feat: read prefill/decode worker counts from DGD status (#5934)

components/src/dynamo/planner/kube.py

@@ -153,6 +153,44 @@ class KubernetesAPI:
 
         return ready_condition is not None and ready_condition.get("status") == "True"
 
+    def get_service_replica_status(
+        self, deployment: dict, service_name: str
+    ) -> tuple[int, bool]:
+        """
+        Get the actual ready replica count for a service from DGD status.
+
+        Returns:
+            tuple[int, bool]: (replica_count, is_stable)
+            - replica_count: number of replicas serving traffic (availableReplicas if present, else readyReplicas)
+            - is_stable: no rollout is in progress (desired == updated == ready/available)
+        """
+        # Get desired replicas from spec
+        service_spec = (
+            deployment.get("spec", {}).get("services", {}).get(service_name, {})
+        )
+        desired_replicas = service_spec.get("replicas", 0)
+
+        # Get status fields
+        service_status = (
+            deployment.get("status", {}).get("services", {}).get(service_name, {})
+        )
+        available = service_status.get("availableReplicas")
+        ready = service_status.get("readyReplicas", 0)
+        updated = service_status.get("updatedReplicas", 0)
+
+        # availableReplicas takes precedence over readyReplicas for the count
+        # refer to ServiceReplicaStatus type (https://github.com/ai-dynamo/dynamo/blob/main/deploy/operator/api/v1alpha1/dynamographdeployment_types.go#L157)
+        if available is not None:
+            traffic_serving_replicas = available
+        else:
+            traffic_serving_replicas = ready
+
+        # Stable means: desired == updated == ready/available
+        # This ensures we're not in a scale-up, scale-down, or rollout
+        is_stable = desired_replicas == updated == traffic_serving_replicas
+
+        return traffic_serving_replicas, is_stable
+
     async def wait_for_graph_deployment_ready(
         self,
         graph_deployment_name: str,

components/src/dynamo/planner/kubernetes_connector.py

@@ -284,6 +284,52 @@ class KubernetesConnector(PlannerConnector):
             self.graph_deployment_name,
         )
 
+    def get_actual_worker_counts(
+        self,
+        prefill_component_name: Optional[str] = None,
+        decode_component_name: Optional[str] = None,
+    ) -> tuple[int, int, bool]:
+        """
+        Get actual ready worker counts for prefill and decode from DGD status.
+
+        Returns:
+            tuple[int, int, bool]: (prefill_count, decode_count, is_stable)
+            - is_stable: False if any service is in a rollout (scaling should be skipped)
+        """
+        deployment = self.kube_api.get_graph_deployment(self.graph_deployment_name)
+
+        prefill_count = 0
+        decode_count = 0
+        all_stable = True
+
+        if prefill_component_name:
+            service = get_service_from_sub_component_type_or_name(
+                deployment,
+                SubComponentType.PREFILL,
+                component_name=prefill_component_name,
+            )
+            ready_replicas, is_stable = self.kube_api.get_service_replica_status(
+                deployment, service.name
+            )
+            if not is_stable:
+                all_stable = False
+            prefill_count = ready_replicas
+
+        if decode_component_name:
+            service = get_service_from_sub_component_type_or_name(
+                deployment,
+                SubComponentType.DECODE,
+                component_name=decode_component_name,
+            )
+            ready_replicas, is_stable = self.kube_api.get_service_replica_status(
+                deployment, service.name
+            )
+            if not is_stable:
+                all_stable = False
+            decode_count = ready_replicas
+
+        return prefill_count, decode_count, all_stable
+
     async def set_component_replicas(
         self, target_replicas: list[TargetReplica], blocking: bool = True
     ):

components/src/dynamo/planner/utils/planner_core.py

@@ -123,8 +123,8 @@ class PlannerPrometheusMetrics:
 @dataclass
 class PlannerSharedState:
     last_metrics: Metrics = field(default_factory=Metrics)
-    p_endpoints: list = field(default_factory=list)
-    d_endpoints: list = field(default_factory=list)
+    num_p_workers: int = 0
+    num_d_workers: int = 0
     cumulative_gpu_hours: float = 0.0
     last_adjustment_time: float = 0.0
 
@@ -444,12 +444,41 @@ class BasePlanner:
 
     async def get_workers_info(
         self, require_prefill: bool = True, require_decode: bool = True
-    ):
+    ) -> tuple[int, int, bool]:
+        """
+        Get worker counts for prefill and decode components.
+
+        Returns:
+            tuple[int, int, bool]: (num_p_workers, num_d_workers, is_stable)
+            - is_stable: False if rollout in progress (scaling should be skipped)
+        """
+        num_p_workers = 0
+        num_d_workers = 0
+
+        # For Kubernetes, use DGD status instead of runtime client
+        if hasattr(self, "connector") and isinstance(
+            self.connector, KubernetesConnector
+        ):
+            (
+                prefill_count,
+                decode_count,
+                is_stable,
+            ) = self.connector.get_actual_worker_counts(
+                prefill_component_name=(
+                    self.prefill_component_name if require_prefill else None
+                ),
+                decode_component_name=(
+                    self.decode_component_name if require_decode else None
+                ),
+            )
+            num_p_workers = prefill_count if require_prefill else 0
+            num_d_workers = decode_count if require_decode else 0
+            return num_p_workers, num_d_workers, is_stable
+
+        # Fall back to runtime client for non-Kubernetes environments
         if self.runtime is None:
             raise RuntimeError("Runtime is not initialized")
 
-        p_endpoints = []
-        d_endpoints = []
         worker_names = WORKER_COMPONENT_NAMES[self.args.backend]
 
         if require_prefill:
@@ -459,9 +488,9 @@ class BasePlanner:
                         worker_names.prefill_worker_component_name,
                         worker_names.prefill_worker_endpoint,
                     )
-                p_endpoints = self.prefill_client.instance_ids()  # type: ignore
+                num_p_workers = len(self.prefill_client.instance_ids())  # type: ignore
             except Exception:
-                p_endpoints = []
+                num_p_workers = 0
                 logger.warning(
                     "No prefill workers found, aggregated mode is not supported yet"
                 )
@@ -473,35 +502,39 @@ class BasePlanner:
                         worker_names.decode_worker_component_name,
                         worker_names.decode_worker_endpoint,
                     )
-                d_endpoints = self.workers_client.instance_ids()  # type: ignore
+                num_d_workers = len(self.workers_client.instance_ids())  # type: ignore
             except Exception as e:
                 raise RuntimeError(f"Failed to get decode worker endpoints: {e}")
 
-        return p_endpoints, d_endpoints
+        return num_p_workers, num_d_workers, True  # Always stable for non-K8s
 
     async def observe_metrics(
         self, require_prefill: bool = True, require_decode: bool = True
-    ):
-        p_endpoints, d_endpoints = await self.get_workers_info(
+    ) -> None:
+        """
+        Observe metrics from Prometheus and update shared state.
+        """
+        num_p_workers, num_d_workers, _ = await self.get_workers_info(
             require_prefill=require_prefill, require_decode=require_decode
         )
-        self.shared_state.p_endpoints = p_endpoints
-        self.shared_state.d_endpoints = d_endpoints
+
+        self.shared_state.num_p_workers = num_p_workers
+        self.shared_state.num_d_workers = num_d_workers
         logger.debug(
-            f"Number of prefill workers: {len(p_endpoints)}, number of decode workers: {len(d_endpoints)}"
+            f"Number of prefill workers: {num_p_workers}, number of decode workers: {num_d_workers}"
         )
 
         # Update Prometheus metrics if server is running
         if self.prometheus_port != 0 and self.prometheus_metrics is not None:
-            self.prometheus_metrics.num_p_workers.set(len(p_endpoints))
-            self.prometheus_metrics.num_d_workers.set(len(d_endpoints))
+            self.prometheus_metrics.num_p_workers.set(num_p_workers)
+            self.prometheus_metrics.num_d_workers.set(num_d_workers)
 
             # Calculate and accumulate GPU hours for this interval
             # TODO: track startup and shutdown times to get more accurate GPU hours
             interval_gpu_hours = (
                 (
-                    len(p_endpoints) * self.args.prefill_engine_num_gpu
-                    + len(d_endpoints) * self.args.decode_engine_num_gpu
+                    num_p_workers * self.args.prefill_engine_num_gpu
+                    + num_d_workers * self.args.decode_engine_num_gpu
                 )
                 * self.args.adjustment_interval
                 / 3600
@@ -769,14 +802,14 @@ class DecodePlanner(BasePlanner):
     component_type = SubComponentType.DECODE
 
     def _update_correction_factor(self) -> bool:
-        if not self.shared_state.d_endpoints:
+        if self.shared_state.num_d_workers == 0:
             logger.warning(
                 "No decode workers found for correction factor, skipping correction update"
             )
             return True
         expect_itl = self.decode_interpolator.interpolate_itl(
             concurrency=self.last_metrics.num_req  # type: ignore
-            / len(self.shared_state.d_endpoints)
+            / self.shared_state.num_d_workers
             * self.last_metrics.request_duration  # type: ignore
             / self.args.adjustment_interval,
             context_length=self.last_metrics.isl + self.last_metrics.osl / 2,  # type: ignore

tests/planner/unit/kube.py

@@ -346,3 +346,162 @@ async def test_get_graph_deployment_not_found(k8s_api, mock_custom_api):
     exception = exc_info.value
     assert exception.deployment_name == "parent-dgd"
     assert exception.namespace == "default"
+
+
+# Tests for get_service_replica_status
+
+
+def test_get_service_replica_status_stable_with_available_replicas(
+    k8s_api, mock_custom_api
+):
+    """Test stable case with availableReplicas present (takes precedence over readyReplicas)"""
+    deployment: Dict[str, Any] = {
+        "spec": {"services": {"prefill-worker": {"replicas": 2}}},
+        "status": {
+            "services": {
+                "prefill-worker": {
+                    "availableReplicas": 2,
+                    "readyReplicas": 2,
+                    "updatedReplicas": 2,
+                }
+            }
+        },
+    }
+
+    count, is_stable = k8s_api.get_service_replica_status(deployment, "prefill-worker")
+
+    assert count == 2
+    assert is_stable is True
+
+
+def test_get_service_replica_status_stable_with_ready_replicas_fallback(
+    k8s_api, mock_custom_api
+):
+    """Test stable case falling back to readyReplicas when availableReplicas is not present"""
+    deployment: Dict[str, Any] = {
+        "spec": {"services": {"decode-worker": {"replicas": 4}}},
+        "status": {
+            "services": {
+                "decode-worker": {
+                    "readyReplicas": 4,
+                    "updatedReplicas": 4,
+                }
+            }
+        },
+    }
+
+    count, is_stable = k8s_api.get_service_replica_status(deployment, "decode-worker")
+
+    assert count == 4
+    assert is_stable is True
+
+
+def test_get_service_replica_status_scale_up_in_progress(k8s_api, mock_custom_api):
+    """Test scale-up in progress: desired=4, updated=2, ready=2"""
+    deployment: Dict[str, Any] = {
+        "spec": {"services": {"prefill-worker": {"replicas": 4}}},
+        "status": {
+            "services": {
+                "prefill-worker": {
+                    "availableReplicas": 2,
+                    "readyReplicas": 2,
+                    "updatedReplicas": 2,
+                }
+            }
+        },
+    }
+
+    count, is_stable = k8s_api.get_service_replica_status(deployment, "prefill-worker")
+
+    assert count == 2
+    assert is_stable is False
+
+
+def test_get_service_replica_status_scale_down_in_progress(k8s_api, mock_custom_api):
+    """Test scale-down in progress: desired=2, updated=4, ready=4"""
+    deployment: Dict[str, Any] = {
+        "spec": {"services": {"decode-worker": {"replicas": 2}}},
+        "status": {
+            "services": {
+                "decode-worker": {
+                    "availableReplicas": 4,
+                    "readyReplicas": 4,
+                    "updatedReplicas": 4,
+                }
+            }
+        },
+    }
+
+    count, is_stable = k8s_api.get_service_replica_status(deployment, "decode-worker")
+
+    assert count == 4
+    assert is_stable is False
+
+
+def test_get_service_replica_status_rollout_in_progress(k8s_api, mock_custom_api):
+    """Test rollout in progress: desired=4, updated=2, ready=4 (old replicas still running)"""
+    deployment: Dict[str, Any] = {
+        "spec": {"services": {"prefill-worker": {"replicas": 4}}},
+        "status": {
+            "services": {
+                "prefill-worker": {
+                    "availableReplicas": 4,
+                    "readyReplicas": 4,
+                    "updatedReplicas": 2,
+                }
+            }
+        },
+    }
+
+    count, is_stable = k8s_api.get_service_replica_status(deployment, "prefill-worker")
+
+    assert count == 4
+    assert is_stable is False
+
+
+def test_get_service_replica_status_missing_status_fields(k8s_api, mock_custom_api):
+    """Test handling when status fields are missing"""
+    deployment: Dict[str, Any] = {
+        "spec": {"services": {"prefill-worker": {"replicas": 2}}},
+        "status": {"services": {}},
+    }
+
+    count, is_stable = k8s_api.get_service_replica_status(deployment, "prefill-worker")
+
+    # Should default to 0 for missing fields
+    assert count == 0
+    # desired=2, updated=0, count=0 -> not stable
+    assert is_stable is False
+
+
+def test_get_service_replica_status_empty_deployment(k8s_api, mock_custom_api):
+    """Test handling when deployment has no spec or status"""
+    deployment: Dict[str, Any] = {}
+
+    count, is_stable = k8s_api.get_service_replica_status(deployment, "prefill-worker")
+
+    # All values default to 0, which makes it "stable" (0 == 0 == 0)
+    assert count == 0
+    assert is_stable is True
+
+
+def test_get_service_replica_status_available_replicas_zero(k8s_api, mock_custom_api):
+    """Test when availableReplicas is explicitly 0 (should use 0, not fall back to ready)"""
+    deployment: Dict[str, Any] = {
+        "spec": {"services": {"prefill-worker": {"replicas": 0}}},
+        "status": {
+            "services": {
+                "prefill-worker": {
+                    "availableReplicas": 0,
+                    "readyReplicas": 2,  # Should be ignored
+                    "updatedReplicas": 0,
+                }
+            }
+        },
+    }
+
+    count, is_stable = k8s_api.get_service_replica_status(deployment, "prefill-worker")
+
+    # availableReplicas=0 should be used (not readyReplicas)
+    assert count == 0
+    assert is_stable is True

tests/planner/unit/kubernetes_connector.py

@@ -888,3 +888,147 @@ def test_get_gpu_counts_service_not_found_raises_error(
         kubernetes_connector.get_gpu_counts()
 
     assert "decode GPU count" in str(exc_info.value)
+
+
+# Tests for get_actual_worker_counts
+
+
+def test_get_actual_worker_counts_stable(kubernetes_connector, mock_kube_api):
+    """Test get_actual_worker_counts when both services are stable"""
+    mock_deployment = {
+        "metadata": {"name": "test-graph"},
+        "spec": {
+            "services": {
+                "prefill-component": {},
+                "decode-component": {},
+            }
+        },
+    }
+    mock_kube_api.get_graph_deployment.return_value = mock_deployment
+    mock_kube_api.get_service_replica_status.side_effect = [(2, True), (4, True)]
+
+    (
+        prefill_count,
+        decode_count,
+        is_stable,
+    ) = kubernetes_connector.get_actual_worker_counts(
+        prefill_component_name="prefill-component",
+        decode_component_name="decode-component",
+    )
+
+    assert prefill_count == 2
+    assert decode_count == 4
+    assert is_stable is True
+
+
+def test_get_actual_worker_counts_prefill_rollout_in_progress(
+    kubernetes_connector, mock_kube_api
+):
+    """Test get_actual_worker_counts when prefill has rollout in progress"""
+    mock_deployment = {
+        "metadata": {"name": "test-graph"},
+        "spec": {
+            "services": {
+                "prefill-component": {},
+                "decode-component": {},
+            }
+        },
+    }
+    mock_kube_api.get_graph_deployment.return_value = mock_deployment
+    mock_kube_api.get_service_replica_status.side_effect = [(2, False), (4, True)]
+
+    (
+        prefill_count,
+        decode_count,
+        is_stable,
+    ) = kubernetes_connector.get_actual_worker_counts(
+        prefill_component_name="prefill-component",
+        decode_component_name="decode-component",
+    )
+
+    assert prefill_count == 2
+    assert decode_count == 4
+    assert is_stable is False
+
+
+def test_get_actual_worker_counts_prefill_only(kubernetes_connector, mock_kube_api):
+    """Test get_actual_worker_counts with only prefill component"""
+    mock_deployment = {
+        "metadata": {"name": "test-graph"},
+        "spec": {
+            "services": {
+                "prefill-component": {
+                    "replicas": 2,
+                    "subComponentType": "prefill",
+                },
+            }
+        },
+    }
+    mock_kube_api.get_graph_deployment.return_value = mock_deployment
+    mock_kube_api.get_service_replica_status.return_value = (2, True)
+
+    (
+        prefill_count,
+        decode_count,
+        is_stable,
+    ) = kubernetes_connector.get_actual_worker_counts(
+        prefill_component_name="prefill-component",
+        decode_component_name=None,
+    )
+
+    assert prefill_count == 2
+    assert decode_count == 0
+    assert is_stable is True
+
+
+def test_get_actual_worker_counts_decode_only(kubernetes_connector, mock_kube_api):
+    """Test get_actual_worker_counts with only decode component"""
+    mock_deployment = {
+        "metadata": {"name": "test-graph"},
+        "spec": {
+            "services": {
+                "decode-component": {
+                    "replicas": 4,
+                    "subComponentType": "decode",
+                },
+            }
+        },
+    }
+    mock_kube_api.get_graph_deployment.return_value = mock_deployment
+    mock_kube_api.get_service_replica_status.return_value = (4, True)
+
+    (
+        prefill_count,
+        decode_count,
+        is_stable,
+    ) = kubernetes_connector.get_actual_worker_counts(
+        prefill_component_name=None,
+        decode_component_name="decode-component",
+    )
+
+    assert prefill_count == 0
+    assert decode_count == 4
+    assert is_stable is True
+
+
+def test_get_actual_worker_counts_no_components(kubernetes_connector, mock_kube_api):
+    """Test get_actual_worker_counts with no components specified"""
+    mock_deployment = {
+        "metadata": {"name": "test-graph"},
+        "spec": {"services": {}},
+        "status": {"services": {}},
+    }
+    mock_kube_api.get_graph_deployment.return_value = mock_deployment
+
+    (
+        prefill_count,
+        decode_count,
+        is_stable,
+    ) = kubernetes_connector.get_actual_worker_counts(
+        prefill_component_name=None,
+        decode_component_name=None,
+    )
+
+    assert prefill_count == 0
+    assert decode_count == 0
+    assert is_stable is True

tests/planner/unit/test_sla_planner_scaling.py

@@ -88,8 +88,9 @@ def _build_planners(args, prometheus_client):
 
     async def mock_get_workers_info(require_prefill=True, require_decode=True):
         return (
-            ["prefill-0"] if require_prefill else [],
-            ["decode-0"] if require_decode else [],
+            1 if require_prefill else 0,
+            1 if require_decode else 0,
+            True,  # is_stable
         )
 
     prefill_planner.get_workers_info = mock_get_workers_info