skypilot-nightly 1.0.0.dev20250329__py3-none-any.whl → 1.0.0.dev20250331__py3-none-any.whl

sky/__init__.py

@@ -5,7 +5,7 @@ from typing import Optional
 import urllib.request
 
 # Replaced with the current commit when building the wheels.
-_SKYPILOT_COMMIT_SHA = '9e6cef22ae679a1fef4539cbfc02a6d5cdc405ed'
+_SKYPILOT_COMMIT_SHA = 'f90ccc1757680ccbff2fb8d86fc6dfd4242bd182'
 
 
 def _get_git_commit():
@@ -35,7 +35,7 @@ def _get_git_commit():
 
 
 __commit__ = _get_git_commit()
-__version__ = '1.0.0.dev20250329'
+__version__ = '1.0.0.dev20250331'
 __root_dir__ = os.path.dirname(os.path.abspath(__file__))
 
 

sky/cli.py

@@ -3943,6 +3943,7 @@ def jobs_launch(
               required=False,
               help='Show jobs from all users.')
 @click.option('--all',
+              '-a',
               default=False,
               is_flag=True,
               required=False,
@@ -4394,6 +4395,7 @@ def serve_up(
     )
     click.secho('Service spec:', fg='cyan')
     click.echo(task.service)
+    serve_lib.validate_service_task(task)
 
     click.secho('Each replica will use the following resources (estimated):',
                 fg='cyan')
@@ -4493,6 +4495,7 @@ def serve_update(service_name: str, service_yaml: Tuple[str, ...],
     )
     click.secho('Service spec:', fg='cyan')
     click.echo(task.service)
+    serve_lib.validate_service_task(task)
 
     click.secho('New replica will use the following resources (estimated):',
                 fg='cyan')

sky/client/cli.py

@@ -3943,6 +3943,7 @@ def jobs_launch(
               required=False,
               help='Show jobs from all users.')
 @click.option('--all',
+              '-a',
               default=False,
               is_flag=True,
               required=False,
@@ -4394,6 +4395,7 @@ def serve_up(
     )
     click.secho('Service spec:', fg='cyan')
     click.echo(task.service)
+    serve_lib.validate_service_task(task)
 
     click.secho('Each replica will use the following resources (estimated):',
                 fg='cyan')
@@ -4493,6 +4495,7 @@ def serve_update(service_name: str, service_yaml: Tuple[str, ...],
     )
     click.secho('Service spec:', fg='cyan')
     click.echo(task.service)
+    serve_lib.validate_service_task(task)
 
     click.secho('New replica will use the following resources (estimated):',
                 fg='cyan')

sky/data/storage_utils.py

@@ -18,12 +18,6 @@ from sky.utils import log_utils
 
 logger = sky_logging.init_logger(__name__)
 
-_FILE_EXCLUSION_FROM_GITIGNORE_FAILURE_MSG = (
-    f'{colorama.Fore.YELLOW}Warning: Files/dirs '
-    'specified in .gitignore will be uploaded '
-    'to the cloud storage for {path!r}'
-    'due to the following error: {error_msg!r}')
-
 _USE_SKYIGNORE_HINT = (
     'To avoid using .gitignore, you can create a .skyignore file instead.')
 
@@ -172,7 +166,7 @@ def get_excluded_files_from_gitignore(src_dir_path: str) -> List[str]:
     submodules = submodules_output.stdout.split('\0')[:-1]
 
     # The empty string is the relative reference to the src_dir_path.
-    all_git_repos = ['.'] + [
+    all_git_repos = [''] + [
         # We only care about submodules that are a subdirectory of src_dir_path.
         submodule for submodule in submodules if not submodule.startswith('../')
     ]
@@ -208,7 +202,7 @@ def get_excluded_files_from_gitignore(src_dir_path: str) -> List[str]:
 
         for item in output_list:
 
-            if repo == '.' and item == './':
+            if repo == '' and item == './':
                 logger.warning(f'{src_dir_path} is within a git repo, but the '
                                'entire directory is ignored by git. We will '
                                'ignore all git exclusions. '

sky/optimizer.py

@@ -1140,50 +1140,6 @@ class DummyCloud(clouds.Cloud):
     pass
 
 
-def _make_launchables_for_valid_region_zones(
-    launchable_resources: resources_lib.Resources
-) -> List[resources_lib.Resources]:
-    assert launchable_resources.is_launchable()
-    # In principle, all provisioning requests should be made at the granularity
-    # of a single zone. However, for on-demand instances, we batch the requests
-    # to the zones in the same region in order to leverage the region-level
-    # provisioning APIs of AWS and Azure. This way, we can reduce the number of
-    # API calls, and thus the overall failover time. Note that this optimization
-    # does not affect the user cost since the clouds charge the same prices for
-    # on-demand instances in the same region regardless of the zones. On the
-    # other hand, for spot instances, we do not batch the requests because the
-    # "AWS" spot prices may vary across zones.
-    # For GCP, we do not batch the requests because GCP reservation system is
-    # zone based. Therefore, price estimation is potentially different across
-    # zones.
-
-    # NOTE(woosuk): GCP does not support region-level provisioning APIs. Thus,
-    # while we return per-region resources here, the provisioner will still
-    # issue the request for one zone at a time.
-    # NOTE(woosuk): If we support Azure spot instances, we should batch the
-    # requests since Azure spot prices are region-level.
-    # TODO(woosuk): Batch the per-zone AWS spot instance requests if they are
-    # in the same region and have the same price.
-    # TODO(woosuk): A better design is to implement batching at a higher level
-    # (e.g., in provisioner or optimizer), not here.
-    launchables = []
-    regions = launchable_resources.get_valid_regions_for_launchable()
-    for region in regions:
-        if (launchable_resources.use_spot and region.zones is not None or
-                launchable_resources.cloud.optimize_by_zone()):
-            # Spot instances.
-            # Do not batch the per-zone requests.
-            for zone in region.zones:
-                launchables.append(
-                    launchable_resources.copy(region=region.name,
-                                              zone=zone.name))
-        else:
-            # On-demand instances.
-            # Batch the requests at the granularity of a single region.
-            launchables.append(launchable_resources.copy(region=region.name))
-    return launchables
-
-
 def _filter_out_blocked_launchable_resources(
         launchable_resources: Iterable[resources_lib.Resources],
         blocked_resources: Iterable[resources_lib.Resources]):
@@ -1313,7 +1269,8 @@ def _fill_in_launchable_resources(
                 cheapest = feasible_resources.resources_list[0]
                 # Generate region/zone-specified resources.
                 launchable[resources].extend(
-                    _make_launchables_for_valid_region_zones(cheapest))
+                    resources_utils.make_launchables_for_valid_region_zones(
+                        cheapest))
                 cloud_candidates[cloud] = feasible_resources.resources_list
             else:
                 all_fuzzy_candidates.update(

sky/serve/__init__.py

@@ -21,6 +21,7 @@ from sky.serve.serve_utils import generate_service_name
 from sky.serve.serve_utils import ServeCodeGen
 from sky.serve.serve_utils import ServiceComponent
 from sky.serve.serve_utils import UpdateMode
+from sky.serve.serve_utils import validate_service_task
 from sky.serve.service_spec import SkyServiceSpec
 
 os.makedirs(os.path.expanduser(SKYSERVE_METADATA_DIR), exist_ok=True)

sky/serve/autoscalers.py

@@ -1,5 +1,6 @@
 """Autoscalers: perform autoscaling by monitoring metrics."""
 import bisect
+import copy
 import dataclasses
 import enum
 import math
@@ -56,8 +57,8 @@ class AutoscalerDecision:
 def _generate_scale_up_decisions(
         num: int, target: Optional[Dict[str, Any]]) -> List[AutoscalerDecision]:
     return [
-        AutoscalerDecision(AutoscalerDecisionOperator.SCALE_UP, target)
-        for _ in range(num)
+        AutoscalerDecision(AutoscalerDecisionOperator.SCALE_UP,
+                           copy.copy(target)) for _ in range(num)
     ]
 
 
@@ -134,6 +135,7 @@ class Autoscaler:
         self.min_replicas: int = spec.min_replicas
         self.max_replicas: int = (spec.max_replicas if spec.max_replicas
                                   is not None else spec.min_replicas)
+        self.num_overprovision: Optional[int] = spec.num_overprovision
         # Target number of replicas is initialized to min replicas
         self.target_num_replicas: int = spec.min_replicas
         self.latest_version: int = constants.INITIAL_VERSION
@@ -143,6 +145,12 @@ class Autoscaler:
         self.latest_version_ever_ready: int = self.latest_version - 1
         self.update_mode = serve_utils.DEFAULT_UPDATE_MODE
 
+    def get_final_target_num_replicas(self) -> int:
+        """Get the final target number of replicas."""
+        if self.num_overprovision is None:
+            return self.target_num_replicas
+        return self.target_num_replicas + self.num_overprovision
+
     def _calculate_target_num_replicas(self) -> int:
         """Calculate target number of replicas."""
         raise NotImplementedError
@@ -207,7 +215,7 @@ class Autoscaler:
         0, to make the service scale faster when the service is not running.
         This will happen when min_replicas = 0 and no traffic.
         """
-        if self.target_num_replicas == 0:
+        if self.get_final_target_num_replicas() == 0:
             return constants.AUTOSCALER_NO_REPLICA_DECISION_INTERVAL_SECONDS
         else:
             return constants.AUTOSCALER_DEFAULT_DECISION_INTERVAL_SECONDS
@@ -236,13 +244,14 @@ class Autoscaler:
             # old and latest versions are allowed in rolling update, this will
             # not affect the time it takes for the service to updated to the
             # latest version.
-            if num_latest_ready_replicas >= self.target_num_replicas:
+            if (num_latest_ready_replicas >=
+                    self.get_final_target_num_replicas()):
                 # Once the number of ready new replicas is greater than or equal
                 # to the target, we can scale down all old replicas.
                 return [info.replica_id for info in old_nonterminal_replicas]
             # If rolling update is in progress, we scale down old replicas
             # based on the number of ready new replicas.
-            num_old_replicas_to_keep = (self.target_num_replicas -
+            num_old_replicas_to_keep = (self.get_final_target_num_replicas() -
                                         num_latest_ready_replicas)
             # Remove old replicas (especially old launching replicas) and only
             # keep the required number of replicas, as we want to let the new
@@ -422,6 +431,7 @@ class _AutoscalerWithHysteresis(Autoscaler):
             f'Old target number of replicas: {old_target_num_replicas}. '
             f'Current target number of replicas: {target_num_replicas}. '
             f'Final target number of replicas: {self.target_num_replicas}. '
+            f'Num overprovision: {self.num_overprovision}. '
             f'Upscale counter: {self.upscale_counter}/'
             f'{self.scale_up_threshold}. '
             f'Downscale counter: {self.downscale_counter}/'
@@ -505,8 +515,9 @@ class RequestRateAutoscaler(_AutoscalerWithHysteresis):
 
         # Case 1. when latest_nonterminal_replicas is less
         # than num_to_provision, we always scale up new replicas.
-        if len(latest_nonterminal_replicas) < self.target_num_replicas:
-            num_replicas_to_scale_up = (self.target_num_replicas -
+        target_num_replicas = self.get_final_target_num_replicas()
+        if len(latest_nonterminal_replicas) < target_num_replicas:
+            num_replicas_to_scale_up = (target_num_replicas -
                                         len(latest_nonterminal_replicas))
             logger.info('Number of replicas to scale up: '
                         f'{num_replicas_to_scale_up}')
@@ -514,11 +525,11 @@ class RequestRateAutoscaler(_AutoscalerWithHysteresis):
                 _generate_scale_up_decisions(num_replicas_to_scale_up, None))
 
         # Case 2: when latest_nonterminal_replicas is more
-        # than self.target_num_replicas, we scale down new replicas.
+        # than target_num_replicas, we scale down new replicas.
         replicas_to_scale_down = []
-        if len(latest_nonterminal_replicas) > self.target_num_replicas:
+        if len(latest_nonterminal_replicas) > target_num_replicas:
             num_replicas_to_scale_down = (len(latest_nonterminal_replicas) -
-                                          self.target_num_replicas)
+                                          target_num_replicas)
             replicas_to_scale_down = (
                 _select_nonterminal_replicas_to_scale_down(
                     num_replicas_to_scale_down, latest_nonterminal_replicas))
@@ -633,7 +644,7 @@ class FallbackRequestRateAutoscaler(RequestRateAutoscaler):
         all_replica_ids_to_scale_down: List[int] = []
 
         # Decide how many spot instances to launch.
-        num_spot_to_provision = (self.target_num_replicas -
+        num_spot_to_provision = (self.get_final_target_num_replicas() -
                                  self.base_ondemand_fallback_replicas)
         if num_nonterminal_spot < num_spot_to_provision:
             # Not enough spot instances, scale up.
@@ -668,6 +679,10 @@ class FallbackRequestRateAutoscaler(RequestRateAutoscaler):
             num_ondemand_to_provision += (num_spot_to_provision -
                                           num_ready_spot)
 
+        # Make sure we don't launch on-demand fallback for
+        # overprovisioned replicas.
+        num_ondemand_to_provision = min(num_ondemand_to_provision,
+                                        self.target_num_replicas)
         if num_ondemand_to_provision > num_nonterminal_ondemand:
             num_ondemand_to_scale_up = (num_ondemand_to_provision -
                                         num_nonterminal_ondemand)

sky/serve/replica_managers.py

@@ -27,6 +27,7 @@ from sky.serve import constants as serve_constants
 from sky.serve import serve_state
 from sky.serve import serve_utils
 from sky.serve import service
+from sky.serve import spot_placer
 from sky.skylet import constants
 from sky.skylet import job_lib
 from sky.usage import usage_lib
@@ -60,6 +61,7 @@ def launch_cluster(replica_id: int,
                    task_yaml_path: str,
                    cluster_name: str,
                    resources_override: Optional[Dict[str, Any]] = None,
+                   retry_until_up: bool = True,
                    max_retry: int = 3) -> None:
     """Launch a sky serve replica cluster.
 
@@ -71,6 +73,10 @@ def launch_cluster(replica_id: int,
             or some error happened before provisioning and will happen again
             if retry.
     """
+    if resources_override is not None:
+        logger.info(f'Scaling up replica (id: {replica_id}) cluster '
+                    f'{cluster_name} with resources override: '
+                    f'{resources_override}')
     try:
         config = common_utils.read_yaml(os.path.expanduser(task_yaml_path))
         task = sky.Task.from_yaml_config(config)
@@ -98,7 +104,7 @@ def launch_cluster(replica_id: int,
             usage_lib.messages.usage.set_internal()
             execution.launch(task,
                              cluster_name,
-                             retry_until_up=True,
+                             retry_until_up=retry_until_up,
                              _is_launched_by_sky_serve_controller=True)
             logger.info(f'Replica cluster {cluster_name} launched.')
         except (exceptions.InvalidClusterNameError,
@@ -246,6 +252,10 @@ class ReplicaStatusProperty:
     preempted: bool = False
     # Whether the replica is purged.
     purged: bool = False
+    # Whether the replica failed to launch due to spot availability.
+    # This is only possible when spot placer is enabled, so the retry until up
+    # is set to True and it can fail immediately due to spot availability.
+    failed_spot_availability: bool = False
 
     def remove_terminated_replica(self) -> bool:
         """Whether to remove the replica record from the replica table.
@@ -385,10 +395,11 @@ class ReplicaStatusProperty:
 class ReplicaInfo:
     """Replica info for each replica."""
 
-    _VERSION = 0
+    _VERSION = 1
 
     def __init__(self, replica_id: int, cluster_name: str, replica_port: str,
-                 is_spot: bool, version: int) -> None:
+                 is_spot: bool, location: Optional[spot_placer.Location],
+                 version: int) -> None:
         self._version = self._VERSION
         self.replica_id: int = replica_id
         self.cluster_name: str = cluster_name
@@ -398,6 +409,11 @@ class ReplicaInfo:
         self.consecutive_failure_times: List[float] = []
         self.status_property: ReplicaStatusProperty = ReplicaStatusProperty()
         self.is_spot: bool = is_spot
+        self.location: Optional[Dict[str, Optional[str]]] = (
+            location.to_pickleable() if location is not None else None)
+
+    def get_spot_location(self) -> Optional[spot_placer.Location]:
+        return spot_placer.Location.from_pickleable(self.location)
 
     def handle(
         self,
@@ -483,6 +499,7 @@ class ReplicaInfo:
                 f'version={self.version}, '
                 f'replica_port={self.replica_port}, '
                 f'is_spot={self.is_spot}, '
+                f'location={self.location}, '
                 f'status={self.status}, '
                 f'launched_at={info_dict["launched_at"]}{handle_str})')
         return info
@@ -557,6 +574,9 @@ class ReplicaInfo:
             # Treated similar to on-demand instances.
             self.is_spot = False
 
+        if version < 1:
+            self.location = None
+
         self.__dict__.update(state)
 
 
@@ -620,6 +640,9 @@ class SkyPilotReplicaManager(ReplicaManager):
                  task_yaml_path: str) -> None:
         super().__init__(service_name, spec)
         self._task_yaml_path = task_yaml_path
+        task = sky.Task.from_yaml(task_yaml_path)
+        self._spot_placer: Optional[spot_placer.SpotPlacer] = (
+            spot_placer.SpotPlacer.from_task(spec, task))
         # TODO(tian): Store launch/down pid in the replica table, to make the
         # manager more persistent. Current blocker is that we need to manually
         # poll the Process (by join or is_launch), otherwise, it will never
@@ -639,6 +662,9 @@ class SkyPilotReplicaManager(ReplicaManager):
     # Replica management functions #
     ################################
 
+    # Adding lock here to make sure spot placer's current locations are
+    # consistent with the replicas' status.
+    @with_lock
     def _launch_replica(
         self,
         replica_id: int,
@@ -653,19 +679,41 @@ class SkyPilotReplicaManager(ReplicaManager):
             self._service_name, replica_id)
         log_file_name = serve_utils.generate_replica_launch_log_file_name(
             self._service_name, replica_id)
+        use_spot = _should_use_spot(self._task_yaml_path, resources_override)
+        retry_until_up = True
+        location = None
+        if use_spot and self._spot_placer is not None:
+            # For spot placer, we don't retry until up so any launch failed
+            # due to availability issue will be handled by the placer.
+            retry_until_up = False
+            # TODO(tian): Currently, the resources_override can only be
+            # `use_spot=True/False`, which will not cause any conflict with
+            # spot placer's cloud, region & zone. When we add more resources
+            # to the resources_override, we need to make sure they won't
+            # conflict with the spot placer's selection.
+            if resources_override is None:
+                resources_override = {}
+            current_spot_locations: List[spot_placer.Location] = []
+            for info in serve_state.get_replica_infos(self._service_name):
+                if info.is_spot:
+                    spot_location = info.get_spot_location()
+                    if spot_location is not None:
+                        current_spot_locations.append(spot_location)
+            location = self._spot_placer.select_next_location(
+                current_spot_locations)
+            resources_override.update(location.to_dict())
         p = multiprocessing.Process(
             target=ux_utils.RedirectOutputForProcess(
                 launch_cluster,
                 log_file_name,
             ).run,
             args=(replica_id, self._task_yaml_path, cluster_name,
-                  resources_override),
+                  resources_override, retry_until_up),
         )
         replica_port = _get_resources_ports(self._task_yaml_path)
-        use_spot = _should_use_spot(self._task_yaml_path, resources_override)
 
         info = ReplicaInfo(replica_id, cluster_name, replica_port, use_spot,
-                           self.latest_version)
+                           location, self.latest_version)
         serve_state.add_or_update_replica(self._service_name, replica_id, info)
         # Don't start right now; we will start it later in _refresh_process_pool
         # to avoid too many sky.launch running at the same time.
@@ -814,6 +862,10 @@ class SkyPilotReplicaManager(ReplicaManager):
         logger.info(
             f'Replica {info.replica_id} is preempted{cluster_status_str}.')
         info.status_property.preempted = True
+        if self._spot_placer is not None:
+            spot_location = info.get_spot_location()
+            assert spot_location is not None
+            self._spot_placer.set_preemptive(spot_location)
         serve_state.add_or_update_replica(self._service_name, info.replica_id,
                                           info)
         self._terminate_replica(info.replica_id,
@@ -868,6 +920,23 @@ class SkyPilotReplicaManager(ReplicaManager):
                     else:
                         info.status_property.sky_launch_status = (
                             ProcessStatus.SUCCEEDED)
+                    if self._spot_placer is not None and info.is_spot:
+                        # TODO(tian): Currently, we set the location to
+                        # preemptive if the launch process failed. This is
+                        # because if the error is not related to the
+                        # availability of the location, then all locations
+                        # should failed for same reason. So it does not matter
+                        # which location is preemptive or not, instead, all
+                        # locations would fail. We should implement a log parser
+                        # to detect if the error is actually related to the
+                        # availability of the location later.
+                        location = info.get_spot_location()
+                        assert location is not None
+                        if p.exitcode != 0:
+                            self._spot_placer.set_preemptive(location)
+                            info.status_property.failed_spot_availability = True
+                        else:
+                            self._spot_placer.set_active(location)
                 serve_state.add_or_update_replica(self._service_name,
                                                   replica_id, info)
                 if error_in_sky_launch:
@@ -918,6 +987,8 @@ class SkyPilotReplicaManager(ReplicaManager):
                     removal_reason = 'for version outdated'
                 elif info.status_property.purged:
                     removal_reason = 'for purge'
+                elif info.status_property.failed_spot_availability:
+                    removal_reason = 'for spot availability failure'
                 else:
                     logger.info(f'Termination of replica {replica_id} '
                                 'finished. Replica info is kept since some '

sky/serve/serve_utils.py

@@ -25,6 +25,7 @@ from sky import global_user_state
 from sky.adaptors import common as adaptors_common
 from sky.serve import constants
 from sky.serve import serve_state
+from sky.serve import spot_placer
 from sky.skylet import constants as skylet_constants
 from sky.skylet import job_lib
 from sky.utils import annotations
@@ -40,6 +41,7 @@ if typing.TYPE_CHECKING:
     import psutil
     import requests
 
+    import sky
     from sky.serve import replica_managers
 else:
     psutil = adaptors_common.LazyImport('psutil')
@@ -210,6 +212,84 @@ class RequestTimestamp(RequestsAggregator):
         return f'RequestTimestamp(timestamps={self.timestamps})'
 
 
+def validate_service_task(task: 'sky.Task') -> None:
+    """Validate the task for Sky Serve.
+
+    Args:
+        task: sky.Task to validate
+
+    Raises:
+        ValueError: if the arguments are invalid.
+        RuntimeError: if the task.serve is not found.
+    """
+    spot_resources: List['sky.Resources'] = [
+        resource for resource in task.resources if resource.use_spot
+    ]
+    # TODO(MaoZiming): Allow mixed on-demand and spot specification in resources
+    # On-demand fallback should go to the resources specified as on-demand.
+    if len(spot_resources) not in [0, len(task.resources)]:
+        with ux_utils.print_exception_no_traceback():
+            raise ValueError(
+                'Resources must either all use spot or none use spot. '
+                'To use on-demand and spot instances together, '
+                'use `dynamic_ondemand_fallback` or set '
+                'base_ondemand_fallback_replicas.')
+
+    if task.service is None:
+        with ux_utils.print_exception_no_traceback():
+            raise RuntimeError('Service section not found.')
+
+    policy_description = ('on-demand'
+                          if task.service.dynamic_ondemand_fallback else 'spot')
+    for resource in list(task.resources):
+        if resource.job_recovery is not None:
+            with ux_utils.print_exception_no_traceback():
+                raise ValueError('job_recovery is disabled for SkyServe. '
+                                 'SkyServe will replenish preempted spot '
+                                 f'with {policy_description} instances.')
+
+    # Try to create a spot placer from the task yaml. Check if the task yaml
+    # is valid for spot placer.
+    spot_placer.SpotPlacer.from_task(task.service, task)
+
+    replica_ingress_port: Optional[int] = int(
+        task.service.ports) if (task.service.ports is not None) else None
+    for requested_resources in task.resources:
+        if (task.service.use_ondemand_fallback and
+                not requested_resources.use_spot):
+            with ux_utils.print_exception_no_traceback():
+                raise ValueError(
+                    '`use_ondemand_fallback` is only supported '
+                    'for spot resources. Please explicitly specify '
+                    '`use_spot: true` in resources for on-demand fallback.')
+        if (task.service.spot_placer is not None and
+                not requested_resources.use_spot):
+            with ux_utils.print_exception_no_traceback():
+                raise ValueError(
+                    '`spot_placer` is only supported for spot resources. '
+                    'Please explicitly specify `use_spot: true` in resources.')
+        if task.service.ports is None:
+            requested_ports = list(
+                resources_utils.port_ranges_to_set(requested_resources.ports))
+            if len(requested_ports) != 1:
+                with ux_utils.print_exception_no_traceback():
+                    raise ValueError(
+                        'To open multiple ports on the replica, please set the '
+                        '`service.ports` field to specify a main service port. '
+                        'Must only specify one port in resources otherwise. '
+                        'Each replica will use the port specified as '
+                        'application ingress port.')
+            service_port = requested_ports[0]
+            if replica_ingress_port is None:
+                replica_ingress_port = service_port
+            elif service_port != replica_ingress_port:
+                with ux_utils.print_exception_no_traceback():
+                    raise ValueError(
+                        f'Got multiple ports: {service_port} and '
+                        f'{replica_ingress_port} in different resources. '
+                        'Please specify the same port instead.')
+
+
 def generate_service_name():
     return f'sky-service-{uuid.uuid4().hex[:4]}'