DataLab crashes (access violation) when a computation requested remotely fails #404

Describe the bug

When a computation requested through remote control fails, DataLab crashes with a native access violation (0xC0000005) instead of reporting the error. This happens with XML-RPC (calc) and with the Web API (POST /api/v1/calc). The client gets ConnectionRefusedError (XML-RPC) or ConnectionResetError (Web API).

The exception raised by the computation is caught by the qt_try_except decorator of BaseProcessor.run_feature, which calls qt_handle_error_message: the crash occurs in its QMessageBox.critical(...) call.

Windows fatal exception: access violation

Current thread 0x0001c72c (most recent call first):
  File "C:\Dev\Projets\SigimaX\sigimax\utils\qthelpers.py", line 359 in qt_handle_error_message
  File "C:\Dev\Projets\DataLab\datalab\utils\qthelpers.py", line 128 in method_wrapper
  File "C:\Dev\Projets\DataLab\datalab\gui\main.py", line 1879 in calc
  File "C:\Dev\Projets\DataLab\datalab\gui\main.py", line 131 in method_wrapper
  ...

To Reproduce

"""DataLab crashes when a computation requested remotely fails"""

import os

import numpy as np

if __name__ == "__main__":
    # Imported here: the DataLab process must not import the test helpers (pytest)
    from datalab.env import execenv
    from datalab.tests import datalab_in_background_context

    os.environ.setdefault("QT_QPA_PLATFORM", "offscreen")
    execenv.unattended = True
    with datalab_in_background_context() as proxy:
        proxy.add_signal("One point", np.array([0.0]), np.array([1.0]))
        proxy.select_objects([1], "signal")
        # The derivative of a one-point signal raises IndexError in Sigima
        proxy.calc("derivative")

Result:

• the script fails with ConnectionRefusedError: [WinError 10061];
• the DataLab process exits with code 3221225477 (0xC0000005);
• ~/.DataLab_v1/.DataLab_faulthandler.log contains the traceback above.

The same crash occurs with the Web API, started with DATALAB_WEBAPI_ENABLED=1, when calling POST /api/v1/calc with {"name": "derivative"} on the same selection (checked on the branch of #402, which does not change this code path; the XML-RPC case was checked on both branches).

Expected behavior

DataLab keeps running, and the error is reported to the remote client: as an xmlrpc.client.Fault with XML-RPC, as an HTTP error with the Web API. A remote request should not open a modal message box.

Installation information

• Python package (development checkout), develop branch at c800d5e
• DataLab 1.3.0, Sigima 1.3.0, SigimaX 1.1.0, PlotPy 2.11.1, guidata 3.16.0, PyQt5 5.15.11 (Qt 5.15.2), FastAPI 0.136.0, uvicorn 0.45.0, Python 3.11.6, NumPy 2.4.4, Windows 10.0.26200

Additional context

• Observed with QT_QPA_PLATFORM=offscreen in unattended mode, i.e. a background DataLab as used by tests and automation. Not checked with a visible window, where the modal message box would at least block the remote call until it is closed, without reporting the error to the client.
• The test suite cannot catch this: when the DataLab process imports pytest (e.g. through datalab.tests), is_running_tests() makes qt_try_except re-raise the exception instead of showing the message box. With Fix remote file loading intermittently ignored at DataLab startup #402, the error then reaches the client as xmlrpc.client.Fault: <class 'IndexError'>:index 0 is out of bounds for axis 0 with size 0.
• Found while investigating Remote load_from_files / load_from_directory intermittently ignored #398 (see Fix remote file loading intermittently ignored at DataLab startup #402), which does not change this behavior: the crash also occurs on its branch.