unilab.ipc.collector_error

Cross-process error propagation for collector subprocesses.

Uses a Pipe + ExceptionWrapper pattern (proven by PyTorch DataLoader and CPython’s ProcessPoolExecutor) to ensure collector tracebacks always reach the parent process — even when stderr is lost or interleaved.

Functions

collector_error_guard([error_conn, ...])

Context manager that catches all exceptions in collector subprocess.

create_error_pipe()

Create a unidirectional pipe for error reporting.

format_collector_death(exitcode[, ...])

Format a human-readable death report for a collector process.

Classes

ExceptionWrapper

Picklable exception + traceback for cross-process propagation.
class unilab.ipc.collector_error.ExceptionWrapper[source]

Bases: object

Picklable exception + traceback for cross-process propagation.

Stores the exception type and a pre-formatted traceback string (not exc_info — that would create reference cycles preventing GC of objects in the exception scope).

Parameters:

where (str)

__init__(where='in collector')[source]
Parameters:

where (str)

reraise()[source]
Return type:

None

unilab.ipc.collector_error.create_error_pipe()[source]

Create a unidirectional pipe for error reporting.

Returns (recv_conn, send_conn). Parent keeps recv, child gets send.

Return type:

tuple[Any, Any]

unilab.ipc.collector_error.collector_error_guard(error_conn=None, metrics_queue=None, stop_event=None, label='collector')[source]

Context manager that catches all exceptions in collector subprocess.

Sends a picklable ExceptionWrapper through the error pipe so the parent process can surface the full traceback. Also pushes to metrics_queue for fast-path detection by the training loop.

Parameters:
unilab.ipc.collector_error.format_collector_death(exitcode, traceback_text=None)[source]

Format a human-readable death report for a collector process.

Parameters:
Return type:

str