Python control surface#
What it exposes#
A small, explicit API to start logging for a ProcMesh, stream remote stdout/stderr back to your notebook/REPL, tune aggregation & verbosity, and perform a versioned sync flush before teardown.
Placement: There is exactly one
LogClientActor, and it runs inside the Python/driver process (your notebook/REPL). Python calls fan out to per-proc forwarders and back.
API surface#
LoggingMeshClient.spawn(instance, proc_mesh) -> PyPythonTask#
Spawns:
a single
LogClientActor(in the Python/driver process),per-proc
LogForwardActors (in remote procs),per-proc
LoggerRuntimeActors (Python logging runtime).
LoggingMeshClient.set_mode(instance, stream_to_client: bool, aggregate_window_sec: Optional[int], level: int) -> None#
Turns streaming on/off (
stream_to_client).Sets aggregation window (seconds) only if streaming is enabled.
Sets Python logging level in the per-proc
LoggerRuntimeActor.
Constraint: if
aggregate_window_secisSomewhilestream_to_clientisFalse,set_moderaises a runtime error (see source).
LoggingMeshClient.flush(instance) -> PyPythonTask#
Performs a versioned sync flush across all forwarders so the caller deterministically waits until “all logs up to now” are delivered. See: Client actor → Barrier protocol and Forwarder internals.
Minimal example#
from monarch._rust_bindings.monarch_hyperactor.logging import LoggingMeshClient
import logging
client = await LoggingMeshClient.spawn(instance, proc_mesh) # one client actor in driver
client.set_mode(instance, stream_to_client=True, aggregate_window_sec=3, level=logging.INFO)
# … run workload producing stdout/stderr in remote procs …
await client.flush(instance) # barrier: all logs observed up to this point
