skypilot-nightly 1.0.0.dev20251005__py3-none-any.whl → 1.0.0.dev20251008__py3-none-any.whl

sky/execution.py

@@ -112,7 +112,6 @@ def _execute(
     stages: Optional[List[Stage]] = None,
     cluster_name: Optional[str] = None,
     detach_setup: bool = False,
-    detach_run: bool = False,
     idle_minutes_to_autostop: Optional[int] = None,
     no_setup: bool = False,
     clone_disk_from: Optional[str] = None,
@@ -157,8 +156,6 @@ def _execute(
         job itself. You can safely ctrl-c to detach from logging, and it will
         not interrupt the setup process. To see the logs again after detaching,
         use `sky logs`. To cancel setup, cancel the job via `sky cancel`.
-      detach_run: If True, as soon as a job is submitted, return from this
-        function and do not stream execution logs.
       idle_minutes_to_autostop: int; if provided, the cluster will be set to
         autostop after this many minutes of idleness.
       no_setup: bool; whether to skip setup commands or not when (re-)launching.
@@ -217,7 +214,6 @@ def _execute(
             stages=stages,
             cluster_name=cluster_name,
             detach_setup=detach_setup,
-            detach_run=detach_run,
             no_setup=no_setup,
             clone_disk_from=clone_disk_from,
             skip_unnecessary_provisioning=skip_unnecessary_provisioning,
@@ -239,7 +235,6 @@ def _execute_dag(
     stages: Optional[List[Stage]],
     cluster_name: Optional[str],
     detach_setup: bool,
-    detach_run: bool,
     no_setup: bool,
     clone_disk_from: Optional[str],
     skip_unnecessary_provisioning: bool,
@@ -507,10 +502,7 @@ def _execute_dag(
         if Stage.EXEC in stages:
             try:
                 global_user_state.update_last_use(handle.get_cluster_name())
-                job_id = backend.execute(handle,
-                                         task,
-                                         detach_run,
-                                         dryrun=dryrun)
+                job_id = backend.execute(handle, task, dryrun=dryrun)
             finally:
                 # Enables post_execute() to be run after KeyboardInterrupt.
                 backend.post_execute(handle, down)
@@ -707,7 +699,6 @@ def launch(
         stages=stages,
         cluster_name=cluster_name,
         detach_setup=detach_setup,
-        detach_run=True,
         idle_minutes_to_autostop=idle_minutes_to_autostop,
         no_setup=no_setup,
         clone_disk_from=clone_disk_from,
@@ -802,6 +793,5 @@ def exec(  # pylint: disable=redefined-builtin
             Stage.EXEC,
         ],
         cluster_name=cluster_name,
-        detach_run=True,
         job_logger=job_logger,
     )

sky/global_user_state.py

@@ -2495,11 +2495,22 @@ def _set_cluster_yaml_from_file(cluster_yaml_path: str,
     # on the local file system and migrate it to the database.
     # TODO(syang): remove this check once we have a way to migrate the
     # cluster from file to database. Remove on v0.12.0.
-    if cluster_yaml_path is not None and os.path.exists(cluster_yaml_path):
-        with open(cluster_yaml_path, 'r', encoding='utf-8') as f:
-            yaml_str = f.read()
-        set_cluster_yaml(cluster_name, yaml_str)
-        return yaml_str
+    if cluster_yaml_path is not None:
+        # First try the exact path
+        path_to_read = None
+        if os.path.exists(cluster_yaml_path):
+            path_to_read = cluster_yaml_path
+        # Fallback: try with .debug suffix (when debug logging was enabled)
+        # Debug logging causes YAML files to be saved with .debug suffix
+        # but the path stored in the handle doesn't include it
+        debug_path = cluster_yaml_path + '.debug'
+        if os.path.exists(debug_path):
+            path_to_read = debug_path
+        if path_to_read is not None:
+            with open(path_to_read, 'r', encoding='utf-8') as f:
+                yaml_str = f.read()
+            set_cluster_yaml(cluster_name, yaml_str)
+            return yaml_str
     return None
 
 

sky/jobs/constants.py

@@ -15,16 +15,10 @@ JOB_CONTROLLER_INDICATOR_FILE = '~/.sky/is_jobs_controller'
 CONSOLIDATED_SIGNAL_PATH = os.path.expanduser('~/.sky/signals/')
 SIGNAL_FILE_PREFIX = '/tmp/sky_jobs_controller_signal_{}'
 # Resources as a dict for the jobs controller.
-# Use smaller CPU instance type for jobs controller, but with more memory, i.e.
-# r6i.xlarge (4vCPUs, 32 GB) for AWS, Standard_E4s_v5 (4vCPUs, 32 GB) for Azure,
-# and n2-highmem-4 (4 vCPUs, 32 GB) for GCP, etc.
-# Concurrently limits are set based on profiling. 4x num vCPUs is the launch
-# parallelism limit, and memory / 350MB is the limit to concurrently running
-# jobs. See _get_launch_parallelism and _get_job_parallelism in scheduler.py.
 # We use 50 GB disk size to reduce the cost.
 CONTROLLER_RESOURCES: Dict[str, Union[str, int]] = {
     'cpus': '4+',
-    'memory': '8x',
+    'memory': '4x',
     'disk_size': 50
 }
 

sky/jobs/controller.py

@@ -1144,7 +1144,15 @@ class Controller:
                 await asyncio.sleep(30)
                 continue
 
-            if len(running_tasks) >= scheduler.JOBS_PER_WORKER:
+            # Normally, 200 jobs can run on each controller. But if we have a
+            # ton of controllers, we need to limit the number of jobs that can
+            # run on each controller, to achieve a total of 2000 jobs across all
+            # controllers.
+            max_jobs = min(scheduler.MAX_JOBS_PER_WORKER,
+                           (scheduler.MAX_TOTAL_RUNNING_JOBS //
+                            scheduler.get_number_of_controllers()))
+
+            if len(running_tasks) >= max_jobs:
                 await asyncio.sleep(60)
                 continue
 

sky/jobs/scheduler.py

@@ -63,7 +63,9 @@ from sky.jobs import state
 from sky.jobs import utils as managed_job_utils
 from sky.server import config as server_config
 from sky.skylet import constants
+from sky.utils import annotations
 from sky.utils import common_utils
+from sky.utils import controller_utils
 from sky.utils import subprocess_utils
 
 if typing.TYPE_CHECKING:
@@ -91,20 +93,29 @@ JOB_MEMORY_MB = 400
 LAUNCHES_PER_WORKER = 8
 # this can probably be increased to around 300-400 but keeping it lower to just
 # to be safe
-JOBS_PER_WORKER = 200
-
-# keep 1GB reserved after the controllers
-MAXIMUM_CONTROLLER_RESERVED_MEMORY_MB = 2048
-
-CURRENT_HASH = os.path.expanduser('~/.sky/wheels/current_sky_wheel_hash')
-
+MAX_JOBS_PER_WORKER = 200
+# Maximum number of controllers that can be running. Hard to handle more than
+# 512 launches at once.
+MAX_CONTROLLERS = 512 // LAUNCHES_PER_WORKER
+# Limit the number of jobs that can be running at once on the entire jobs
+# controller cluster. It's hard to handle cancellation of more than 2000 jobs at
+# once.
+# TODO(cooperc): Once we eliminate static bottlenecks (e.g. sqlite), remove this
+# hardcoded max limit.
+MAX_TOTAL_RUNNING_JOBS = 2000
 # Maximum values for above constants. There will start to be lagging issues
 # at these numbers already.
 # JOB_MEMORY_MB = 200
 # LAUNCHES_PER_WORKER = 16
 # JOBS_PER_WORKER = 400
 
+# keep 2GB reserved after the controllers
+MAXIMUM_CONTROLLER_RESERVED_MEMORY_MB = 2048
+
+CURRENT_HASH = os.path.expanduser('~/.sky/wheels/current_sky_wheel_hash')
+
 
+@annotations.lru_cache(scope='global')
 def get_number_of_controllers() -> int:
     """Returns the number of controllers that should be running.
 
@@ -123,7 +134,7 @@ def get_number_of_controllers() -> int:
     consolidation_mode = skypilot_config.get_nested(
         ('jobs', 'controller', 'consolidation_mode'), default_value=False)
 
-    total_memory_mb = common_utils.get_mem_size_gb() * 1024
+    total_memory_mb = controller_utils.get_controller_mem_size_gb() * 1024
     if consolidation_mode:
         config = server_config.compute_server_config(deploy=True, quiet=True)
 
@@ -136,13 +147,16 @@ def get_number_of_controllers() -> int:
                     config.short_worker_config.burstable_parallelism) * \
             server_config.SHORT_WORKER_MEM_GB * 1024
 
-        return max(1, int((total_memory_mb - used) // JOB_MEMORY_MB))
+        return min(MAX_CONTROLLERS,
+                   max(1, int((total_memory_mb - used) // JOB_MEMORY_MB)))
     else:
-        return max(
-            1,
-            int((total_memory_mb - MAXIMUM_CONTROLLER_RESERVED_MEMORY_MB) /
-                ((LAUNCHES_PER_WORKER * server_config.LONG_WORKER_MEM_GB) * 1024
-                 + JOB_MEMORY_MB)))
+        return min(
+            MAX_CONTROLLERS,
+            max(
+                1,
+                int((total_memory_mb - MAXIMUM_CONTROLLER_RESERVED_MEMORY_MB) /
+                    ((LAUNCHES_PER_WORKER * server_config.LONG_WORKER_MEM_GB) *
+                     1024 + JOB_MEMORY_MB))))
 
 
 def start_controller() -> None:
@@ -280,7 +294,8 @@ def submit_job(job_id: int, dag_yaml_path: str, original_user_yaml_path: str,
                                 common_utils.get_user_hash(), priority)
     if state.get_ha_recovery_script(job_id) is None:
         # the run command is just the command that called scheduler
-        run = (f'{sys.executable} -m sky.jobs.scheduler {dag_yaml_path} '
+        run = (f'source {env_file_path} && '
+               f'{sys.executable} -m sky.jobs.scheduler {dag_yaml_path} '
                f'--job-id {job_id} --env-file {env_file_path} '
                f'--user-yaml-path {original_user_yaml_path} '
                f'--priority {priority}')

sky/jobs/server/core.py

@@ -407,9 +407,12 @@ def launch(
             job_identity = ''
             if job_rank is not None:
                 job_identity = f' (rank: {job_rank})'
-            logger.info(f'{colorama.Fore.YELLOW}'
-                        f'Launching managed job {dag.name!r}{job_identity} '
-                        f'from jobs controller...{colorama.Style.RESET_ALL}')
+            job_controller_postfix = (' from jobs controller' if
+                                      consolidation_mode_job_id is None else '')
+            logger.info(
+                f'{colorama.Fore.YELLOW}'
+                f'Launching managed job {dag.name!r}{job_identity}'
+                f'{job_controller_postfix}...{colorama.Style.RESET_ALL}')
 
             # Launch with the api server's user hash, so that sky status does
             # not show the owner of the controller as whatever user launched
@@ -456,6 +459,8 @@ def launch(
                     managed_job_state.set_ha_recovery_script(
                         consolidation_mode_job_id, run_script)
                     backend.run_on_head(local_handle, run_script)
+                    ux_utils.starting_message(
+                        f'Job submitted, ID: {consolidation_mode_job_id}')
                     return consolidation_mode_job_id, local_handle
 
     if pool is None:

sky/jobs/utils.py

@@ -11,6 +11,7 @@ import enum
 import logging
 import os
 import pathlib
+import re
 import shlex
 import textwrap
 import time
@@ -299,8 +300,10 @@ async def get_job_status(
                 job_logger.info(f'Job status: {status}')
             job_logger.info('=' * 34)
             return status
-        except (exceptions.CommandError, grpc.RpcError,
-                grpc.FutureTimeoutError) as e:
+        except (exceptions.CommandError, grpc.RpcError, grpc.FutureTimeoutError,
+                ValueError, TypeError) as e:
+            # Note: Each of these exceptions has some additional conditions to
+            # limit how we handle it and whether or not we catch it.
             # Retry on k8s transient network errors. This is useful when using
             # coreweave which may have transient network issue sometimes.
             is_transient_error = False
@@ -319,6 +322,31 @@ async def get_job_status(
                     is_transient_error = True
             elif isinstance(e, grpc.FutureTimeoutError):
                 detailed_reason = 'Timeout'
+            # TODO(cooperc): Gracefully handle these exceptions in the backend.
+            elif isinstance(e, ValueError):
+                # If the cluster yaml is deleted in the middle of getting the
+                # SSH credentials, we could see this. See
+                # sky/global_user_state.py get_cluster_yaml_dict.
+                if re.search(r'Cluster yaml .* not found', str(e)):
+                    detailed_reason = 'Cluster yaml was deleted'
+                else:
+                    raise
+            elif isinstance(e, TypeError):
+                # We will grab the SSH credentials from the cluster yaml, but if
+                # handle.cluster_yaml is None, we will just return an empty dict
+                # for the credentials. See
+                # backend_utils.ssh_credential_from_yaml. Then, the credentials
+                # are passed as kwargs to SSHCommandRunner.__init__ - see
+                # cloud_vm_ray_backend.get_command_runners. So we can hit this
+                # TypeError if the cluster yaml is removed from the handle right
+                # when we pull it before the cluster is fully deleted.
+                error_msg_to_check = (
+                    'SSHCommandRunner.__init__() missing 2 required positional '
+                    'arguments: \'ssh_user\' and \'ssh_private_key\'')
+                if str(e) == error_msg_to_check:
+                    detailed_reason = 'SSH credentials were already cleaned up'
+                else:
+                    raise
             if is_transient_error:
                 logger.info('Failed to connect to the cluster. Retrying '
                             f'({i + 1}/{_JOB_STATUS_FETCH_MAX_RETRIES})...')

sky/metrics/utils.py

@@ -11,7 +11,9 @@ from typing import List, Optional, Tuple
 import httpx
 import prometheus_client as prom
 
+from sky import sky_logging
 from sky.skylet import constants
+from sky.utils import common_utils
 from sky.utils import context_utils
 
 _SELECT_TIMEOUT = 1
@@ -35,6 +37,8 @@ _MEM_BUCKETS = [
     float('inf'),
 ]
 
+logger = sky_logging.init_logger(__name__)
+
 # Whether the metrics are enabled, cannot be changed at runtime.
 METRICS_ENABLED = os.environ.get(constants.ENV_VAR_SERVER_METRICS_ENABLED,
                                  'false').lower() == 'true'
@@ -188,53 +192,61 @@ def start_svc_port_forward(context: str, namespace: str, service: str,
     if 'KUBECONFIG' not in env:
         env['KUBECONFIG'] = os.path.expanduser('~/.kube/config')
 
-    # start the port forward process
-    port_forward_process = subprocess.Popen(cmd,
-                                            stdout=subprocess.PIPE,
-                                            stderr=subprocess.STDOUT,
-                                            text=True,
-                                            env=env)
-
+    port_forward_process = None
+    port_forward_exit = False
     local_port = None
-    start_time = time.time()
-
-    buffer = ''
-    # wait for the port forward to start and extract the local port
-    while time.time() - start_time < start_port_forward_timeout:
-        if port_forward_process.poll() is not None:
-            # port forward process has terminated
-            if port_forward_process.returncode != 0:
-                raise RuntimeError(
-                    f'Port forward failed for service {service} in namespace '
-                    f'{namespace} on context {context}')
-            break
-
-        # read output line by line to find the local port
-        if port_forward_process.stdout:
-            # Wait up to 1s for data to be available without blocking
-            r, _, _ = select.select([port_forward_process.stdout], [], [],
-                                    _SELECT_TIMEOUT)
-            if r:
-                # Read available bytes from the FD without blocking
-                fd = port_forward_process.stdout.fileno()
-                raw = os.read(fd, _SELECT_BUFFER_SIZE)
-                chunk = raw.decode(errors='ignore')
-                buffer += chunk
-                match = re.search(r'Forwarding from 127\.0\.0\.1:(\d+)', buffer)
-                if match:
-                    local_port = int(match.group(1))
-                    break
-
-        # sleep for 100ms to avoid busy-waiting
-        time.sleep(0.1)
 
+    try:
+        # start the port forward process
+        port_forward_process = subprocess.Popen(cmd,
+                                                stdout=subprocess.PIPE,
+                                                stderr=subprocess.STDOUT,
+                                                text=True,
+                                                env=env)
+
+        start_time = time.time()
+
+        buffer = ''
+        # wait for the port forward to start and extract the local port
+        while time.time() - start_time < start_port_forward_timeout:
+            if port_forward_process.poll() is not None:
+                # port forward process has terminated
+                if port_forward_process.returncode != 0:
+                    port_forward_exit = True
+                break
+
+            # read output line by line to find the local port
+            if port_forward_process.stdout:
+                # Wait up to 1s for data to be available without blocking
+                r, _, _ = select.select([port_forward_process.stdout], [], [],
+                                        _SELECT_TIMEOUT)
+                if r:
+                    # Read available bytes from the FD without blocking
+                    fd = port_forward_process.stdout.fileno()
+                    raw = os.read(fd, _SELECT_BUFFER_SIZE)
+                    chunk = raw.decode(errors='ignore')
+                    buffer += chunk
+                    match = re.search(r'Forwarding from 127\.0\.0\.1:(\d+)',
+                                      buffer)
+                    if match:
+                        local_port = int(match.group(1))
+                        break
+
+            # sleep for 100ms to avoid busy-waiting
+            time.sleep(0.1)
+    except BaseException:  # pylint: disable=broad-exception-caught
+        if port_forward_process:
+            stop_svc_port_forward(port_forward_process,
+                                  timeout=terminate_port_forward_timeout)
+        raise
+    if port_forward_exit:
+        raise RuntimeError(f'Port forward failed for service {service} in '
+                           f'namespace {namespace} on context {context}')
     if local_port is None:
         try:
-            port_forward_process.terminate()
-            port_forward_process.wait(timeout=terminate_port_forward_timeout)
-        except subprocess.TimeoutExpired:
-            port_forward_process.kill()
-            port_forward_process.wait()
+            if port_forward_process:
+                stop_svc_port_forward(port_forward_process,
+                                      timeout=terminate_port_forward_timeout)
         finally:
             raise RuntimeError(
                 f'Failed to extract local port for service {service} in '
@@ -243,14 +255,15 @@ def start_svc_port_forward(context: str, namespace: str, service: str,
     return port_forward_process, local_port
 
 
-def stop_svc_port_forward(port_forward_process: subprocess.Popen) -> None:
+def stop_svc_port_forward(port_forward_process: subprocess.Popen,
+                          timeout: int = 5) -> None:
     """Stops a port forward to a service in a Kubernetes cluster.
     Args:
         port_forward_process: The subprocess.Popen process to terminate
     """
     try:
         port_forward_process.terminate()
-        port_forward_process.wait(timeout=5)
+        port_forward_process.wait(timeout=timeout)
     except subprocess.TimeoutExpired:
         port_forward_process.kill()
         port_forward_process.wait()
@@ -301,6 +314,10 @@ async def send_metrics_request_with_port_forward(
             response.raise_for_status()
             return response.text
 
+    except Exception as e:  # pylint: disable=broad-exception-caught
+        logger.error(f'Failed to send metrics request with port forward: '
+                     f'{common_utils.format_exception(e)}')
+        raise
     finally:
         # Always clean up port forward
         if port_forward_process:

sky/provision/instance_setup.py

@@ -10,6 +10,7 @@ from typing import Any, Callable, Dict, List, Optional, Tuple
 from sky import exceptions
 from sky import logs
 from sky import provision
+from sky import resources as resources_lib
 from sky import sky_logging
 from sky.provision import common
 from sky.provision import docker_utils
@@ -92,12 +93,6 @@ def _set_usage_run_id_cmd() -> str:
         f'{usage_constants.USAGE_RUN_ID_FILE}')
 
 
-def _set_skypilot_env_var_cmd() -> str:
-    """Sets the skypilot environment variables on the remote machine."""
-    env_vars = env_options.Options.all_options()
-    return '; '.join([f'export {k}={v}' for k, v in env_vars.items()])
-
-
 def _auto_retry(should_retry: Callable[[Exception], bool] = lambda _: True):
     """Decorator that retries the function if it fails.
 
@@ -482,11 +477,38 @@ def start_ray_on_worker_nodes(cluster_name: str, no_restart: bool,
 @common.log_function_start_end
 @_auto_retry()
 @timeline.event
-def start_skylet_on_head_node(cluster_name: str,
-                              cluster_info: common.ClusterInfo,
-                              ssh_credentials: Dict[str, Any]) -> None:
+def start_skylet_on_head_node(
+        cluster_name: resources_utils.ClusterName,
+        cluster_info: common.ClusterInfo, ssh_credentials: Dict[str, Any],
+        launched_resources: resources_lib.Resources) -> None:
     """Start skylet on the head node."""
-    del cluster_name
+    # Avoid circular import.
+    # pylint: disable=import-outside-toplevel
+    from sky.utils import controller_utils
+
+    def _set_skypilot_env_var_cmd() -> str:
+        """Sets the skypilot environment variables on the remote machine."""
+        env_vars = {
+            k: str(v) for (k, v) in env_options.Options.all_options().items()
+        }
+        is_controller = controller_utils.Controllers.from_name(
+            cluster_name.display_name) is not None
+        is_kubernetes = cluster_info.provider_name == 'kubernetes'
+        if is_controller and is_kubernetes:
+            # For jobs/serve controller, we pass in the CPU and memory limits
+            # when starting the skylet to handle cases where these env vars
+            # are not set on the cluster's pod spec. The skylet will read
+            # these env vars when starting (ManagedJobEvent.start()) and write
+            # it to disk.
+            resources = launched_resources.assert_launchable()
+            vcpus, mem = resources.cloud.get_vcpus_mem_from_instance_type(
+                resources.instance_type)
+            if vcpus is not None:
+                env_vars['SKYPILOT_POD_CPU_CORE_LIMIT'] = str(vcpus)
+            if mem is not None:
+                env_vars['SKYPILOT_POD_MEMORY_GB_LIMIT'] = str(mem)
+        return '; '.join([f'export {k}={v}' for k, v in env_vars.items()])
+
     runners = provision.get_command_runners(cluster_info.provider_name,
                                             cluster_info, **ssh_credentials)
     head_runner = runners[0]

sky/provision/kubernetes/utils.py

@@ -1688,7 +1688,10 @@ def check_credentials(context: Optional[str],
     try:
         namespace = get_kube_config_context_namespace(context)
         kubernetes.core_api(context).list_namespaced_pod(
-            namespace, _request_timeout=timeout)
+            namespace, limit=1, _request_timeout=timeout)
+        # This call is "free" because this function is a cached call,
+        # and it will not be called again in this function.
+        get_kubernetes_nodes(context=context)
     except ImportError:
         # TODO(romilb): Update these error strs to also include link to docs
         #  when docs are ready.

sky/provision/provisioner.py

@@ -18,6 +18,7 @@ from sky import exceptions
 from sky import global_user_state
 from sky import logs
 from sky import provision
+from sky import resources as resources_lib
 from sky import sky_logging
 from sky import skypilot_config
 from sky.adaptors import aws
@@ -428,13 +429,14 @@ def wait_for_ssh(cluster_info: provision_common.ClusterInfo,
 
 
 def _post_provision_setup(
-        cloud_name: str, cluster_name: resources_utils.ClusterName,
-        handle_cluster_yaml: str,
+        launched_resources: resources_lib.Resources,
+        cluster_name: resources_utils.ClusterName, handle_cluster_yaml: str,
         provision_record: provision_common.ProvisionRecord,
         custom_resource: Optional[str]) -> provision_common.ClusterInfo:
     config_from_yaml = global_user_state.get_cluster_yaml_dict(
         handle_cluster_yaml)
     provider_config = config_from_yaml.get('provider')
+    cloud_name = repr(launched_resources.cloud)
     cluster_info = provision.get_cluster_info(cloud_name,
                                               provision_record.region,
                                               cluster_name.name_on_cloud,
@@ -694,8 +696,9 @@ def _post_provision_setup(
                                                     cluster_info,
                                                     ssh_credentials)
 
-        instance_setup.start_skylet_on_head_node(cluster_name.name_on_cloud,
-                                                 cluster_info, ssh_credentials)
+        instance_setup.start_skylet_on_head_node(cluster_name, cluster_info,
+                                                 ssh_credentials,
+                                                 launched_resources)
 
     logger.info(
         ux_utils.finishing_message(f'Cluster launched: {cluster_name}.',
@@ -706,8 +709,8 @@ def _post_provision_setup(
 
 @timeline.event
 def post_provision_runtime_setup(
-        cloud_name: str, cluster_name: resources_utils.ClusterName,
-        handle_cluster_yaml: str,
+        launched_resources: resources_lib.Resources,
+        cluster_name: resources_utils.ClusterName, handle_cluster_yaml: str,
         provision_record: provision_common.ProvisionRecord,
         custom_resource: Optional[str],
         log_dir: str) -> provision_common.ClusterInfo:
@@ -728,7 +731,7 @@ def post_provision_runtime_setup(
         try:
             logger.debug(_TITLE.format('System Setup After Provision'))
             return _post_provision_setup(
-                cloud_name,
+                launched_resources,
                 cluster_name,
                 handle_cluster_yaml=handle_cluster_yaml,
                 provision_record=provision_record,

sky/schemas/db/global_user_state/010_save_ssh_key.py

@@ -0,0 +1,66 @@
+"""Add ssh keys in filesystem to global user state.
+
+Revision ID: 010
+Revises: 009
+Create Date: 2025-10-07
+
+"""
+import glob
+# pylint: disable=invalid-name
+import os
+from typing import Sequence, Union
+
+from alembic import op
+import sqlalchemy as sa
+
+# revision identifiers, used by Alembic.
+revision: str = '010'
+down_revision: Union[str, Sequence[str], None] = '009'
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade():
+    """Add last_activity_time and launched_at columns to cluster history."""
+    connection = op.get_bind()
+
+    match_dirs = glob.glob(os.path.expanduser('~/.sky/clients/*/ssh'))
+    file_user_hashes = set()
+    for match_dir in match_dirs:
+        user_hash = match_dir.split('/')[-2]
+        file_user_hashes.add(user_hash)
+
+    # Get all existing ssh keys
+    existing_user_hashes = set()
+    result = connection.execute(sa.text('SELECT user_hash FROM ssh_key'))
+    for row in result:
+        existing_user_hashes.add(row[0])
+
+    user_hashes_to_add = file_user_hashes - existing_user_hashes
+    for user_hash in user_hashes_to_add:
+        match_dir = os.path.join(os.path.expanduser('~/.sky/clients'),
+                                 user_hash, 'ssh')
+        public_key_path = os.path.join(match_dir, 'sky-key.pub')
+        private_key_path = os.path.join(match_dir, 'sky-key')
+        try:
+            with open(public_key_path, 'r', encoding='utf-8') as f:
+                public_key = f.read().strip()
+            with open(private_key_path, 'r', encoding='utf-8') as f:
+                private_key = f.read().strip()
+        except FileNotFoundError:
+            # Skip if the key files are not found
+            continue
+        connection.execute(
+            sa.text('INSERT INTO ssh_key '
+                    '(user_hash, ssh_public_key, ssh_private_key) '
+                    'VALUES (:user_hash, :ssh_public_key, :ssh_private_key) '
+                    'ON CONFLICT DO NOTHING'), {
+                        'user_hash': user_hash,
+                        'ssh_public_key': public_key,
+                        'ssh_private_key': private_key
+                    })
+
+
+def downgrade():
+    """No-op for backward compatibility."""
+    pass

sky/server/common.py

@@ -950,6 +950,7 @@ def clear_local_api_server_database() -> None:
     db_path = os.path.expanduser(server_constants.API_SERVER_REQUEST_DB_PATH)
     for extension in ['', '-shm', '-wal']:
         try:
+            logger.debug(f'Removing database file {db_path}{extension}')
             os.remove(f'{db_path}{extension}')
         except FileNotFoundError:
             logger.debug(f'Database file {db_path}{extension} not found.')

sky/server/config.py

@@ -111,7 +111,9 @@ def compute_server_config(deploy: bool,
     process after API server was introduced.
     """
     cpu_count = common_utils.get_cpu_count()
+    logger.debug(f'CPU count: {cpu_count}')
     mem_size_gb = common_utils.get_mem_size_gb()
+    logger.debug(f'Memory size: {mem_size_gb}GB')
     max_parallel_for_long = _max_long_worker_parallism(cpu_count,
                                                        mem_size_gb,
                                                        local=not deploy)