chore: add limits and requests configuration in values (#867)

backend/backend/settings/common.py

@@ -197,6 +197,7 @@
     "KANIKO_MIRROR": to_bool(os.environ.get("KANIKO_MIRROR", False)),
     "KANIKO_IMAGE": os.environ.get("KANIKO_IMAGE"),
     "KANIKO_DOCKER_CONFIG_SECRET_NAME": os.environ.get("KANIKO_DOCKER_CONFIG_SECRET_NAME"),
+    "KANIKO_RESOURCES": os.environ.get("KANIKO_RESOURCES"),
     "COMPUTE_POD_STARTUP_TIMEOUT_SECONDS": int(os.environ.get("COMPUTE_POD_STARTUP_TIMEOUT_SECONDS", 300)),
     "PRIVATE_CA_ENABLED": to_bool(os.environ.get("PRIVATE_CA_ENABLED")),
 }
@@ -223,6 +224,7 @@
 COMPUTE_POD_RUN_AS_GROUP = os.environ.get("COMPUTE_POD_RUN_AS_GROUP")
 COMPUTE_POD_FS_GROUP = os.environ.get("COMPUTE_POD_FS_GROUP")
 COMPUTE_POD_GKE_GPUS_LIMITS = int(os.environ.get("COMPUTE_POD_GKE_GPUS_LIMITS", 0))
+COMPUTE_POD_RESOURCES = os.environ.get("COMPUTE_POD_RESOURCES")
 
 # Prometheus configuration
 ENABLE_METRICS = to_bool(os.environ.get("ENABLE_METRICS", False))

backend/builder/image_builder/image_builder.py

@@ -23,6 +23,7 @@
 from substrapp.compute_tasks.volumes import get_worker_subtuple_pvc_name
 from substrapp.docker_registry import USER_IMAGE_REPOSITORY
 from substrapp.kubernetes_utils import delete_pod
+from substrapp.kubernetes_utils import get_resources_requirements_from_yaml
 from substrapp.kubernetes_utils import get_security_context
 from substrapp.lock_local import lock_resource
 from substrapp.utils import timeit
@@ -42,6 +43,7 @@
 IMAGE_BUILD_TIMEOUT = settings.IMAGE_BUILD_TIMEOUT
 KANIKO_CONTAINER_NAME = "kaniko"
 HOSTNAME = settings.HOSTNAME
+KANIKO_RESOURCES = settings.TASK["KANIKO_RESOURCES"]
 
 
 def container_image_tag_from_function(function: orchestrator.Function) -> str:
@@ -306,6 +308,7 @@ def _build_container(dockerfile_mount_path: str, image_tag: str) -> kubernetes.c
         args=args,
         volume_mounts=volume_mounts,
         security_context=container_security_context,
+        resources=get_resources_requirements_from_yaml(yaml_resources=KANIKO_RESOURCES),
     )
 
 

backend/substrapp/compute_tasks/compute_pod.py

@@ -6,9 +6,11 @@
 
 from substrapp.kubernetes_utils import delete_pod
 from substrapp.kubernetes_utils import get_pod_security_context
+from substrapp.kubernetes_utils import get_resources_requirements_from_yaml
 from substrapp.kubernetes_utils import get_security_context
 
 NAMESPACE = settings.NAMESPACE
+COMPUTE_POD_RESOURCES = settings.COMPUTE_POD_RESOURCES
 logger = structlog.get_logger(__name__)
 
 
@@ -112,6 +114,7 @@ def create_pod(
         args=None,
         volume_mounts=volume_mounts + gpu_volume_mounts,
         security_context=get_security_context(),
+        resources=get_resources_requirements_from_yaml(yaml_resources=COMPUTE_POD_RESOURCES),
         env=[kubernetes.client.V1EnvVar(name=env_name, value=env_value) for env_name, env_value in environment.items()],
         **container_optional_kwargs,
     )

backend/substrapp/kubernetes_utils.py

@@ -1,5 +1,6 @@
 import kubernetes
 import structlog
+import yaml
 from django.conf import settings
 
 from substrapp.exceptions import KubernetesError
@@ -47,6 +48,18 @@ def get_security_context(root: bool = False, capabilities: list[str] = None) ->
     return security_context
 
 
+def get_resources_requirements_from_yaml(
+    *,
+    yaml_resources: str,
+) -> kubernetes.client.V1ResourceRequirements:
+    """Return a kubernetes.client.V1ResourceRequirements object from a yaml string."""
+    resources_dict = yaml.safe_load(yaml_resources)
+
+    return kubernetes.client.V1ResourceRequirements(
+        requests=resources_dict["requests"], limits=resources_dict["limits"]
+    )
+
+
 def pod_exists_by_label_selector(k8s_client: kubernetes.client.CoreV1Api, label_selector: str) -> bool:
     """Return True if the pod exists, else False.
 

backend/substrapp/tasks/task.py

@@ -73,7 +73,7 @@ def on_success(self, retval: dict[str, Any], task_id: str, args: tuple, kwargs:
     # Celery does not provide unpacked arguments, we are doing it in `split_args`
     def on_retry(self, exc: Exception, task_id: str, args: tuple, kwargs: dict[str, Any], einfo: ExceptionInfo) -> None:
         _, task = self.split_args(args)
-        # delete compute pod to reset hardware ressources
+        # delete compute pod to reset hardware resources
         delete_compute_plan_pods(task.compute_plan_key)
         logger.info(
             "Retrying task",

changes/892.added

@@ -0,0 +1 @@
+Resources to Kaniko pods and computeTask pods

charts/substra-backend/CHANGELOG.md

@@ -2,6 +2,12 @@
 
 <!-- towncrier release notes start -->
 
+## [26.4.0] - 2024-04-25
+
+### Added
+
+- Resources limits and requests (CPU and memory) for all containers are set and configurable in values.
+
 ## [26.3.0] - 2024-04-19
 
 ### Removed

charts/substra-backend/Chart.yaml

@@ -1,7 +1,7 @@
 apiVersion: v2
 name: substra-backend
 home: https://github.com/Substra
-version: 26.3.0
+version: 26.4.0
 appVersion: 0.45.0
 kubeVersion: ">= 1.19.0-0"
 description: Main package for Substra