azure-ai-evaluation: maximum recursion depth exceeded

[Wixee]

• Package Name: azure.ai.evaluation
• Package Version: 1.13.7
• Operating System: Windows 11
• Python Version: 3.11.11

Describe the bug
Please find the details in #43040

In a scheduled job / long-running environment, azure-ai-evaluation can fail with a RecursionError that is triggered by the SDK’s stdout/stderr interception logic. This ultimately makes evaluate(...) fail (surfaced as an internal evaluation error).

Expected behavior

evaluate(...) completes and returns results/metrics reliably in batch/concurrent runs.

Actual behavior

Evaluation occasionally fails due to RecursionError originating from the SDK’s stdout/stderr wrapper, which recursively forwards write() calls through a deep chain of wrapped streams.

Suspected root cause

The legacy logging layer replaces sys.stdout and sys.stderr with NodeLogWriter via NodeLogManager. NodeLogWriter.write() scrubs credentials and then forwards output to a “previous” stream self._prev_out.write(s) when no node context is present.

Under concurrency / multi-threaded execution, sys.stdout may be wrapped multiple times (wrapper → wrapper → wrapper …), because each NodeLogManager captures the current sys.stdout as its prev_stdout. When later writing output (e.g., run summary / streaming output), the call chain through self._prev_out.write(...) becomes deep enough to overflow recursion depth.

Relevant code locations:

stdout/stderr replacement: _logging.py:162-188
recursive forwarding in write(): _logging.py:245-256
run streaming writes to sys.stdout: _run_submitter.py:204-263
NodeLogManager used per-line (concurrent): _engine.py:443

Why this matters

This can make evaluation flaky/non-deterministic in production scheduled jobs, especially when concurrency is enabled.

[github-actions]

Thanks for the feedback! We are routing this to the appropriate team for follow-up. cc @luigiw @needuv @singankit.