fix(system-services-monitor): address architectural review findings

- Makefile: set MODULE_NAME := $(PYTHON_PACKAGE_NAME). common.mk defaults
  MODULE_NAME to the hyphenated dir basename, so python.mk's
  `coverage run --source=$(MODULE_NAME)` never matched the importable
  package and reported 0% coverage. Matches preflight-checks/nccl-allreduce.
- pyproject.toml: relax python floor ^3.13 -> ^3.10 to match siblings
  (nccl-allreduce, dcgm-diag). Package uses no 3.11+/3.13-only syntax or
  stdlib (verified: no tomllib, PEP 695 type params, StrEnum, datetime.UTC,
  ExceptionGroup, etc.).
- Dockerfile: pin poetry==2.3.3 + poetry-plugin-export==1.10.0 per
  .versions.yaml (was poetry==1.8.2 with no export plugin while running
  `poetry export`). Export invocation already matches the poetry-2.x
  sibling Dockerfiles.
- cli.py: rename `exit = Event()` -> `stop_event` to stop shadowing the
  builtin (CodeRabbit flagged this on the prior #891).
- platform_connector/event_processor.py: guard the UDS dial with
  _is_platform_connector_socket_present() before/between retries, mirroring
  gpu-health-monitor; log-and-skip instead of noisy gRPC stack traces when
  the socket isn't present yet. Adds events_sent_skipped_pc_unavailable
  counter.

Author: dmvevents

health-monitors/system-services-monitor/Dockerfile

@@ -19,7 +19,7 @@ FROM public.ecr.aws/docker/library/python:3.13-bookworm AS build
 ARG VERSION
 
 RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install poetry==1.8.2
+    pip install poetry==2.3.3 poetry-plugin-export==1.10.0
 
 ENV POETRY_NO_INTERACTION=1 \
     POETRY_VIRTUALENVS_IN_PROJECT=1 \

health-monitors/system-services-monitor/Makefile

@@ -25,6 +25,10 @@ HAS_DOCKER := 1
 
 # Python package name (with underscores, not hyphens)
 PYTHON_PACKAGE_NAME := system_services_monitor
+# Coverage in make/python.mk uses --source=$(MODULE_NAME); common.mk defaults
+# MODULE_NAME to the (hyphenated) directory basename, which does not match the
+# importable package, yielding 0% coverage. Pin it to the package name.
+MODULE_NAME := $(PYTHON_PACKAGE_NAME)
 
 # system-services-monitor specific settings (Python module)
 CLEAN_EXTRA_FILES := system_services_monitor.egg-info

health-monitors/system-services-monitor/pyproject.toml

@@ -6,7 +6,7 @@ authors = ["Community Contributors"]
 readme = "README.md"
 
 [tool.poetry.dependencies]
-python = "^3.13"
+python = "^3.10"
 click = "^8.3.1"
 grpcio = "^1.78.0"
 prometheus-client = "^0.24.1"

health-monitors/system-services-monitor/system_services_monitor/cli.py

@@ -66,7 +66,7 @@ def cli(
     processing_strategy,
     verbose,
 ):
-    exit = Event()
+    stop_event = Event()
 
     # Resolve node name from CLI or environment
     if node_name is None:
@@ -108,7 +108,7 @@ def cli(
     prom_server, t = start_http_server(port)
 
     def process_exit_signal(signum, frame):
-        exit.set()
+        stop_event.set()
         prom_server.shutdown()
         t.join()
 
@@ -126,7 +126,7 @@ def process_exit_signal(signum, frame):
         enable_fabric_check=enable_fabric_check,
     )
 
-    watcher.start(exit)
+    watcher.start(stop_event)
 
 
 if __name__ == "__main__":

health-monitors/system-services-monitor/system_services_monitor/platform_connector/event_processor.py

@@ -20,6 +20,7 @@
 
 import dataclasses
 import logging as log
+import os
 import threading
 from time import sleep
 from typing import List
@@ -162,15 +163,44 @@ def health_check_completed(self, results: List[CheckResult]) -> None:
                         for key in pending_cache_updates:
                             self.entity_cache.pop(key, None)
 
+    def _is_platform_connector_socket_present(self) -> bool:
+        # platform-connector removes the socket file on shutdown and on
+        # startup before binding, so file-presence is a faithful proxy
+        # for "PC is up" on this node.
+        return os.path.exists(self._socket_path)
+
     def send_health_event_with_retries(self, health_events: list[platformconnector_pb2.HealthEvent]) -> bool:
         """Send health events to the platform connector with retries.
 
+        If the platform-connector Unix socket is absent at send time the send
+        is skipped immediately (no gRPC call, no buffering, no cache mutation)
+        and ``False`` is returned so the caller re-emits on the next cycle.
+
         Returns:
-            True if the send was successful, False if all retries were exhausted.
+            True if the send was successful, False if the socket was missing or
+            all retries were exhausted.
             Cache updates should only be performed by the caller when this returns True.
         """
+        if not self._is_platform_connector_socket_present():
+            metrics.events_sent_skipped_pc_unavailable.inc()
+            log.warning(
+                "Platform-connector socket %s is missing; skipping send.",
+                self._socket_path,
+            )
+            return False
+
         delay = INITIAL_DELAY
-        for _ in range(MAX_RETRIES):
+        for attempt in range(MAX_RETRIES):
+            # Re-check between retries so a connector that disappears
+            # mid-flight short-circuits instead of burning the budget.
+            if attempt > 0 and not self._is_platform_connector_socket_present():
+                metrics.events_sent_skipped_pc_unavailable.inc()
+                log.warning(
+                    "Platform-connector socket %s disappeared mid-retry; aborting send.",
+                    self._socket_path,
+                )
+                return False
+
             with grpc.insecure_channel(f"unix://{self._socket_path}") as chan:
                 stub = platformconnector_pb2_grpc.PlatformConnectorStub(chan)
                 try:

health-monitors/system-services-monitor/system_services_monitor/platform_connector/metrics.py

@@ -31,3 +31,8 @@
     "fabric_monitor_events_sent_error",
     "Total number of failed health event sends to platform-connector UDS",
 )
+
+events_sent_skipped_pc_unavailable = Counter(
+    "fabric_monitor_events_sent_skipped_pc_unavailable",
+    "Total number of health event sends skipped because the platform-connector Unix socket was missing",
+)