delete template and registry after termination

sky/backends/cloud_vm_ray_backend.py

@@ -1570,6 +1570,17 @@ def _retry_zones(
                     config_dict['provision_record'] = provision_record
                     config_dict['resources_vars'] = resources_vars
                     config_dict['handle'] = handle
+                    if provision_record.ephemeral_resources:
+                        # Some ephemeral resources are created during the launch
+                        # process. Add them to the provider config so that they
+                        # can be cleaned up later.
+                        original_config_content = common_utils.read_yaml(
+                            cluster_config_file)
+                        original_config_content['provider'][
+                            'ephemeral_resources'] = (
+                                provision_record.ephemeral_resources)
+                        common_utils.dump_yaml(cluster_config_file,
+                                               original_config_content)
                     return config_dict
                 except provision_common.StopFailoverError:
                     with ux_utils.print_exception_no_traceback():

sky/provision/common.py

@@ -78,6 +78,7 @@ class ProvisionRecord:
     resumed_instance_ids: List[InstanceId]
     # The IDs of all just created instances.
     created_instance_ids: List[InstanceId]
+    ephemeral_resources: List[Any] = dataclasses.field(default_factory=list)
 
     def is_instance_just_booted(self, instance_id: InstanceId) -> bool:
         """Whether or not the instance is just booted.

sky/provision/runpod/instance.py

@@ -79,10 +79,11 @@ def run_instances(region: str, cluster_name_on_cloud: str,
                                       created_instance_ids=[])
 
     created_instance_ids = []
+    ephemeral_resources = []
     for _ in range(to_start_count):
         node_type = 'head' if head_instance_id is None else 'worker'
         try:
-            instance_id = utils.launch(
+            instance_id, ers = utils.launch(
                 cluster_name=cluster_name_on_cloud,
                 node_type=node_type,
                 instance_type=config.node_config['InstanceType'],
@@ -96,6 +97,9 @@ def run_instances(region: str, cluster_name_on_cloud: str,
                 docker_login_config=config.provider_config.get(
                     'docker_login_config'),
             )
+            for er in ers:
+                if er is not None:
+                    ephemeral_resources.append(er)
         except Exception as e:  # pylint: disable=broad-except
             logger.warning(f'run_instances error: {e}')
             raise
@@ -124,7 +128,8 @@ def run_instances(region: str, cluster_name_on_cloud: str,
                                   zone=None,
                                   head_instance_id=head_instance_id,
                                   resumed_instance_ids=[],
-                                  created_instance_ids=created_instance_ids)
+                                  created_instance_ids=created_instance_ids,
+                                  ephemeral_resources=ephemeral_resources)
 
 
 def wait_instances(region: str, cluster_name_on_cloud: str,
@@ -146,7 +151,8 @@ def terminate_instances(
     worker_only: bool = False,
 ) -> None:
     """See sky/provision/__init__.py"""
-    del provider_config  # unused
+    assert provider_config is not None, (cluster_name_on_cloud, provider_config)
+    ephemeral_resources = provider_config.get('ephemeral_resources', [])
     instances = _filter_instances(cluster_name_on_cloud, None)
     for inst_id, inst in instances.items():
         logger.debug(f'Terminating instance {inst_id}: {inst}')
@@ -160,6 +166,12 @@ def terminate_instances(
                     f'Failed to terminate instance {inst_id}: '
                     f'{common_utils.format_exception(e, use_bracket=False)}'
                 ) from e
+    if ephemeral_resources:
+        # See sky/provision/runpod/utils.py::launch for details
+        assert len(ephemeral_resources) == 2, ephemeral_resources
+        template_name, registry_auth_id = ephemeral_resources
+        utils.delete_pod_template(template_name)
+        utils.delete_register_auth(registry_auth_id)
 
 
 def get_cluster_info(

sky/provision/runpod/utils.py

@@ -3,7 +3,6 @@
 import base64
 import time
 from typing import Any, Dict, List, Optional, Tuple
-import uuid
 
 from sky import sky_logging
 from sky.adaptors import runpod
@@ -102,19 +101,44 @@ def list_instances() -> Dict[str, Dict[str, Any]]:
     return instance_dict
 
 
+def delete_pod_template(template_name: str) -> None:
+    """Deletes a pod template."""
+    try:
+        runpod.runpod.api.graphql.run_graphql_query(
+            f'mutation {{deleteTemplate(templateName: "{template_name}")}}')
+    except runpod.runpod.error.QueryError as e:
+        logger.warning(f'Failed to delete template {template_name}: {e}'
+                       'Please delete it manually.')
+
+
+def delete_register_auth(registry_auth_id: str) -> None:
+    """Deletes a registry auth."""
+    try:
+        runpod.runpod.delete_container_registry_auth(registry_auth_id)
+    except runpod.runpod.error.QueryError as e:
+        logger.warning(f'Failed to delete registry auth {registry_auth_id}: {e}'
+                       'Please delete it manually.')
+
+
 def _create_template_for_docker_login(
-        cluster_name: str, image_name: str,
-        docker_login_config: Optional[Dict[str,
-                                           str]]) -> Tuple[str, Optional[str]]:
+    cluster_name: str,
+    image_name: str,
+    docker_login_config: Optional[Dict[str, str]],
+) -> Tuple[str, Optional[str], Optional[str], Optional[str]]:
+    """Creates a template for the given image with the docker login config.
+
+    Returns:
+        formatted_image_name: The formatted image name.
+        # following fields are None for no docker login config.
+        template_id: The template ID.
+        template_name: The template name.
+        registry_auth_id: The registry auth ID.
+    """
     if docker_login_config is None:
-        return image_name, None
-    # We add a uuid here to avoid the name conflict for terminating and
-    # launching with the same cluster name. Please see the comments below
-    # for reason we cannot cleanup the old resources.
-    name = f'{cluster_name}-{str(uuid.uuid4())[:4]}'
+        return image_name, None, None, None
     login_config = docker_utils.DockerLoginConfig(**docker_login_config)
-    container_registry_auth_name = f'{name}-registry-auth'
-    container_template_name = f'{name}-docker-login-template'
+    container_registry_auth_name = f'{cluster_name}-registry-auth'
+    container_template_name = f'{cluster_name}-docker-login-template'
     # The `name` argument is only for display purpose and the registry server
     # will be splitted from the docker image name (Tested with AWS ECR).
     # Here we only need the username and password to create the registry auth.
@@ -123,15 +147,6 @@ def _create_template_for_docker_login(
     # we create it. So we use a separate auth and template per cluster. This
     # also assumes that every cluster has only one node, so no extra worker
     # nodes will reuse the same auth and template name.
-    # TODO(tian): RunPod python API does not provide a way to delete the
-    # template. So we skip the deletion of template for now. We should
-    # implement this once they provide the API.
-    # TODO(tian): We also skipped the deletion of the auth for now, as the
-    # RunPod python API does not provide a way to delete the auth with the
-    # name (nor to get the id by the name), which requires we store the id
-    # at creation somewhere, and returning this value to outer caller will
-    # increase the call chain complexity. We should implement this once they
-    # provide the API.
     create_auth_resp = runpod.runpod.create_container_registry_auth(
         name=container_registry_auth_name,
         username=login_config.username,
@@ -143,17 +158,23 @@ def _create_template_for_docker_login(
         image_name=None,
         registry_auth_id=registry_auth_id,
     )
-    return login_config.format_image(image_name), create_template_resp['id']
+    return (login_config.format_image(image_name), create_template_resp['id'],
+            container_template_name, registry_auth_id)
 
 
-def launch(cluster_name: str, node_type: str, instance_type: str, region: str,
-           disk_size: int, image_name: str, ports: Optional[List[int]],
-           public_key: str, preemptible: Optional[bool], bid_per_gpu: float,
-           docker_login_config: Optional[Dict[str, str]]) -> str:
+def launch(
+        cluster_name: str, node_type: str, instance_type: str, region: str,
+        disk_size: int, image_name: str, ports: Optional[List[int]],
+        public_key: str, preemptible: Optional[bool], bid_per_gpu: float,
+        docker_login_config: Optional[Dict[str, str]]) -> Tuple[str, List[Any]]:
     """Launches an instance with the given parameters.
 
     Converts the instance_type to the RunPod GPU name, finds the specs for the
     GPU, and launches the instance.
+
+    Returns:
+        instance_id: The instance ID.
+        ephemeral_resources: A list of ephemeral resources.
     """
     name = f'{cluster_name}-{node_type}'
     gpu_type = GPU_NAME_MAP[instance_type.split('_')[1]]
@@ -199,8 +220,9 @@ def launch(cluster_name: str, node_type: str, instance_type: str, region: str,
                  f'{constants.SKY_REMOTE_RAY_DASHBOARD_PORT}/http,'
                  f'{constants.SKY_REMOTE_RAY_PORT}/http')
 
-    image_name_formatted, template_id = (_create_template_for_docker_login(
-        cluster_name, image_name, docker_login_config))
+    image_name_formatted, template_id, template_name, registry_auth_id = (
+        _create_template_for_docker_login(cluster_name, image_name,
+                                          docker_login_config))
 
     params = {
         'name': name,
@@ -226,7 +248,7 @@ def launch(cluster_name: str, node_type: str, instance_type: str, region: str,
             **params,
         )
 
-    return new_instance['id']
+    return new_instance['id'], [template_name, registry_auth_id]
 
 
 def remove(instance_id: str) -> None: