🐛🎨External Clusters: improve work stealing (#5474)

Author: sanderegg

.env-devel

@@ -36,6 +36,7 @@ CATALOG_SERVICES_DEFAULT_SPECIFICATIONS='{}'
 
 CLUSTERS_KEEPER_COMPUTATIONAL_BACKEND_DOCKER_IMAGE_TAG=master-github-latest
 CLUSTERS_KEEPER_DASK_NTHREADS=0
+CLUSTERS_KEEPER_DASK_WORKER_SATURATION=inf
 CLUSTERS_KEEPER_EC2_ACCESS=null
 CLUSTERS_KEEPER_MAX_MISSED_HEARTBEATS_BEFORE_CLUSTER_TERMINATION=5
 CLUSTERS_KEEPER_PRIMARY_EC2_INSTANCES=null

Makefile

@@ -523,14 +523,13 @@ nodenv: node_modules ## builds node_modules local environ (TODO)
 
 pylint: ## python linting
 	# pylint version info
-	@/bin/bash -c "pylint --version"
+	@pylint --version
 	# Running linter in packages and services (except director)
 	@folders=$$(find $(CURDIR)/services $(CURDIR)/packages  -type d -not -path "*/director/*" -name 'src' -exec dirname {} \; | sort -u); \
 	exit_status=0; \
 	for folder in $$folders; do \
-		pushd "$$folder"; \
-		make pylint || exit_status=1; \
-		popd; \
+		echo "Linting $$folder"; \
+		$(MAKE_C) "$$folder" pylint || exit_status=1; \
 	done;\
 	exit $$exit_status
 	# Running linter elsewhere

packages/settings-library/src/settings_library/base.py

@@ -13,8 +13,9 @@
 )
 from pydantic.error_wrappers import ErrorList, ErrorWrapper
 from pydantic.fields import ModelField, Undefined
+from pydantic.typing import is_literal_type
 
-logger = logging.getLogger(__name__)
+_logger = logging.getLogger(__name__)
 
 _DEFAULTS_TO_NONE_MSG: Final[
     str
@@ -39,7 +40,7 @@ def _default_factory():
         except ValidationError as err:
             if field.allow_none:
                 # e.g. Optional[PostgresSettings] would warn if defaults to None
-                logger.warning(
+                _logger.warning(
                     _DEFAULTS_TO_NONE_MSG,
                     field.name,
                 )
@@ -101,8 +102,14 @@ def prepare_field(cls, field: ModelField) -> None:
             is_not_composed = (
                 get_origin(field_type) is None
             )  # is not composed as dict[str, Any] or Generic[Base]
-
-            if is_not_composed and issubclass(field_type, BaseCustomSettings):
+            # avoid literals raising TypeError
+            is_not_literal = is_literal_type(field.type_) is False
+
+            if (
+                is_not_literal
+                and is_not_composed
+                and issubclass(field_type, BaseCustomSettings)
+            ):
                 if auto_default_from_env:
                     assert field.field_info.default is Undefined
                     assert field.field_info.default_factory is None
@@ -112,7 +119,11 @@ def prepare_field(cls, field: ModelField) -> None:
                     field.default = None
                     field.required = False  # has a default now
 
-            elif is_not_composed and issubclass(field_type, BaseSettings):
+            elif (
+                is_not_literal
+                and is_not_composed
+                and issubclass(field_type, BaseSettings)
+            ):
                 msg = f"{cls}.{field.name} of type {field_type} must inherit from BaseCustomSettings"
                 raise ConfigError(msg)
 

packages/settings-library/tests/test_base.py

@@ -2,6 +2,7 @@
 # pylint: disable=unused-argument
 # pylint: disable=unused-variable
 # pylint: disable=too-many-arguments
+# pylint: disable=protected-access
 
 import inspect
 import json
@@ -196,7 +197,7 @@ def test_auto_default_to_none_logs_a_warning(
     create_settings_class: Callable[[str], type[BaseCustomSettings]],
     mocker: MockerFixture,
 ):
-    logger_warn = mocker.spy(settings_library.base.logger, "warning")
+    logger_warn = mocker.spy(settings_library.base._logger, "warning")  # noqa: SLF001
 
     S = create_settings_class("S")
 

services/autoscaling/src/simcore_service_autoscaling/modules/auto_scaling_mode_computational.py

@@ -115,16 +115,30 @@ async def compute_node_used_resources(
         app: FastAPI, instance: AssociatedInstance
     ) -> Resources:
         try:
-            num_results_in_memory = await dask.get_worker_still_has_results_in_memory(
+            resource = await dask.get_worker_used_resources(
                 _scheduler_url(app), _scheduler_auth(app), instance.ec2_instance
             )
-            if num_results_in_memory > 0:
-                # NOTE: this is a trick to consider the node still useful
-                return Resources(cpus=0, ram=ByteSize(1024 * 1024 * 1024))
-            return await dask.get_worker_used_resources(
-                _scheduler_url(app), _scheduler_auth(app), instance.ec2_instance
+            if resource == Resources.create_as_empty():
+                num_results_in_memory = (
+                    await dask.get_worker_still_has_results_in_memory(
+                        _scheduler_url(app), _scheduler_auth(app), instance.ec2_instance
+                    )
+                )
+                if num_results_in_memory > 0:
+                    _logger.debug(
+                        "found %s for %s",
+                        f"{num_results_in_memory=}",
+                        f"{instance.ec2_instance.id}",
+                    )
+                    # NOTE: this is a trick to consider the node still useful
+                    return Resources(cpus=0, ram=ByteSize(1024 * 1024 * 1024))
+
+            _logger.debug(
+                "found %s for %s", f"{resource=}", f"{instance.ec2_instance.id}"
             )
+            return resource
         except (DaskWorkerNotFoundError, DaskNoWorkersError):
+            _logger.debug("no resource found for %s", f"{instance.ec2_instance.id}")
             return Resources.create_as_empty()
 
     @staticmethod

services/autoscaling/src/simcore_service_autoscaling/modules/dask.py

@@ -1,3 +1,4 @@
+import collections
 import contextlib
 import logging
 import re
@@ -217,30 +218,39 @@ async def get_worker_used_resources(
         DaskNoWorkersError
     """
 
-    def _get_worker_used_resources(
+    def _list_processing_tasks_on_worker(
         dask_scheduler: distributed.Scheduler, *, worker_url: str
-    ) -> dict[str, float] | None:
-        for worker_name, worker_state in dask_scheduler.workers.items():
-            if worker_url != worker_name:
-                continue
-            if worker_state.status is distributed.Status.closing_gracefully:
-                # NOTE: when a worker was retired it is in this state
-                return {}
-            return dict(worker_state.used_resources)
-        return None
+    ) -> list[tuple[DaskTaskId, DaskTaskResources]]:
+        processing_tasks = []
+        for task_key, task_state in dask_scheduler.tasks.items():
+            if task_state.processing_on and (
+                task_state.processing_on.address == worker_url
+            ):
+                processing_tasks.append((task_key, task_state.resource_restrictions))
+        return processing_tasks
 
     async with _scheduler_client(scheduler_url, authentication) as client:
         worker_url, _ = _dask_worker_from_ec2_instance(client, ec2_instance)
 
+        _logger.debug("looking for processing tasksfor %s", f"{worker_url=}")
+
         # now get the used resources
-        worker_used_resources: dict[str, Any] | None = await _wrap_client_async_routine(
-            client.run_on_scheduler(_get_worker_used_resources, worker_url=worker_url),
+        worker_processing_tasks: list[
+            tuple[DaskTaskId, DaskTaskResources]
+        ] = await _wrap_client_async_routine(
+            client.run_on_scheduler(
+                _list_processing_tasks_on_worker, worker_url=worker_url
+            ),
         )
-        if worker_used_resources is None:
-            raise DaskWorkerNotFoundError(worker_host=worker_url, url=scheduler_url)
+
+        total_resources_used: collections.Counter[str] = collections.Counter()
+        for _, task_resources in worker_processing_tasks:
+            total_resources_used.update(task_resources)
+
+        _logger.debug("found %s for %s", f"{total_resources_used=}", f"{worker_url=}")
         return Resources(
-            cpus=worker_used_resources.get("CPU", 0),
-            ram=parse_obj_as(ByteSize, worker_used_resources.get("RAM", 0)),
+            cpus=total_resources_used.get("CPU", 0),
+            ram=parse_obj_as(ByteSize, total_resources_used.get("RAM", 0)),
         )
 
 

services/autoscaling/src/simcore_service_autoscaling/utils/utils_docker.py

@@ -59,23 +59,41 @@
 _PENDING_DOCKER_TASK_MESSAGE: Final[str] = "pending task scheduling"
 _INSUFFICIENT_RESOURCES_DOCKER_TASK_ERR: Final[str] = "insufficient resources on"
 _NOT_SATISFIED_SCHEDULING_CONSTRAINTS_TASK_ERR: Final[str] = "no suitable node"
+_OSPARC_SERVICE_READY_LABEL_KEY: Final[DockerLabelKey] = parse_obj_as(
+    DockerLabelKey, "io.simcore.osparc-services-ready"
+)
+_OSPARC_SERVICES_READY_DATETIME_LABEL_KEY: Final[DockerLabelKey] = parse_obj_as(
+    DockerLabelKey, f"{_OSPARC_SERVICE_READY_LABEL_KEY}-last-changed"
+)
+_OSPARC_SERVICE_READY_LABEL_KEYS: Final[list[DockerLabelKey]] = [
+    _OSPARC_SERVICE_READY_LABEL_KEY,
+    _OSPARC_SERVICES_READY_DATETIME_LABEL_KEY,
+]
 
 
 async def get_monitored_nodes(
     docker_client: AutoscalingDocker, node_labels: list[DockerLabelKey]
 ) -> list[Node]:
+    node_label_filters = [f"{label}=true" for label in node_labels] + [
+        f"{label}" for label in _OSPARC_SERVICE_READY_LABEL_KEYS
+    ]
     return parse_obj_as(
         list[Node],
-        await docker_client.nodes.list(
-            filters={"node.label": [f"{label}=true" for label in node_labels]}
-        ),
+        await docker_client.nodes.list(filters={"node.label": node_label_filters}),
     )
 
 
 async def get_worker_nodes(docker_client: AutoscalingDocker) -> list[Node]:
     return parse_obj_as(
         list[Node],
-        await docker_client.nodes.list(filters={"role": ["worker"]}),
+        await docker_client.nodes.list(
+            filters={
+                "role": ["worker"],
+                "node.label": [
+                    f"{label}" for label in _OSPARC_SERVICE_READY_LABEL_KEYS
+                ],
+            }
+        ),
     )
 
 
@@ -550,14 +568,6 @@ def is_node_ready_and_available(node: Node, *, availability: Availability) -> bo
     )
 
 
-_OSPARC_SERVICE_READY_LABEL_KEY: Final[DockerLabelKey] = parse_obj_as(
-    DockerLabelKey, "io.simcore.osparc-services-ready"
-)
-_OSPARC_SERVICES_READY_DATETIME_LABEL_KEY: Final[DockerLabelKey] = parse_obj_as(
-    DockerLabelKey, f"{_OSPARC_SERVICE_READY_LABEL_KEY}-last-changed"
-)
-
-
 def is_node_osparc_ready(node: Node) -> bool:
     if not is_node_ready_and_available(node, availability=Availability.active):
         return False

services/autoscaling/tests/unit/test_utils_docker.py

@@ -37,6 +37,7 @@
 from simcore_service_autoscaling.modules.docker import AutoscalingDocker
 from simcore_service_autoscaling.utils.utils_docker import (
     _OSPARC_SERVICE_READY_LABEL_KEY,
+    _OSPARC_SERVICES_READY_DATETIME_LABEL_KEY,
     Node,
     _by_created_dt,
     attach_node,
@@ -133,7 +134,13 @@ async def test_get_monitored_nodes_with_valid_label(
     create_node_labels: Callable[[list[str]], Awaitable[None]],
 ):
     labels = faker.pylist(allowed_types=(str,))
-    await create_node_labels(labels)
+    await create_node_labels(
+        [
+            *labels,
+            _OSPARC_SERVICE_READY_LABEL_KEY,
+            _OSPARC_SERVICES_READY_DATETIME_LABEL_KEY,
+        ]
+    )
     monitored_nodes = await get_monitored_nodes(autoscaling_docker, node_labels=labels)
     assert len(monitored_nodes) == 1
 

services/clusters-keeper/src/simcore_service_clusters_keeper/core/settings.py

@@ -1,6 +1,6 @@
 import datetime
 from functools import cached_property
-from typing import Any, ClassVar, Final, cast
+from typing import Any, ClassVar, Final, Literal, cast
 
 from aws_library.ec2.models import EC2InstanceBootSpecific, EC2Tags
 from fastapi import FastAPI
@@ -11,7 +11,14 @@
     VersionTag,
 )
 from models_library.clusters import InternalClusterAuthentication
-from pydantic import Field, NonNegativeInt, PositiveInt, parse_obj_as, validator
+from pydantic import (
+    Field,
+    NonNegativeFloat,
+    NonNegativeInt,
+    PositiveInt,
+    parse_obj_as,
+    validator,
+)
 from settings_library.base import BaseCustomSettings
 from settings_library.docker_registry import RegistrySettings
 from settings_library.ec2 import EC2Settings
@@ -270,7 +277,9 @@ class ApplicationSettings(BaseCustomSettings, MixinLoggingSettings):
         description="defines the image tag to use for the computational backend sidecar image (NOTE: it currently defaults to use itisfoundation organisation in Dockerhub)",
     )
 
-    CLUSTERS_KEEPER_COMPUTATIONAL_BACKEND_DEFAULT_CLUSTER_AUTH: InternalClusterAuthentication = Field(
+    CLUSTERS_KEEPER_COMPUTATIONAL_BACKEND_DEFAULT_CLUSTER_AUTH: (
+        InternalClusterAuthentication
+    ) = Field(
         ...,
         description="defines the authentication of the clusters created via clusters-keeper (can be None or TLS)",
     )
@@ -280,6 +289,12 @@ class ApplicationSettings(BaseCustomSettings, MixinLoggingSettings):
         description="overrides the default number of threads in the dask-sidecars, setting it to 0 will use the default (see description in dask-sidecar)",
     )
 
+    CLUSTERS_KEEPER_DASK_WORKER_SATURATION: NonNegativeFloat | Literal["inf"] = Field(
+        default="inf",
+        description="override the dask scheduler 'worker-saturation' field"
+        ", see https://selectfrom.dev/deep-dive-into-dask-distributed-scheduler-9fdb3b36b7c7",
+    )
+
     SWARM_STACK_NAME: str = Field(
         ..., description="Stack name defined upon deploy (see main Makefile)"
     )

services/clusters-keeper/src/simcore_service_clusters_keeper/data/docker-compose.yml

@@ -11,6 +11,7 @@ services:
       DASK_TLS_KEY: ${DASK_TLS_KEY}
       DASK_SCHEDULER_URL: tls://dask-scheduler:8786
       DASK_START_AS_SCHEDULER: 1
+      DASK_WORKER_SATURATION: ${DASK_WORKER_SATURATION}
       LOG_LEVEL: ${LOG_LEVEL}
     ports:
       - 8786:8786 # dask-scheduler access
@@ -49,6 +50,7 @@ services:
       DASK_TLS_CA_FILE: ${DASK_TLS_CA_FILE}
       DASK_TLS_CERT: ${DASK_TLS_CERT}
       DASK_TLS_KEY: ${DASK_TLS_KEY}
+      DASK_WORKER_SATURATION: ${DASK_WORKER_SATURATION}
       LOG_LEVEL: ${LOG_LEVEL}
       SIDECAR_COMP_SERVICES_SHARED_FOLDER: /home/scu/computational_shared_data
       SIDECAR_COMP_SERVICES_SHARED_VOLUME_NAME: computational_shared_data