Do not create clusters from burstable AWS instances

Fixes #2872

Author: r4victor

src/dstack/_internal/core/backends/aws/compute.py

@@ -33,7 +33,12 @@
     merge_tags,
 )
 from dstack._internal.core.backends.base.offers import get_catalog_offers
-from dstack._internal.core.errors import ComputeError, NoCapacityError, PlacementGroupInUseError
+from dstack._internal.core.errors import (
+    ComputeError,
+    NoCapacityError,
+    PlacementGroupInUseError,
+    PlacementGroupNotSupportedError,
+)
 from dstack._internal.core.models.backends.base import BackendType
 from dstack._internal.core.models.common import CoreModel
 from dstack._internal.core.models.gateways import (
@@ -46,7 +51,11 @@
     InstanceOffer,
     InstanceOfferWithAvailability,
 )
-from dstack._internal.core.models.placement import PlacementGroup, PlacementGroupProvisioningData
+from dstack._internal.core.models.placement import (
+    PlacementGroup,
+    PlacementGroupProvisioningData,
+    PlacementStrategy,
+)
 from dstack._internal.core.models.resources import Memory, Range
 from dstack._internal.core.models.runs import JobProvisioningData, Requirements
 from dstack._internal.core.models.volumes import (
@@ -334,6 +343,8 @@ def create_placement_group(
         placement_group: PlacementGroup,
         master_instance_offer: InstanceOffer,
     ) -> PlacementGroupProvisioningData:
+        if not _offer_supports_placement_group(master_instance_offer, placement_group):
+            raise PlacementGroupNotSupportedError()
         ec2_client = self.session.client("ec2", region_name=placement_group.configuration.region)
         logger.debug("Creating placement group %s...", placement_group.name)
         ec2_client.create_placement_group(
@@ -370,6 +381,8 @@ def is_suitable_placement_group(
         placement_group: PlacementGroup,
         instance_offer: InstanceOffer,
     ) -> bool:
+        if not _offer_supports_placement_group(instance_offer, placement_group):
+            return False
         return (
             placement_group.configuration.backend == BackendType.AWS
             and placement_group.configuration.region == instance_offer.region
@@ -1059,6 +1072,15 @@ def _supported_instances(offer: InstanceOffer) -> bool:
     return False
 
 
+def _offer_supports_placement_group(offer: InstanceOffer, placement_group: PlacementGroup) -> bool:
+    if placement_group.configuration.placement_strategy != PlacementStrategy.CLUSTER:
+        return True
+    for family in ["t3.", "t2."]:
+        if offer.instance.name.startswith(family):
+            return False
+    return True
+
+
 def _get_maximum_efa_interfaces(ec2_client: botocore.client.BaseClient, instance_type: str) -> int:
     try:
         response = ec2_client.describe_instance_types(

src/dstack/_internal/core/errors.py

@@ -110,6 +110,10 @@ class PlacementGroupInUseError(ComputeError):
     pass
 
 
+class PlacementGroupNotSupportedError(ComputeError):
+    pass
+
+
 class NotYetTerminated(ComputeError):
     """
     Used by Compute.terminate_instance to signal that instance termination is not complete

src/dstack/_internal/server/background/__init__.py

@@ -109,7 +109,7 @@ def start_background_tasks() -> AsyncIOScheduler:
             process_instances,
             IntervalTrigger(seconds=4, jitter=2),
             kwargs={"batch_size": 5},
-            max_instances=4 if replica == 0 else 1,
+            max_instances=2 if replica == 0 else 1,
         )
         _scheduler.add_job(
             process_fleets,

src/dstack/_internal/server/background/tasks/process_instances.py

@@ -45,6 +45,7 @@
 from dstack._internal.core.errors import (
     BackendError,
     NotYetTerminated,
+    PlacementGroupNotSupportedError,
     ProvisioningError,
 )
 from dstack._internal.core.models.backends.base import BackendType
@@ -1067,6 +1068,12 @@ async def _create_placement_group(
             placement_group_model_to_placement_group(placement_group_model),
             master_instance_offer,
         )
+    except PlacementGroupNotSupportedError:
+        logger.debug(
+            "Skipping offer %s because placement group not supported",
+            master_instance_offer.instance.name,
+        )
+        return None
     except BackendError as e:
         logger.warning(
             "Failed to create placement group %s in %s/%s: %r",