_file_py_autodetects_python(self, lease_scope, tmp_path) ->
             # Verify it used the exporter's own Python interpreter
             assert any(sys.executable in call for call in debug_calls)
 
+    @macos_pty_xfail
     async def test_script_file_py_exec_override(self, lease_scope, tmp_path) -> None:
         """Test that explicit exec overrides .py auto-detection."""
         script_file = tmp_path / "hook_script.py"
@@ -451,6 +470,7 @@ async def test_script_file_py_exec_override(self, lease_scope, tmp_path) -> None
             debug_calls = [str(call) for call in mock_logger.debug.call_args_list]
             assert not any("Auto-detected" in call for call in debug_calls)
 
+    @macos_pty_xfail
     async def test_noninteractive_environment(self, lease_scope) -> None:
         """Test that hooks receive noninteractive environment variables.
 
@@ -710,6 +730,7 @@ async def test_drain_handles_oserror_gracefully(self) -> None:
         assert output_lines == []
         assert drained == 0
 
+    @macos_pty_xfail
     async def test_drain_captures_output_without_trailing_newline(self, lease_scope) -> None:
         """Verify output without a trailing newline is still captured."""
         hook_config = HookConfigV1Alpha1(
@@ -726,6 +747,7 @@ async def test_drain_captures_output_without_trailing_newline(self, lease_scope)
             info_calls = [str(call) for call in mock_logger.info.call_args_list]
             assert any("NO_NEWLINE_OUTPUT" in call for call in info_calls)
 
+    @macos_pty_xfail
     async def test_drain_reads_data_remaining_in_pty_buffer(self, lease_scope) -> None:
         """Verify the drain loop inside read_pty_output reads data left in the
         PTY kernel buffer after the main read loop exits.
@@ -790,6 +812,7 @@ def os_read_with_drain_data(fd, size):
             info_calls = [str(call) for call in mock_logger.info.call_args_list]
             assert any("DRAIN_CAPTURED" in call for call in info_calls)
 
+    @macos_pty_xfail
     async def test_drain_select_oserror_exits_gracefully(self, lease_scope) -> None:
         """Verify the drain loop exits gracefully when select.select() raises
         OSError (e.g. fd closed during drain).
@@ -826,6 +849,7 @@ def select_with_oserror(rlist, wlist, xlist, timeout=None):
             info_calls = [str(call) for call in mock_logger.info.call_args_list]
             assert any("SELECT_ERROR_TEST" in call for call in info_calls)
 
+    @macos_pty_xfail
     async def test_drain_select_valueerror_exits_gracefully(self, lease_scope) -> None:
         """Verify the drain loop exits gracefully when select.select() raises
         ValueError (e.g. negative fd).
@@ -860,6 +884,7 @@ def select_with_valueerror(rlist, wlist, xlist, timeout=None):
             info_calls = [str(call) for call in mock_logger.info.call_args_list]
             assert any("VALUEERROR_TEST" in call for call in info_calls)
 
+    @macos_pty_xfail
     async def test_drain_exits_when_deadline_exceeded_before_select(self, lease_scope) -> None:
         """Verify the drain loop exits when the deadline is exceeded between the
         while condition and the remaining-time check (line: if remaining <= 0).
@@ -894,6 +919,7 @@ async def test_drain_exits_when_deadline_exceeded_before_select(self, lease_scop
             # exited early due to remaining <= 0 before select could run
             assert not any("SHOULD_NOT_APPEAR" in call for call in info_calls)
 
+    @macos_pty_xfail
     async def test_drain_exception_is_suppressed(self, lease_scope) -> None:
         """Verify that an unexpected exception raised during the drain is caught
         by the except-Exception handler and does not propagate to the caller.
@@ -928,6 +954,7 @@ def flush_lines_with_drain_error(buffer, output_lines):
             result = await executor.execute_before_lease_hook(lease_scope)
             assert result is None
 
+    @macos_pty_xfail
     async def test_drain_retries_empty_select_then_captures_data(self, lease_scope) -> None:
         """Verify that the drain retries after empty select() calls and still
         captures data that arrives later.
@@ -969,6 +996,7 @@ def select_with_delayed_ready(rlist, wlist, xlist, timeout=None):
             info_calls = [str(call) for call in mock_logger.info.call_args_list]
             assert any("DELAYED_DRAIN_OK" in call for call in info_calls)
 
+    @macos_pty_xfail
     async def test_drain_terminates_after_max_empty_polls(self, lease_scope) -> None:
         """Verify the drain loop terminates after DRAIN_MAX_EMPTY_POLLS
         consecutive empty select() results.
@@ -1006,6 +1034,7 @@ def select_always_empty(rlist, wlist, xlist, timeout=None):
             info_calls = [str(call) for call in mock_logger.info.call_args_list]
             assert any("MAX_EMPTY_TEST" in call for call in info_calls)
 
+    @macos_pty_xfail
     async def test_drain_empty_counter_resets_on_data(self, lease_scope) -> None:
         """Verify the consecutive empty poll counter resets when data arrives.
 
@@ -1064,6 +1093,7 @@ async def test_exec_default_is_none(self) -> None:
 class TestHookExecutorPRRegressions:
     """Regression tests for issues reported during PR review of hooks feature."""
 
+    @macos_pty_xfail
     async def test_infrastructure_messages_at_debug_not_info(self, lease_scope) -> None:
         """Issue A1: Hook infrastructure messages should be at DEBUG, not INFO.