dstack 0.18.43__tar.gz → 0.18.44__tar.gz

{dstack-0.18.43 → dstack-0.18.44}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: dstack
-Version: 0.18.43
+Version: 0.18.44
 Summary: dstack is an open-source orchestration engine for running AI workloads on any cloud or on-premises.
 Home-page: https://dstack.ai
 Author: Andrey Cheptsov

{dstack-0.18.43 → dstack-0.18.44}/setup.py

@@ -76,6 +76,7 @@ SERVER_DEPS = GATEWAY_AND_SERVER_COMMON_DEPS + [
     "asyncpg",
     "cachetools",
     "python-json-logger>=3.1.0",
+    "prometheus-client",
     "grpcio>=1.50",  # indirect
 ]
 

{dstack-0.18.43 → dstack-0.18.44}/src/dstack/_internal/cli/services/configurators/run.py

@@ -95,6 +95,7 @@ class BaseRunConfigurator(ApplyEnvVarsConfiguratorMixin, BaseApplyConfigurator):
                 reservation=profile.reservation,
                 spot_policy=profile.spot_policy,
                 retry_policy=profile.retry_policy,
+                utilization_policy=profile.utilization_policy,
                 max_duration=profile.max_duration,
                 stop_duration=profile.stop_duration,
                 max_price=profile.max_price,

{dstack-0.18.43 → dstack-0.18.44}/src/dstack/_internal/cli/utils/run.py

@@ -4,6 +4,8 @@ from rich.markup import escape
 from rich.table import Table
 
 from dstack._internal.cli.utils.common import NO_OFFERS_WARNING, add_row_from_dict, console
+from dstack._internal.core.models.common import is_core_model_instance
+from dstack._internal.core.models.configurations import DevEnvironmentConfiguration
 from dstack._internal.core.models.instances import InstanceAvailability
 from dstack._internal.core.models.profiles import (
     DEFAULT_RUN_TERMINATION_IDLE_TIME,
@@ -38,6 +40,13 @@ def print_run_plan(run_plan: RunPlan, offers_limit: int = 3):
         if job_plan.job_spec.max_duration
         else "-"
     )
+    inactivity_duration = None
+    if is_core_model_instance(run_plan.run_spec.configuration, DevEnvironmentConfiguration):
+        inactivity_duration = "-"
+        if isinstance(run_plan.run_spec.configuration.inactivity_duration, int):
+            inactivity_duration = format_pretty_duration(
+                run_plan.run_spec.configuration.inactivity_duration
+            )
     if job_plan.job_spec.retry is None:
         retry = "-"
     else:
@@ -72,6 +81,8 @@ def print_run_plan(run_plan: RunPlan, offers_limit: int = 3):
     props.add_row(th("Resources"), pretty_req)
     props.add_row(th("Max price"), max_price)
     props.add_row(th("Max duration"), max_duration)
+    if inactivity_duration is not None:  # None means n/a
+        props.add_row(th("Inactivity duration"), inactivity_duration)
     props.add_row(th("Spot policy"), spot_policy)
     props.add_row(th("Retry policy"), retry)
     props.add_row(th("Creation policy"), creation_policy)

{dstack-0.18.43 → dstack-0.18.44}/src/dstack/_internal/core/backends/aws/compute.py

@@ -506,6 +506,7 @@ class AWSCompute(Compute):
                 "Failed to terminate all gateway %s resources. backend_data parsing error.",
                 configuration.instance_name,
             )
+            return
 
         elb_client = self.session.client("elbv2", region_name=configuration.region)
 

{dstack-0.18.43 → dstack-0.18.44}/src/dstack/_internal/core/backends/azure/compute.py

@@ -312,7 +312,7 @@ def get_resource_group_network_subnet_or_error(
         except Exception:
             raise ComputeError(
                 "Network specified in incorrect format."
-                " Supported format for `vps_ids` values: 'networkResourceGroupName/networkName'"
+                " Supported format for `vpc_ids` values: 'networkResourceGroupName/networkName'"
             )
     elif resource_group is not None:
         network_name = azure_utils.get_default_network_name(resource_group, location)

{dstack-0.18.43 → dstack-0.18.44}/src/dstack/_internal/core/backends/gcp/compute.py

@@ -580,7 +580,7 @@ class GCPCompute(Compute):
             operation = self.disk_client.delete(
                 project=self.config.project_id,
                 zone=get_or_error(volume.provisioning_data).availability_zone,
-                disk=volume.name,
+                disk=volume.volume_id,
             )
             gcp_resources.wait_for_extended_operation(operation, "persistent disk deletion")
         except google.api_core.exceptions.NotFound:

{dstack-0.18.43 → dstack-0.18.44}/src/dstack/_internal/core/backends/runpod/compute.py

@@ -52,8 +52,9 @@ class RunpodCompute(Compute):
     ) -> List[InstanceOfferWithAvailability]:
         offers = get_catalog_offers(
             backend=BackendType.RUNPOD,
-            locations=self.config.regions,
+            locations=self.config.regions or None,
             requirements=requirements,
+            extra_filter=lambda o: _is_secure_cloud(o.region) or self.config.allow_community_cloud,
         )
         offers = [
             InstanceOfferWithAvailability(
@@ -102,13 +103,22 @@ class RunpodCompute(Compute):
         bid_per_gpu = None
         if instance_offer.instance.resources.spot and gpu_count:
             bid_per_gpu = instance_offer.price / gpu_count
+        if _is_secure_cloud(instance_offer.region):
+            cloud_type = "SECURE"
+            data_center_id = instance_offer.region
+            country_code = None
+        else:
+            cloud_type = "COMMUNITY"
+            data_center_id = None
+            country_code = instance_offer.region
 
         resp = self.api_client.create_pod(
             name=pod_name,
             image_name=job.job_spec.image_name,
             gpu_type_id=instance_offer.instance.name,
-            cloud_type="SECURE",  # ["ALL", "COMMUNITY", "SECURE"]:
-            data_center_id=instance_offer.region,
+            cloud_type=cloud_type,
+            data_center_id=data_center_id,
+            country_code=country_code,
             gpu_count=gpu_count,
             container_disk_in_gb=disk_size,
             min_vcpu_count=instance_offer.instance.resources.cpus,
@@ -257,3 +267,11 @@ def _get_volume_price(size: int) -> float:
     if size < 1000:
         return 0.07 * size
     return 0.05 * size
+
+
+def _is_secure_cloud(region: str) -> str:
+    """
+    Secure cloud regions are datacenter IDs: CA-MTL-1, EU-NL-1, etc.
+    Community cloud regions are country codes: CA, NL, etc.
+    """
+    return "-" in region

{dstack-0.18.43 → dstack-0.18.44}/src/dstack/_internal/core/backends/runpod/config.py

@@ -4,6 +4,14 @@ from dstack._internal.core.models.backends.runpod import (
     RunpodStoredConfig,
 )
 
+RUNPOD_COMMUNITY_CLOUD_DEFAULT = True
+
 
 class RunpodConfig(RunpodStoredConfig, BackendConfig):
     creds: AnyRunpodCreds
+
+    @property
+    def allow_community_cloud(self) -> bool:
+        if self.community_cloud is not None:
+            return self.community_cloud
+        return RUNPOD_COMMUNITY_CLOUD_DEFAULT

{dstack-0.18.43 → dstack-0.18.44}/src/dstack/_internal/core/models/backends/runpod.py

@@ -10,6 +10,7 @@ from dstack._internal.core.models.common import CoreModel
 class RunpodConfigInfo(CoreModel):
     type: Literal["runpod"] = "runpod"
     regions: Optional[List[str]] = None
+    community_cloud: Optional[bool] = None
 
 
 class RunpodStoredConfig(RunpodConfigInfo):
@@ -33,6 +34,7 @@ class RunpodConfigInfoWithCredsPartial(CoreModel):
     type: Literal["runpod"] = "runpod"
     creds: Optional[AnyRunpodCreds]
     regions: Optional[List[str]]
+    community_cloud: Optional[bool]
 
 
 class RunpodConfigValues(CoreModel):

{dstack-0.18.43 → dstack-0.18.44}/src/dstack/_internal/core/models/configurations.py

@@ -221,7 +221,8 @@ class DevEnvironmentConfigurationParams(CoreModel):
                 " Inactivity is defined as the absence of SSH connections to the"
                 " dev environment, including VS Code connections, `ssh <run name>`"
                 " shells, and attached `dstack apply` or `dstack attach` commands."
-                " Use `off` for unlimited duration. Defaults to `off`"
+                " Use `off` for unlimited duration. Can be updated in-place."
+                " Defaults to `off`"
             )
         ),
     ]

{dstack-0.18.43 → dstack-0.18.44}/src/dstack/_internal/core/models/profiles.py

@@ -1,5 +1,5 @@
 from enum import Enum
-from typing import List, Optional, Union
+from typing import List, Optional, Union, overload
 
 from pydantic import Field, root_validator, validator
 from typing_extensions import Annotated, Literal
@@ -34,6 +34,14 @@ class TerminationPolicy(str, Enum):
     DESTROY_AFTER_IDLE = "destroy-after-idle"
 
 
+@overload
+def parse_duration(v: None) -> None: ...
+
+
+@overload
+def parse_duration(v: Union[int, str]) -> int: ...
+
+
 def parse_duration(v: Optional[Union[int, str]]) -> Optional[int]:
     if v is None:
         return None
@@ -112,6 +120,39 @@ class ProfileRetry(CoreModel):
         return values
 
 
+class UtilizationPolicy(CoreModel):
+    _min_time_window = "5m"
+
+    min_gpu_utilization: Annotated[
+        int,
+        Field(
+            description=(
+                "Minimum required GPU utilization, percent."
+                " If any GPU has utilization below specified value during the whole time window,"
+                " the run is terminated"
+            ),
+            ge=0,
+            le=100,
+        ),
+    ]
+    time_window: Annotated[
+        Union[int, str],
+        Field(
+            description=(
+                "The time window of metric samples taking into account to measure utilization"
+                f" (e.g., `30m`, `1h`). Minimum is `{_min_time_window}`"
+            )
+        ),
+    ]
+
+    @validator("time_window", pre=True)
+    def validate_time_window(cls, v: Union[int, str]) -> int:
+        v = parse_duration(v)
+        if v < parse_duration(cls._min_time_window):
+            raise ValueError(f"Minimum time_window is {cls._min_time_window}")
+        return v
+
+
 class ProfileParams(CoreModel):
     backends: Annotated[
         Optional[List[BackendType]],
@@ -194,6 +235,10 @@ class ProfileParams(CoreModel):
             )
         ),
     ]
+    utilization_policy: Annotated[
+        Optional[UtilizationPolicy],
+        Field(description="Run termination policy based on utilization"),
+    ]
     # Deprecated:
     termination_policy: Annotated[
         Optional[TerminationPolicy],

{dstack-0.18.43 → dstack-0.18.44}/src/dstack/_internal/core/models/runs.py

@@ -23,6 +23,7 @@ from dstack._internal.core.models.profiles import (
     ProfileRetryPolicy,
     RetryEvent,
     SpotPolicy,
+    UtilizationPolicy,
 )
 from dstack._internal.core.models.repos import AnyRunRepoData
 from dstack._internal.core.models.resources import Memory, ResourcesSpec
@@ -114,6 +115,7 @@ class JobTerminationReason(str, Enum):
     ABORTED_BY_USER = "aborted_by_user"
     TERMINATED_BY_SERVER = "terminated_by_server"
     INACTIVITY_DURATION_EXCEEDED = "inactivity_duration_exceeded"
+    TERMINATED_DUE_TO_UTILIZATION_POLICY = "terminated_due_to_utilization_policy"
     # Set by the runner
     CONTAINER_EXITED_WITH_ERROR = "container_exited_with_error"
     PORTS_BINDING_FAILED = "ports_binding_failed"
@@ -135,6 +137,7 @@ class JobTerminationReason(str, Enum):
             self.ABORTED_BY_USER: JobStatus.ABORTED,
             self.TERMINATED_BY_SERVER: JobStatus.TERMINATED,
             self.INACTIVITY_DURATION_EXCEEDED: JobStatus.TERMINATED,
+            self.TERMINATED_DUE_TO_UTILIZATION_POLICY: JobStatus.TERMINATED,
             self.CONTAINER_EXITED_WITH_ERROR: JobStatus.FAILED,
             self.PORTS_BINDING_FAILED: JobStatus.FAILED,
             self.CREATING_CONTAINER_ERROR: JobStatus.FAILED,
@@ -190,6 +193,7 @@ class JobSpec(CoreModel):
     single_branch: Optional[bool] = None
     max_duration: Optional[int]
     stop_duration: Optional[int] = None
+    utilization_policy: Optional[UtilizationPolicy] = None
     registry_auth: Optional[RegistryAuth]
     requirements: Requirements
     retry: Optional[Retry]

{dstack-0.18.43 → dstack-0.18.44}/src/dstack/_internal/server/app.py

@@ -29,6 +29,7 @@ from dstack._internal.server.routers import (
     metrics,
     pools,
     projects,
+    prometheus,
     repos,
     runs,
     secrets,
@@ -185,6 +186,7 @@ def register_routes(app: FastAPI, ui: bool = True):
     app.include_router(model_proxy.router, prefix="/proxy/models", tags=["model-proxy"])
     app.include_router(pools.root_router)
     app.include_router(pools.router)
+    app.include_router(prometheus.router)
 
     @app.exception_handler(ForbiddenError)
     async def forbidden_error_handler(request: Request, exc: ForbiddenError):
@@ -252,7 +254,11 @@ def register_routes(app: FastAPI, ui: bool = True):
 
         @app.exception_handler(404)
         async def custom_http_exception_handler(request, exc):
-            if request.url.path.startswith("/api") or _is_proxy_request(request):
+            if (
+                request.url.path.startswith("/api")
+                or _is_proxy_request(request)
+                or _is_prometheus_request(request)
+            ):
                 return JSONResponse(
                     {"detail": exc.detail},
                     status_code=status.HTTP_404_NOT_FOUND,
@@ -283,6 +289,10 @@ def _is_proxy_request(request: Request) -> bool:
     ) and referrer.path.startswith("/proxy")
 
 
+def _is_prometheus_request(request: Request) -> bool:
+    return request.url.path.startswith("/metrics")
+
+
 def _print_dstack_logo():
     console.print(
         """[purple]╱╱╭╮╱╱╭╮╱╱╱╱╱╱╭╮

{dstack-0.18.43 → dstack-0.18.44}/src/dstack/_internal/server/background/__init__.py

@@ -1,6 +1,7 @@
 from apscheduler.schedulers.asyncio import AsyncIOScheduler
 from apscheduler.triggers.interval import IntervalTrigger
 
+from dstack._internal.server import settings
 from dstack._internal.server.background.tasks.process_fleets import process_fleets
 from dstack._internal.server.background.tasks.process_gateways import (
     process_gateways_connections,
@@ -16,6 +17,10 @@ from dstack._internal.server.background.tasks.process_metrics import (
 from dstack._internal.server.background.tasks.process_placement_groups import (
     process_placement_groups,
 )
+from dstack._internal.server.background.tasks.process_prometheus_metrics import (
+    collect_prometheus_metrics,
+    delete_prometheus_metrics,
+)
 from dstack._internal.server.background.tasks.process_running_jobs import process_running_jobs
 from dstack._internal.server.background.tasks.process_runs import process_runs
 from dstack._internal.server.background.tasks.process_submitted_jobs import process_submitted_jobs
@@ -43,6 +48,11 @@ def start_background_tasks() -> AsyncIOScheduler:
     # * 150 active instances with up to 2 minutes processing latency
     _scheduler.add_job(collect_metrics, IntervalTrigger(seconds=10), max_instances=1)
     _scheduler.add_job(delete_metrics, IntervalTrigger(minutes=5), max_instances=1)
+    if settings.ENABLE_PROMETHEUS_METRICS:
+        _scheduler.add_job(
+            collect_prometheus_metrics, IntervalTrigger(seconds=10), max_instances=1
+        )
+        _scheduler.add_job(delete_prometheus_metrics, IntervalTrigger(minutes=5), max_instances=1)
     # process_submitted_jobs and process_instances max processing rate is 75 jobs(instances) per minute.
     _scheduler.add_job(
         process_submitted_jobs,

{dstack-0.18.43 → dstack-0.18.44}/src/dstack/_internal/server/background/tasks/process_placement_groups.py

@@ -28,6 +28,7 @@ async def process_placement_groups():
                     PlacementGroupModel.deleted == False,
                     PlacementGroupModel.id.not_in(lockset),
                 )
+                .order_by(PlacementGroupModel.id)  # take locks in order
                 .with_for_update(skip_locked=True)
             )
             placement_group_models = res.scalars().all()

dstack-0.18.44/src/dstack/_internal/server/background/tasks/process_prometheus_metrics.py

@@ -0,0 +1,135 @@
+import uuid
+from datetime import datetime, timedelta
+from typing import Optional
+
+import sqlalchemy.exc
+from sqlalchemy import delete, or_, select, update
+from sqlalchemy.orm import joinedload
+
+from dstack._internal.core.consts import DSTACK_SHIM_HTTP_PORT
+from dstack._internal.core.models.runs import JobStatus
+from dstack._internal.server.db import get_session_ctx
+from dstack._internal.server.models import InstanceModel, JobModel, JobPrometheusMetrics
+from dstack._internal.server.services.jobs import get_job_provisioning_data, get_job_runtime_data
+from dstack._internal.server.services.pools import get_instance_ssh_private_keys
+from dstack._internal.server.services.runner import client
+from dstack._internal.server.services.runner.ssh import runner_ssh_tunnel
+from dstack._internal.server.utils.common import gather_map_async
+from dstack._internal.utils.common import batched, get_current_datetime, get_or_error, run_async
+from dstack._internal.utils.logging import get_logger
+
+logger = get_logger(__name__)
+
+
+MAX_JOBS_FETCHED = 100
+BATCH_SIZE = 10
+MIN_COLLECT_INTERVAL_SECONDS = 9
+# 10 minutes should be more than enough to scrape metrics, and, in any case,
+# 10 minutes old metrics has little to no value
+METRICS_TTL_SECONDS = 600
+
+
+async def collect_prometheus_metrics():
+    now = get_current_datetime()
+    cutoff = now - timedelta(seconds=MIN_COLLECT_INTERVAL_SECONDS)
+    async with get_session_ctx() as session:
+        res = await session.execute(
+            select(JobModel)
+            .join(JobPrometheusMetrics, isouter=True)
+            .where(
+                JobModel.status.in_([JobStatus.RUNNING]),
+                or_(
+                    JobPrometheusMetrics.job_id.is_(None),
+                    JobPrometheusMetrics.collected_at < cutoff,
+                ),
+            )
+            .options(joinedload(JobModel.instance).joinedload(InstanceModel.project))
+            .order_by(JobModel.last_processed_at.asc())
+            .limit(MAX_JOBS_FETCHED)
+        )
+        job_models = res.unique().scalars().all()
+    for batch in batched(job_models, BATCH_SIZE):
+        await _collect_jobs_metrics(batch, now)
+
+
+async def delete_prometheus_metrics():
+    now = get_current_datetime()
+    cutoff = now - timedelta(seconds=METRICS_TTL_SECONDS)
+    async with get_session_ctx() as session:
+        await session.execute(
+            delete(JobPrometheusMetrics).where(JobPrometheusMetrics.collected_at < cutoff)
+        )
+        await session.commit()
+
+
+async def _collect_jobs_metrics(job_models: list[JobModel], collected_at: datetime):
+    results = await gather_map_async(job_models, _collect_job_metrics, return_exceptions=True)
+    async with get_session_ctx() as session:
+        for job_model, result in results:
+            if result is None:
+                continue
+            if isinstance(result, BaseException):
+                logger.error(
+                    "Failed to collect job %s Prometheus metrics: %r", job_model.job_name, result
+                )
+                continue
+            res = await session.execute(
+                update(JobPrometheusMetrics)
+                .where(JobPrometheusMetrics.job_id == job_model.id)
+                .values(
+                    collected_at=collected_at,
+                    text=result,
+                )
+                .returning(JobPrometheusMetrics)
+            )
+            metrics = res.scalar()
+            if metrics is None:
+                metrics = JobPrometheusMetrics(
+                    job_id=job_model.id,
+                    collected_at=collected_at,
+                    text=result,
+                )
+                try:
+                    async with session.begin_nested():
+                        session.add(metrics)
+                except sqlalchemy.exc.IntegrityError:
+                    # Concurrent server replica already committed, ignoring
+                    pass
+        await session.commit()
+
+
+async def _collect_job_metrics(job_model: JobModel) -> Optional[str]:
+    ssh_private_keys = get_instance_ssh_private_keys(get_or_error(job_model.instance))
+    jpd = get_job_provisioning_data(job_model)
+    jrd = get_job_runtime_data(job_model)
+    if jpd is None:
+        return None
+    try:
+        res = await run_async(
+            _pull_job_metrics,
+            ssh_private_keys,
+            jpd,
+            jrd,
+            job_model.id,
+        )
+    except Exception:
+        logger.exception("Failed to collect job %s Prometheus metrics", job_model.job_name)
+        return None
+
+    if isinstance(res, bool):
+        logger.warning(
+            "Failed to connect to job %s to collect Prometheus metrics", job_model.job_name
+        )
+        return None
+
+    if res is None:
+        # Either not supported by shim or exporter is not available
+        return None
+
+    return res
+
+
+@runner_ssh_tunnel(ports=[DSTACK_SHIM_HTTP_PORT], retries=1)
+def _pull_job_metrics(ports: dict[int, int], task_id: uuid.UUID) -> Optional[str]:
+    shim_client = client.ShimClient(port=ports[DSTACK_SHIM_HTTP_PORT])
+    return shim_client.get_task_metrics(task_id)

{dstack-0.18.43 → dstack-0.18.44}/src/dstack/_internal/server/background/tasks/process_running_jobs.py

@@ -1,4 +1,6 @@
 import asyncio
+from collections.abc import Iterable
+from datetime import timedelta
 from typing import Dict, List, Optional
 
 from sqlalchemy import select
@@ -15,6 +17,7 @@ from dstack._internal.core.models.instances import (
     RemoteConnectionInfo,
     SSHConnectionParams,
 )
+from dstack._internal.core.models.metrics import Metric
 from dstack._internal.core.models.repos import RemoteRepoCreds
 from dstack._internal.core.models.runs import (
     ClusterInfo,
@@ -48,6 +51,7 @@ from dstack._internal.server.services.jobs import (
 )
 from dstack._internal.server.services.locking import get_locker
 from dstack._internal.server.services.logging import fmt
+from dstack._internal.server.services.metrics import get_job_metrics
 from dstack._internal.server.services.pools import get_instance_ssh_private_keys
 from dstack._internal.server.services.repos import (
     get_code_model,
@@ -343,6 +347,9 @@ async def _process_running_job(session: AsyncSession, job_model: JobModel):
             job_model.status = JobStatus.TERMINATING
             job_model.termination_reason = JobTerminationReason.GATEWAY_ERROR
 
+    if job_model.status == JobStatus.RUNNING:
+        await _check_gpu_utilization(session, job_model, job)
+
     job_model.last_processed_at = common_utils.get_current_datetime()
     await session.commit()
 
@@ -646,27 +653,67 @@ def _terminate_if_inactivity_duration_exceeded(
     run_model: RunModel, job_model: JobModel, no_connections_secs: Optional[int]
 ) -> None:
     conf = RunSpec.__response__.parse_raw(run_model.run_spec).configuration
-    if is_core_model_instance(conf, DevEnvironmentConfiguration) and isinstance(
+    if not is_core_model_instance(conf, DevEnvironmentConfiguration) or not isinstance(
         conf.inactivity_duration, int
     ):
-        logger.debug("%s: no SSH connections for %s seconds", fmt(job_model), no_connections_secs)
-        job_model.inactivity_secs = no_connections_secs
-        if no_connections_secs is None:
-            # TODO(0.19 or earlier): make no_connections_secs required
-            job_model.status = JobStatus.TERMINATING
-            job_model.termination_reason = JobTerminationReason.INTERRUPTED_BY_NO_CAPACITY
-            job_model.termination_reason_message = (
-                "The selected instance was created before dstack 0.18.41"
-                " and does not support inactivity_duration"
-            )
-        elif no_connections_secs >= conf.inactivity_duration:
-            job_model.status = JobStatus.TERMINATING
-            # TODO(0.19 or earlier): set JobTerminationReason.INACTIVITY_DURATION_EXCEEDED
-            job_model.termination_reason = JobTerminationReason.TERMINATED_BY_SERVER
-            job_model.termination_reason_message = (
-                f"The job was inactive for {no_connections_secs} seconds,"
-                f" exceeding the inactivity_duration of {conf.inactivity_duration} seconds"
-            )
+        # reset in case inactivity_duration was disabled via in-place update
+        job_model.inactivity_secs = None
+        return
+    logger.debug("%s: no SSH connections for %s seconds", fmt(job_model), no_connections_secs)
+    job_model.inactivity_secs = no_connections_secs
+    if no_connections_secs is None:
+        # TODO(0.19 or earlier): make no_connections_secs required
+        job_model.status = JobStatus.TERMINATING
+        job_model.termination_reason = JobTerminationReason.INTERRUPTED_BY_NO_CAPACITY
+        job_model.termination_reason_message = (
+            "The selected instance was created before dstack 0.18.41"
+            " and does not support inactivity_duration"
+        )
+    elif no_connections_secs >= conf.inactivity_duration:
+        job_model.status = JobStatus.TERMINATING
+        # TODO(0.19 or earlier): set JobTerminationReason.INACTIVITY_DURATION_EXCEEDED
+        job_model.termination_reason = JobTerminationReason.TERMINATED_BY_SERVER
+        job_model.termination_reason_message = (
+            f"The job was inactive for {no_connections_secs} seconds,"
+            f" exceeding the inactivity_duration of {conf.inactivity_duration} seconds"
+        )
+
+
+async def _check_gpu_utilization(session: AsyncSession, job_model: JobModel, job: Job) -> None:
+    policy = job.job_spec.utilization_policy
+    if policy is None:
+        return
+    after = common_utils.get_current_datetime() - timedelta(seconds=policy.time_window)
+    job_metrics = await get_job_metrics(session, job_model, after=after)
+    gpus_util_metrics: list[Metric] = []
+    for metric in job_metrics.metrics:
+        if metric.name.startswith("gpu_util_percent_gpu"):
+            gpus_util_metrics.append(metric)
+    if not gpus_util_metrics or gpus_util_metrics[0].timestamps[-1] > after + timedelta(minutes=1):
+        # Job has started recently, not enough points collected.
+        # Assuming that metrics collection interval less than 1 minute.
+        logger.debug("%s: GPU utilization check: not enough samples", fmt(job_model))
+        return
+    if _should_terminate_due_to_low_gpu_util(
+        policy.min_gpu_utilization, [m.values for m in gpus_util_metrics]
+    ):
+        logger.info("%s: GPU utilization check: terminating", fmt(job_model))
+        job_model.status = JobStatus.TERMINATING
+        # TODO(0.19 or earlier): set JobTerminationReason.TERMINATED_DUE_TO_UTILIZATION_POLICY
+        job_model.termination_reason = JobTerminationReason.TERMINATED_BY_SERVER
+        job_model.termination_reason_message = (
+            f"The job GPU utilization below {policy.min_gpu_utilization}%"
+            f" for {policy.time_window} seconds"
+        )
+    else:
+        logger.debug("%s: GPU utilization check: OK", fmt(job_model))
+
+
+def _should_terminate_due_to_low_gpu_util(min_util: int, gpus_util: Iterable[Iterable[int]]):
+    for gpu_util in gpus_util:
+        if all(util < min_util for util in gpu_util):
+            return True
+    return False
 
 
 def _get_cluster_info(

{dstack-0.18.43 → dstack-0.18.44}/src/dstack/_internal/server/background/tasks/process_runs.py

@@ -74,6 +74,7 @@ async def _process_next_run():
                     JobModel.run_id == run_model.id,
                     JobModel.id.not_in(job_lockset),
                 )
+                .order_by(JobModel.id)  # take locks in order
                 .with_for_update(skip_locked=True)
             )
             job_models = res.scalars().all()

{dstack-0.18.43 → dstack-0.18.44}/src/dstack/_internal/server/background/tasks/process_submitted_jobs.py

@@ -35,6 +35,7 @@ from dstack._internal.core.models.runs import (
 )
 from dstack._internal.core.models.volumes import Volume
 from dstack._internal.core.services.profiles import get_termination
+from dstack._internal.server import settings
 from dstack._internal.server.db import get_db, get_session_ctx
 from dstack._internal.server.models import (
     FleetModel,
@@ -195,6 +196,7 @@ async def _process_submitted_job(session: AsyncSession, job_model: JobModel):
                 InstanceModel.total_blocks > InstanceModel.busy_blocks,
             )
             .options(lazyload(InstanceModel.jobs))
+            .order_by(InstanceModel.id)  # take locks in order
             .with_for_update()
         )
         pool_instances = list(res.unique().scalars().all())
@@ -319,6 +321,7 @@ async def _process_submitted_job(session: AsyncSession, job_model: JobModel):
         select(VolumeModel)
         .where(VolumeModel.id.in_(volumes_ids))
         .options(selectinload(VolumeModel.user))
+        .order_by(VolumeModel.id)  # take locks in order
         .with_for_update()
     )
     async with get_locker().lock_ctx(VolumeModel.__tablename__, volumes_ids):
@@ -450,7 +453,7 @@ async def _run_job_on_new_instance(
     )
     # Limit number of offers tried to prevent long-running processing
     # in case all offers fail.
-    for backend, offer in offers[:15]:
+    for backend, offer in offers[: settings.MAX_OFFERS_TRIED]:
         logger.debug(
             "%s: trying %s in %s/%s for $%0.4f per hour",
             fmt(job_model),