fix: Catch OSError with exponential backoff

nemo_run/run/torchx_backend/schedulers/slurm.py

@@ -228,7 +228,11 @@ def _cancel_existing(self, app_id: str) -> None:
         self.tunnel.run(f"scancel {app_id}", hide=False)
 
     def describe(self, app_id: str) -> Optional[DescribeAppResponse]:
-        job_dirs = _get_job_dirs()
+        try:
+            job_dirs = _get_job_dirs()
+        except OSError as e:
+            log.warning(f"Could not read job dirs file, returning UNKNOWN status for {app_id}: {e}")
+            return DescribeAppResponse(app_id=app_id, state=AppState.UNKNOWN)
         if app_id in job_dirs:
             _, tunnel_cfg, _ = job_dirs[app_id]
             self._initialize_tunnel(tunnel_cfg)
@@ -426,7 +430,7 @@ def _save_job_dir(
 
 def _get_job_dirs(retries: int = 5) -> dict[str, tuple[str, SSHTunnel | LocalTunnel, str]]:
     last_exc: OSError | None = None
-    for _ in range(retries):
+    for attempt in range(retries):
         try:
             with open(SLURM_JOB_DIRS, "rt") as f:
                 lines = f.readlines()
@@ -435,12 +439,17 @@ def _get_job_dirs(retries: int = 5) -> dict[str, tuple[str, SSHTunnel | LocalTun
             return {}
         except OSError as e:
             last_exc = e
-            time.sleep(1)
+            delay = min(2**attempt, 30)
+            log.warning(
+                f"OSError reading {SLURM_JOB_DIRS} (attempt {attempt + 1}/{retries}): {e}. "
+                f"Retrying in {delay}s..."
+            )
+            time.sleep(delay)
     else:
         if last_exc is not None:
             raise last_exc
-        raise OSError(
-            f"Failed to read SLURM job dirs from {SLURM_JOB_DIRS} after {retries} retries"
+        raise RuntimeError(
+            f"Failed to read {SLURM_JOB_DIRS} after {retries} retries, but no OSError was captured."
         )
 
     out = {}

test/run/torchx_backend/schedulers/test_slurm.py

@@ -340,8 +340,8 @@ def test_get_job_dirs():
 
 
 def test_get_job_dirs_retries_on_permission_error(tmp_path, mocker):
-    """Transient PermissionError should be retried; success on 3rd attempt returns data."""
-    mocker.patch("time.sleep")  # don't actually sleep in tests
+    """Transient PermissionError should be retried with exponential backoff; success on 3rd attempt returns data."""
+    mock_sleep = mocker.patch("time.sleep")
     mock_open = mocker.mock_open(read_data="")
     mock_open.side_effect = [
         PermissionError("[Errno 1] Operation not permitted"),
@@ -353,6 +353,8 @@ def test_get_job_dirs_retries_on_permission_error(tmp_path, mocker):
     result = _get_job_dirs(retries=5)
     assert result == {}
     assert mock_open.call_count == 3
+    # Exponential backoff: attempt 0 -> sleep(1), attempt 1 -> sleep(2)
+    assert mock_sleep.call_args_list == [mock.call(1), mock.call(2)]
 
 
 def test_get_job_dirs_raises_after_exhausting_retries(mocker):
@@ -364,6 +366,20 @@ def test_get_job_dirs_raises_after_exhausting_retries(mocker):
         _get_job_dirs(retries=3)
 
 
+def test_describe_returns_unknown_on_persistent_permission_error(slurm_scheduler, mocker):
+    """describe() should return UNKNOWN state when _get_job_dirs() raises OSError, not propagate."""
+    from torchx.specs import AppState
+
+    mocker.patch(
+        "nemo_run.run.torchx_backend.schedulers.slurm._get_job_dirs",
+        side_effect=PermissionError("[Errno 1] Operation not permitted"),
+    )
+
+    result = slurm_scheduler.describe("12345")
+    assert result is not None
+    assert result.state == AppState.UNKNOWN
+
+
 def test_schedule_with_dependencies(slurm_scheduler, slurm_executor):
     mock_request = mock.MagicMock()
     mock_request.cmd = ["sbatch", "--requeue", "--parsable"]