🔴 Non-daemon background thread can prevent process exit indefinitely

Changing daemon=True to daemon=False means that if a background evaluator hangs (e.g., a network call that never completes, a deadlock, or an unresponsive external service), the Python process will be unable to exit. The main thread and all other work may be done, but the interpreter will block waiting for the non-daemon thread to finish — with no timeout on the thread's actual work (ctx.run(anyio.run, _run) at line 174 has no time bound).

The existing wait_for_evaluations() function (pydantic_evals/pydantic_evals/_online.py:566) already provides an explicit opt-in mechanism (with a configurable timeout) for users who want to ensure evaluations complete before exiting. Making threads non-daemon is redundant for that use case and introduces a hang risk for users who don't call wait_for_evaluations() — their process will silently refuse to exit if any evaluation thread is stuck.

The change from daemon=True to daemon=False in dispatch_in_background_thread (line 179 of _online.py) prevents process exit if any background evaluator thread hangs. The function dispatch_in_background_thread is called from online.py:739 when there is no running event loop, to fire-and-forget evaluation work.

The stated goal is to avoid losing evaluation results when the main program exits quickly. However, non-daemon threads will block process exit indefinitely if any evaluator hangs (no timeout on the thread work itself at line 174: ctx.run(anyio.run, _run)).

Better approaches:
1. Keep daemon=True (the original) and document that users should call wait_for_evaluations() before exit to ensure results are flushed. This is already the documented pattern.
2. If non-daemon threads are truly desired, add an atexit handler that joins threads with a bounded timeout so the process can still exit.
3. Register an atexit hook that calls wait_for_evaluations with a reasonable timeout, keeping daemon=True so the process can always exit after the timeout.

Was this helpful? React with 👍 or 👎 to provide feedback.

🚩 No atexit handler or thread-level timeout to complement the daemon=False change

With daemon=False, the process relies entirely on the user calling wait_for_evaluations() to drain threads before exit. But wait_for_evaluations (_online.py:566-589) is opt-in and only imposes a timeout on thread.join(), not on the thread's work itself (ctx.run(anyio.run, _run) at line 174). If a user never calls wait_for_evaluations(), non-daemon threads will silently keep the process alive. An atexit handler that joins with a bounded timeout — or a per-thread cancellation mechanism — would make the daemon=False change safe without requiring user action.

(Refers to lines 166-187)

Was this helpful? React with 👍 or 👎 to provide feedback.