feat(observability): drop the aiperf/tachometer scrape exclusion

src/srtctl/cli/mixins/benchmark_stage.py

@@ -664,38 +664,6 @@ def _get_aiperf_server_metrics_env(
         urls = list(dict.fromkeys(urls)) if logical_workers_only else sorted(set(urls))
         return {"AIPERF_SERVER_METRICS_URLS": ",".join(urls)}
 
-    def _client_polled_metric_urls(self) -> frozenset[str]:
-        """The ``/metrics`` URLs the benchmark client will poll on its own.
-
-        Tachometer scrapes the complement of this set (see
-        ``TelemetryStageMixin.start_tachometer``), so it is derived from the
-        same logic that injects ``AIPERF_SERVER_METRICS_URLS`` — including the
-        dead-TRT-LLM-worker omission and the explicit recipe override. It is
-        deliberately NOT a second endpoint list to maintain: when the injected
-        set changes, the complement moves with it. A serve-only or manual run
-        has no client, so nothing is polled and Tachometer covers everything.
-        """
-        if bool(getattr(self, "serve_only", False)):
-            return frozenset()
-        explicit = self.runtime.environment.get("AIPERF_SERVER_METRICS_URLS")
-        if explicit is not None:
-            return frozenset(url for url in explicit.split(",") if url)
-        from srtctl.benchmarks.base import AIPerfBenchmarkRunner, get_runner
-
-        benchmark_type = self.config.benchmark.type
-        if benchmark_type == "custom":
-            env = self._get_aiperf_server_metrics_env(logical_workers_only=True)
-        else:
-            try:
-                runner = get_runner(benchmark_type)
-            except ValueError:
-                return frozenset()
-            if not isinstance(runner, AIPerfBenchmarkRunner):
-                return frozenset()
-            env = self._get_aiperf_server_metrics_env()
-        urls = env.get("AIPERF_SERVER_METRICS_URLS", "")
-        return frozenset(url for url in urls.split(",") if url)
-
     def _get_benchmark_env(self, runner: "BenchmarkRunner") -> dict[str, str]:
         """Get environment variables for the benchmark script."""
         from srtctl.benchmarks.base import AIPerfBenchmarkRunner

src/srtctl/cli/mixins/telemetry_stage.py

@@ -295,17 +295,6 @@ def start_tachometer(self) -> list[ManagedProcess]:
 
         logger.info("Starting Tachometer")
 
-        # Scrape the complement of what the benchmark client already polls;
-        # the helper lives on BenchmarkStageMixin (same orchestrator object).
-        client_urls_fn = getattr(self, "_client_polled_metric_urls", None)
-        exclude_urls = client_urls_fn() if client_urls_fn is not None else frozenset()
-        if exclude_urls:
-            logger.info(
-                "Tachometer excludes %d endpoint(s) the benchmark client polls: %s",
-                len(exclude_urls),
-                ",".join(sorted(exclude_urls)),
-            )
-
         power_telemetry = self.config.telemetry
         dcgm_exporter = power_telemetry.dcgm_exporter if power_telemetry.enabled else tachometer.dcgm_exporter
         topology = self._compute_frontend_topology()
@@ -318,7 +307,6 @@ def start_tachometer(self) -> list[ManagedProcess]:
                 tachometer=tachometer,
                 dcgm_exporter=dcgm_exporter,
                 frontend_type=self.config.frontend.type,
-                exclude_urls=exclude_urls,
             )
         )
 

src/srtctl/core/telemetry.py

@@ -47,16 +47,13 @@ def generate_tachometer_config(
     tachometer: TachometerConfig,
     dcgm_exporter: TelemetryExporterConfig | None = None,
     frontend_type: str = "dynamo",
-    exclude_urls: frozenset[str] | set[str] = frozenset(),
 ) -> str:
     """Generate Tachometer TOML from backend and frontend topology.
 
-    ``exclude_urls`` is the set of ``/metrics`` URLs the benchmark client
-    already polls (``AIPERF_SERVER_METRICS_URLS``). Tachometer scrapes the
-    complement so a worker endpoint is never double-polled — the extra scrape
-    load has previously made a submission irreproducible. Frontend, DCGM and
-    node-exporter endpoints are never excluded: the frontend scrape is cheap
-    and Tachometer is the only whole-window, per-replica capture of it.
+    Every endpoint is scraped even when the benchmark client polls the same
+    URL (``AIPERF_SERVER_METRICS_URLS``): double-polling has been validated
+    as harmless, and unconditional coverage keeps Tachometer the one
+    whole-window, per-replica capture regardless of what the client does.
     """
     dcgm_exporter = dcgm_exporter or tachometer.dcgm_exporter
     node_exporter = tachometer.node_exporter
@@ -117,10 +114,6 @@ def generate_tachometer_config(
         else:
             port = process.sys_port
         url = f"http://{node_ip}:{port}/metrics"
-        if url in exclude_urls:
-            # The benchmark client already polls this endpoint on its own
-            # cadence; scrape the complement instead of double-polling.
-            continue
         node_metadata = {
             "hostname": process.node,
             "worker_index": str(process.endpoint_index),

tests/test_telemetry.py

@@ -382,8 +382,9 @@ def test_storage_leaf_is_never_pre_created(self, _mock_get_hostname_ip, tmp_path
         assert not local_dir.is_relative_to(storage_path)
 
     @patch("srtctl.core.telemetry.get_hostname_ip", return_value="10.0.0.1")
-    def test_client_polled_urls_are_excluded_from_backend_targets(self, _mock_get_hostname_ip):
-        """Tachometer scrapes the complement of the client's URL list."""
+    def test_client_polled_urls_are_still_scraped(self, _mock_get_hostname_ip):
+        """Tachometer scrapes every endpoint, including URLs the benchmark
+        client also polls — double-polling has been validated as harmless."""
         tachometer = TachometerConfig(enabled=True)
         runtime = MagicMock(job_id="12345", run_name="test_12345", network_interface="eth0")
         runtime.log_dir = Path("/runs/12345/logs")
@@ -414,12 +415,12 @@ def test_client_polled_urls_are_excluded_from_backend_targets(self, _mock_get_ho
             frontend_topology=topology,
             runtime=runtime,
             tachometer=tachometer,
-            exclude_urls={"http://10.0.0.1:8081/metrics"},
         )
 
-        assert 'url = "http://10.0.0.1:8081/metrics"' not in config_text
+        # Both worker endpoints appear even though a benchmark client would
+        # poll the same sys-port URLs via AIPERF_SERVER_METRICS_URLS.
+        assert 'url = "http://10.0.0.1:8081/metrics"' in config_text
         assert 'url = "http://10.0.0.1:8082/metrics"' in config_text
-        # The frontend endpoint is never excluded (whole-window coverage).
         assert 'url = "http://10.0.0.1:8000/metrics"' in config_text
 
     @patch("srtctl.core.telemetry.get_hostname_ip", return_value="10.0.0.1")