[lucasjia-aws]

Summary

Feature Store pandas ingestion currently creates a multiprocessing pool with the platform default start method and defines its initializer inside the calling method. Fork mode can deadlock on recent macOS releases, while the nested initializer is not safe under spawn.

This change creates the pool from an explicit spawn context and moves the signal initializer to module scope so it can be pickled. Existing asynchronous result, close, and join behavior is preserved.

Testing

• Added coverage for spawn-context selection, process count, initializer identity and behavior, asynchronous result retention, and pool cleanup.
• Ran the complete affected Feature Store unit test file: 35 passed.
• Black, Flake8, and git diff --check passed.
• Built the sagemaker-mlops wheel and source distribution successfully.
• No AWS integration tests or AWS calls were run.

Related change

The companion V2 fix is published on #6465.

Fixes #5312

[github-actions]

🤖 AI code review

This PR switches Feature Store pandas multi-process ingestion from the platform-default multiprocessing start method to an explicit spawn context, and moves the SIGINT-ignoring worker initializer from a nested closure to module scope so it is picklable under spawn. This directly addresses the fork-based deadlock reported in #5312. The change is correct: the staticmethod target and module-level initializer are picklable, the pre-existing from multiprocessing import Pool import is still required (used as the type annotation at line 71), and wait()'s close/join behavior is preserved. Tests cover spawn-context selection, process count, initializer identity/behavior, and pool cleanup, and are consistent with the implementation. No issues found.

_{Reviewed commit 3a6c142. Automated review; verify before acting.}