fix: handle None return from error handler in no-output background callbacks

[dajiaohuang]

Summary

When a background callback with no outputs raises an exception, the error handler is called. If the error handler returns None, the code was incorrectly setting output_value = NoUpdate(), which then triggered an InvalidCallbackReturnValue exception at the output validation step.

Fix

This change adds the same output_spec check that exists in the non-background callback path to the background callback path at _callback.py:635, ensuring that NoUpdate() is only set when there are actual outputs to update.

Testing

The issue includes a minimal reproducible example that can be used to verify the fix.

Related Issue

Fixes #3628

[camdecoster]

Thanks for the PR! Could you please look into the test failures and see what's going on?

[T4rk1n]

Works good, and the lcbc019 test covers it. One gap: the job-cancel branch a few lines above has the same NoUpdate() for no-output callbacks, so cancelling a no-output background callback still raises InvalidCallbackReturnValue. Can you fix that here too, since it's the same bug? Please also add a CHANGELOG entry under Fixed for #3628.

[dajiaohuang]

感谢你的反馈！你说得对，测试确实过度使用了 mock，我会简化测试方法。

[dajiaohuang]

Addressed the review feedback:

1. Cancel branch fix: The cancelled-job branch in _handle_rest_background_callback (dash/_callback.py:689) now applies the same output_spec guard that was added for the error-handler branch — it sets NoUpdate() only when output_spec is truthy (i.e., the callback has outputs), and leaves output_value = None for no-output callbacks. This prevents InvalidCallbackReturnValue from being raised when a no-output background callback is cancelled via cancel= or when its job is gone on poll.

2. Test simplification: Dropped the heavily-mocked test_mcpbg012_tasks_result_passes_output_spec (which monkeypatched 6 internals and copied the whole signature) and replaced it with two integration-style tests that register a real no-output background callback via @app.callback and exercise the actual MCP dispatch path:

   • test_mcpbg012_no_output_task_result_does_not_raise — polls tasks/result after a no-output callback completes successfully.
   • test_mcpbg013_cancel_no_output_task_does_not_raise — calls tasks/cancel and then tasks/result on a no-output task to cover the cancel branch.

3. CHANGELOG: Added a Fixed entry under Unreleased referencing fix: handle None return from error handler in no-output background callbacks #4012 / [BUG] Exceptions in Background Callback Cause InvalidCallbackReturnValue When Using Custom Error Handler #3628.

[dajiaohuang]

Fixed the current-head MCP unit failures in f5cde05. No-output callback IDs are hashes, not component.property IDs, so tasks/result must not split them. It also now formats the framework's normal PreventUpdate result instead of returning an MCP internal error. The existing cancel output_spec guard remains intact.

Validation: reproduced both failures from run 38700521782, then passed all 198 MCP unit tests and a retained headless browser/HTTP cancellation regression. Used the generated components and renderer from that run's official dash-packages artifact; no generated files were edited. Local environment: Windows/Python 3.13, unrelated pytest plugin autoload disabled. A browser completion variant timed out waiting for its Windows background process, so completion is covered locally by the real-dispatch unit test, not claimed as browser-validated.

The benchmark process exited successfully in run 38700521890; that job failed when its comment POST received HTTP 403 (Resource not accessible by integration). No benchmark thresholds or workflow permissions changed.

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud