skypilot-nightly 1.0.0.dev20251203__py3-none-any.whl → 1.0.0.dev20260112__py3-none-any.whl

sky/data/mounting_utils.py

@@ -7,6 +7,7 @@ import textwrap
 from typing import Optional
 
 from sky import exceptions
+from sky import skypilot_config
 from sky.skylet import constants
 from sky.utils import command_runner
 
@@ -223,7 +224,10 @@ def get_gcs_mount_cmd(bucket_name: str,
     """Returns a command to mount a GCS bucket using gcsfuse."""
     bucket_sub_path_arg = f'--only-dir {_bucket_sub_path} '\
         if _bucket_sub_path else ''
-    mount_cmd = ('gcsfuse -o allow_other '
+    log_file = '$(mktemp -t gcsfuse.XXXX.log)'
+    mount_cmd = (f'gcsfuse --log-file {log_file} '
+                 '--debug_fuse_errors '
+                 '-o allow_other '
                  '--implicit-dirs '
                  f'--stat-cache-capacity {_STAT_CACHE_CAPACITY} '
                  f'--stat-cache-ttl {_STAT_CACHE_TTL} '
@@ -470,6 +474,13 @@ def get_mount_cached_cmd(rclone_config: str, rclone_profile_name: str,
                                  f'{hashed_mount_path}.log')
     create_log_cmd = (f'mkdir -p {constants.RCLONE_MOUNT_CACHED_LOG_DIR} && '
                       f'touch {log_file_path}')
+
+    # Check if sequential upload is enabled via config.
+    # Default is False (parallel uploads for better performance).
+    sequential_upload = skypilot_config.get_nested(
+        ('data', 'mount_cached', 'sequential_upload'), False)
+    transfers_flag = '--transfers 1 ' if sequential_upload else ''
+
     # when mounting multiple directories with vfs cache mode, it's handled by
     # rclone to create separate cache directories at ~/.cache/rclone/vfs. It is
     # not necessary to specify separate cache directories.
@@ -488,13 +499,11 @@ def get_mount_cached_cmd(rclone_config: str, rclone_profile_name: str,
         # interval allows for faster detection of new or updated files on the
         # remote, but increases the frequency of metadata lookups.
         '--allow-other --vfs-cache-mode full --dir-cache-time 10s '
-        # '--transfers 1' guarantees the files written at the local mount point
-        # to be uploaded to the backend storage in the order of creation.
         # '--vfs-cache-poll-interval' specifies the frequency of how often
         # rclone checks the local mount point for stale objects in cache.
         # '--vfs-write-back' defines the time to write files on remote storage
         # after last use of the file in local mountpoint.
-        '--transfers 1 --vfs-cache-poll-interval 10s --vfs-write-back 1s '
+        f'{transfers_flag}--vfs-cache-poll-interval 10s --vfs-write-back 1s '
         # Have rclone evict files if the cache size exceeds 10G.
         # This is to prevent cache from growing too large and
         # using up all the disk space. Note that files that opened
@@ -502,6 +511,9 @@ def get_mount_cached_cmd(rclone_config: str, rclone_profile_name: str,
         '--vfs-cache-max-size 10G '
         # give each mount its own cache directory
         f'--cache-dir {constants.RCLONE_CACHE_DIR}/{hashed_mount_path} '
+        # Use a faster fingerprint algorithm to detect changes in files.
+        # Recommended by rclone documentation for buckets like s3.
+        '--vfs-fast-fingerprint '
         # This command produces children processes, which need to be
         # detached from the current process's terminal. The command doesn't
         # produce any output, so we aren't dropping any logs.
@@ -646,8 +658,35 @@ def get_mounting_script(
                 else
                     echo "No goofys log file found in /tmp"
                 fi
+            elif [ "$MOUNT_BINARY" = "gcsfuse" ]; then
+                echo "Looking for gcsfuse log files..."
+                # Find gcsfuse log files in /tmp (created by mktemp -t gcsfuse.XXXX.log)
+                GCSFUSE_LOGS=$(ls -t /tmp/gcsfuse.*.log 2>/dev/null | head -1)
+                if [ -n "$GCSFUSE_LOGS" ]; then
+                    echo "=== GCSFuse log file contents ==="
+                    cat "$GCSFUSE_LOGS"
+                    echo "=== End of gcsfuse log file ==="
+                else
+                    echo "No gcsfuse log file found in /tmp"
+                fi
+            elif [ "$MOUNT_BINARY" = "rclone" ]; then
+                echo "Looking for rclone log files..."
+                # Find rclone log files in ~/.sky/rclone_log/ (for MOUNT_CACHED mode)
+                RCLONE_LOG_DIR={constants.RCLONE_MOUNT_CACHED_LOG_DIR}
+                if [ -d "$RCLONE_LOG_DIR" ]; then
+                    RCLONE_LOGS=$(ls -t "$RCLONE_LOG_DIR"/*.log 2>/dev/null | head -1)
+                    if [ -n "$RCLONE_LOGS" ]; then
+                        echo "=== Rclone log file contents ==="
+                        tail -50 "$RCLONE_LOGS"
+                        echo "=== End of rclone log file ==="
+                    else
+                        echo "No rclone log file found in $RCLONE_LOG_DIR"
+                    fi
+                else
+                    echo "Rclone log directory $RCLONE_LOG_DIR not found"
+                fi
             fi
-            # TODO(kevin): Print logs from rclone, etc too for observability.
+            # TODO(kevin): Print logs from blobfuse2, etc too for observability.
             exit $MOUNT_EXIT_CODE
         fi
         echo "Mounting done."

sky/global_user_state.py

@@ -16,7 +16,7 @@ import re
 import threading
 import time
 import typing
-from typing import Any, Dict, List, Optional, Set, Tuple
+from typing import Any, Dict, List, Literal, Optional, Set, Tuple, Union
 import uuid
 
 import sqlalchemy
@@ -1020,8 +1020,46 @@ async def cluster_event_retention_daemon():
         await asyncio.sleep(sleep_amount)
 
 
-def get_cluster_events(cluster_name: Optional[str], cluster_hash: Optional[str],
-                       event_type: ClusterEventType) -> List[str]:
+@typing.overload
+def get_cluster_events(
+    cluster_name: Optional[str],
+    cluster_hash: Optional[str],
+    event_type: ClusterEventType,
+    include_timestamps: Literal[False],
+    limit: Optional[int] = ...,
+) -> List[str]:
+    ...
+
+
+@typing.overload
+def get_cluster_events(
+    cluster_name: Optional[str],
+    cluster_hash: Optional[str],
+    event_type: ClusterEventType,
+    include_timestamps: Literal[True],
+    limit: Optional[int] = ...,
+) -> List[Dict[str, Union[str, int]]]:
+    ...
+
+
+@typing.overload
+def get_cluster_events(
+    cluster_name: Optional[str],
+    cluster_hash: Optional[str],
+    event_type: ClusterEventType,
+    include_timestamps: bool = ...,
+    limit: Optional[int] = ...,
+) -> Union[List[str], List[Dict[str, Union[str, int]]]]:
+    ...
+
+
+def get_cluster_events(
+    cluster_name: Optional[str],
+    cluster_hash: Optional[str],
+    event_type: ClusterEventType,
+    include_timestamps: bool = False,
+    limit: Optional[int] = None
+) -> Union[List[str], List[Dict[str, Union[str, int]]]]:
     """Returns the cluster events for the cluster.
 
     Args:
@@ -1030,22 +1068,44 @@ def get_cluster_events(cluster_name: Optional[str], cluster_hash: Optional[str],
         cluster_hash: Hash of the cluster. Cannot be specified if cluster_name
             is specified.
         event_type: Type of the event.
+        include_timestamps: If True, returns list of dicts with 'reason' and
+            'transitioned_at' fields. If False, returns list of reason strings.
+        limit: If specified, returns at most this many events (most recent).
+            If None, returns all events.
+
+    Returns:
+        If include_timestamps is False: List of reason strings.
+        If include_timestamps is True: List of dicts with 'reason' and
+            'transitioned_at' (unix timestamp) fields.
+        Events are ordered from oldest to newest.
     """
     assert _SQLALCHEMY_ENGINE is not None
 
-    if cluster_name is not None and cluster_hash is not None:
-        raise ValueError('Cannot specify both cluster_name and cluster_hash')
-    if cluster_name is None and cluster_hash is None:
-        raise ValueError('Must specify either cluster_name or cluster_hash')
-    if cluster_name is not None:
-        cluster_hash = _get_hash_for_existing_cluster(cluster_name)
-        if cluster_hash is None:
-            raise ValueError(f'Hash for cluster {cluster_name} not found.')
-
-    with orm.Session(_SQLALCHEMY_ENGINE) as session:
-        rows = session.query(cluster_event_table).filter_by(
-            cluster_hash=cluster_hash, type=event_type.value).order_by(
-                cluster_event_table.c.transitioned_at.asc()).all()
+    cluster_hash = _resolve_cluster_hash(cluster_hash, cluster_name)
+    if cluster_hash is None:
+        raise ValueError(f'Hash for cluster {cluster_name} not found.')
+
+    with orm.Session(_SQLALCHEMY_ENGINE) as session:
+        if limit is not None:
+            # To get the most recent N events in ASC order, we use a subquery:
+            # 1. Get most recent N events (ORDER BY DESC LIMIT N)
+            # 2. Re-order them by ASC
+            subquery = session.query(cluster_event_table).filter_by(
+                cluster_hash=cluster_hash, type=event_type.value).order_by(
+                    cluster_event_table.c.transitioned_at.desc()).limit(
+                        limit).subquery()
+            rows = session.query(subquery).order_by(
+                subquery.c.transitioned_at.asc()).all()
+        else:
+            rows = session.query(cluster_event_table).filter_by(
+                cluster_hash=cluster_hash, type=event_type.value).order_by(
+                    cluster_event_table.c.transitioned_at.asc()).all()
+
+    if include_timestamps:
+        return [{
+            'reason': row.reason,
+            'transitioned_at': row.transitioned_at
+        } for row in rows]
     return [row.reason for row in rows]
 
 
@@ -1537,6 +1597,38 @@ def _get_hash_for_existing_cluster(cluster_name: str) -> Optional[str]:
     return row.cluster_hash
 
 
+def _resolve_cluster_hash(cluster_hash: Optional[str] = None,
+                          cluster_name: Optional[str] = None) -> Optional[str]:
+    """Resolve cluster_hash from either cluster_hash or cluster_name.
+
+    Validates that exactly one of cluster_hash or cluster_name is provided,
+    then resolves cluster_name to cluster_hash if needed.
+
+    Args:
+        cluster_hash: Direct cluster hash, if known.
+        cluster_name: Cluster name to resolve to hash.
+
+    Returns:
+        The cluster_hash string, or None if cluster_name was provided but
+        the cluster doesn't exist.
+
+    Raises:
+        ValueError: If both or neither of cluster_hash/cluster_name are
+        provided.
+    """
+    if cluster_hash is not None and cluster_name is not None:
+        raise ValueError(f'Cannot specify both cluster_hash ({cluster_hash}) '
+                         f'and cluster_name ({cluster_name})')
+
+    if cluster_hash is None and cluster_name is None:
+        raise ValueError('Must specify either cluster_hash or cluster_name')
+
+    if cluster_name is not None:
+        return _get_hash_for_existing_cluster(cluster_name)
+
+    return cluster_hash
+
+
 @_init_db
 @metrics_lib.time_me
 def get_launched_resources_from_cluster_hash(
@@ -2241,7 +2333,7 @@ def get_volumes(is_ephemeral: Optional[bool] = None) -> List[Dict[str, Any]]:
             rows = session.query(volume_table).all()
         else:
             rows = session.query(volume_table).filter_by(
-                is_ephemeral=is_ephemeral).all()
+                is_ephemeral=int(is_ephemeral)).all()
     records = []
     for row in rows:
         records.append({
@@ -2253,7 +2345,7 @@ def get_volumes(is_ephemeral: Optional[bool] = None) -> List[Dict[str, Any]]:
             'last_attached_at': row.last_attached_at,
             'last_use': row.last_use,
             'status': status_lib.VolumeStatus[row.status],
-            'is_ephemeral': row.is_ephemeral,
+            'is_ephemeral': bool(row.is_ephemeral),
         })
     return records
 
@@ -2316,7 +2408,7 @@ def add_volume(
             last_attached_at=last_attached_at,
             last_use=last_use,
             status=status.value,
-            is_ephemeral=is_ephemeral,
+            is_ephemeral=int(is_ephemeral),
         )
         do_update_stmt = insert_stmnt.on_conflict_do_nothing()
         session.execute(do_update_stmt)

sky/jobs/client/sdk.py

@@ -7,6 +7,7 @@ import click
 
 from sky import sky_logging
 from sky.adaptors import common as adaptors_common
+from sky.backends import backend_utils
 from sky.client import common as client_common
 from sky.client import sdk
 from sky.schemas.api import responses
@@ -100,9 +101,13 @@ def launch(
                 pool_statuses = sdk.get(pool_status_request_id)
                 if not pool_statuses:
                     raise click.UsageError(f'Pool {pool!r} not found.')
-                resources = pool_statuses[0]['requested_resources_str']
-                click.secho(f'Use resources from pool {pool!r}: {resources}.',
-                            fg='green')
+                # Show the job's requested resources, not the pool worker
+                # resources
+                job_resources_str = backend_utils.get_task_resources_str(
+                    dag.tasks[0], is_managed_job=True)
+                click.secho(
+                    f'Use resources from pool {pool!r}: {job_resources_str}.',
+                    fg='green')
                 if num_jobs is not None:
                     job_identity = f'{num_jobs} managed jobs'
             prompt = f'Launching {job_identity} {dag.name!r}. Proceed?'

sky/jobs/controller.py

@@ -2,6 +2,7 @@
 """
 import asyncio
 import io
+import json
 import os
 import pathlib
 import resource
@@ -11,7 +12,7 @@ import threading
 import time
 import traceback
 import typing
-from typing import Dict, Optional, Set
+from typing import Dict, List, Optional, Set
 
 import dotenv
 
@@ -31,6 +32,7 @@ from sky.jobs import recovery_strategy
 from sky.jobs import scheduler
 from sky.jobs import state as managed_job_state
 from sky.jobs import utils as managed_job_utils
+from sky.server import plugins
 from sky.skylet import constants
 from sky.skylet import job_lib
 from sky.usage import usage_lib
@@ -43,11 +45,16 @@ from sky.utils import controller_utils
 from sky.utils import dag_utils
 from sky.utils import status_lib
 from sky.utils import ux_utils
+from sky.utils.plugin_extensions import ExternalClusterFailure
+from sky.utils.plugin_extensions import ExternalFailureSource
 
 if typing.TYPE_CHECKING:
     import psutil
+
+    from sky.schemas.generated import jobsv1_pb2
 else:
     psutil = adaptors_common.LazyImport('psutil')
+    jobsv1_pb2 = adaptors_common.LazyImport('sky.schemas.generated.jobsv1_pb2')
 
 logger = sky_logging.init_logger('sky.jobs.controller')
 
@@ -236,6 +243,64 @@ class JobController:
             await context_utils.to_thread(managed_job_utils.terminate_cluster,
                                           cluster_name)
 
+    async def _get_job_exit_codes(
+        self, job_id: Optional[int],
+        handle: 'cloud_vm_ray_backend.CloudVmRayResourceHandle'
+    ) -> Optional[list]:
+        """Retrieve exit codes from the remote cluster.
+
+        Args:
+            job_id: The job ID on the remote cluster.
+            handle: The handle to the cluster.
+
+        Returns:
+            List of exit codes, or None if not available.
+        """
+        try:
+            use_legacy = not handle.is_grpc_enabled_with_flag
+
+            if not use_legacy:
+                try:
+                    request = jobsv1_pb2.GetJobExitCodesRequest()
+                    if job_id is not None:
+                        request.job_id = job_id
+
+                    response = await context_utils.to_thread(
+                        backend_utils.invoke_skylet_with_retries,
+                        lambda: cloud_vm_ray_backend.SkyletClient(
+                            handle.get_grpc_channel()).get_job_exit_codes(
+                                request))
+
+                    exit_codes = list(
+                        response.exit_codes) if response.exit_codes else None
+                    return exit_codes
+                except exceptions.SkyletMethodNotImplementedError:
+                    # Fall back to legacy if RPC not implemented
+                    use_legacy = True
+
+            if use_legacy:
+                # Use existing SSH-based code generation
+                code = job_lib.JobLibCodeGen.get_job_exit_codes(job_id)
+
+                returncode, stdout, stderr = await context_utils.to_thread(
+                    self._backend.run_on_head,
+                    handle,
+                    code,
+                    stream_logs=False,
+                    require_outputs=True,
+                    separate_stderr=True)
+
+                if returncode != 0:
+                    logger.debug(f'Failed to retrieve exit codes: {stderr}')
+                    return None
+
+                exit_codes = json.loads(stdout.strip())
+                return exit_codes
+        except Exception as e:  # pylint: disable=broad-except
+            logger.debug(f'Failed to retrieve job exit codes: {e}')
+            return None
+        return None
+
     async def _run_one_task(self, task_id: int, task: 'sky.Task') -> bool:
         """Busy loop monitoring cluster status and handling recovery.
 
@@ -334,6 +399,12 @@ class JobController:
             resources_str = backend_utils.get_task_resources_str(
                 task, is_managed_job=True)
 
+            # Get full_resources_json using get_resource_config which handles
+            # heterogeneous resource configurations (any_of/ordered).
+            full_resources_json = None
+            if task.resources:
+                full_resources_json = task.get_resource_config()
+
             await managed_job_state.set_starting_async(
                 self._job_id,
                 task_id,
@@ -342,9 +413,12 @@ class JobController:
                 resources_str=resources_str,
                 specs={
                     'max_restarts_on_errors':
-                        self._strategy_executor.max_restarts_on_errors
+                        self._strategy_executor.max_restarts_on_errors,
+                    'recover_on_exit_codes':
+                        self._strategy_executor.recover_on_exit_codes
                 },
-                callback_func=callback_func)
+                callback_func=callback_func,
+                full_resources_json=full_resources_json)
             logger.info(f'Submitted managed job {self._job_id} '
                         f'(task: {task_id}, name: {task.name!r}); '
                         f'{constants.TASK_ID_ENV_VAR}: {task_id_env_var}')
@@ -365,9 +439,8 @@ class JobController:
             launch_time = time.time() - launch_start
             logger.info(f'Cluster launch completed in {launch_time:.2f}s')
             assert remote_job_submitted_at is not None, remote_job_submitted_at
-        if self._pool is None:
-            job_id_on_pool_cluster = None
-        else:
+        job_id_on_pool_cluster: Optional[int] = None
+        if self._pool:
             # Update the cluster name when using pool.
             cluster_name, job_id_on_pool_cluster = (
                 await
@@ -411,6 +484,8 @@ class JobController:
             except KeyError:
                 pass
 
+        transient_job_check_error_start_time = None
+        job_check_backoff = None
         while True:
             status_check_count += 1
 
@@ -462,19 +537,38 @@ class JobController:
             # recovering, we will set the job status to None, which will force
             # enter the recovering logic.
             job_status = None
+            transient_job_check_error_reason = None
             if not force_transit_to_recovering:
                 try:
-                    job_status = await managed_job_utils.get_job_status(
-                        self._backend,
-                        cluster_name,
-                        job_id=job_id_on_pool_cluster,
-                    )
+                    job_status, transient_job_check_error_reason = await (
+                        managed_job_utils.get_job_status(
+                            self._backend,
+                            cluster_name,
+                            job_id=job_id_on_pool_cluster,
+                        ))
                 except exceptions.FetchClusterInfoError as fetch_e:
                     logger.info(
                         'Failed to fetch the job status. Start recovery.\n'
                         f'Exception: {common_utils.format_exception(fetch_e)}\n'
                         f'Traceback: {traceback.format_exc()}')
 
+            # When job status check fails, we need to retry to avoid false alarm
+            # for job failure, as it could be a transient error for
+            # communication issue.
+            if transient_job_check_error_reason is not None:
+                logger.info(
+                    'Potential transient error when fetching the job '
+                    f'status. Reason: {transient_job_check_error_reason}.\n'
+                    'Check cluster status to determine if the job is '
+                    'preempted or failed.')
+                if transient_job_check_error_start_time is None:
+                    transient_job_check_error_start_time = time.time()
+                    job_check_backoff = common_utils.Backoff(
+                        initial_backoff=1, max_backoff_factor=5)
+            else:
+                transient_job_check_error_start_time = None
+                job_check_backoff = None
+
             if job_status == job_lib.JobStatus.SUCCEEDED:
                 logger.info(f'Task {task_id} succeeded! '
                             'Getting end time and cleaning up')
@@ -550,15 +644,16 @@ class JobController:
 
             # Pull the actual cluster status from the cloud provider to
             # determine whether the cluster is preempted or failed.
-            # TODO(zhwu): For hardware failure, such as GPU failure, it may not
-            # be reflected in the cluster status, depending on the cloud, which
-            # can also cause failure of the job, and we need to recover it
-            # rather than fail immediately.
-            (cluster_status,
-             handle) = backend_utils.refresh_cluster_status_handle(
-                 cluster_name,
-                 force_refresh_statuses=set(status_lib.ClusterStatus))
-
+            # NOTE: Some failures may not be reflected in the cluster status
+            # depending on the cloud, which can also cause failure of the job.
+            # Plugins can report such failures via ExternalFailureSource.
+            # TODO(cooperc): do we need to add this to asyncio thread?
+            (cluster_status, handle) = await context_utils.to_thread(
+                backend_utils.refresh_cluster_status_handle,
+                cluster_name,
+                force_refresh_statuses=set(status_lib.ClusterStatus))
+
+            external_failures: Optional[List[ExternalClusterFailure]] = None
             if cluster_status != status_lib.ClusterStatus.UP:
                 # The cluster is (partially) preempted or failed. It can be
                 # down, INIT or STOPPED, based on the interruption behavior of
@@ -569,6 +664,15 @@ class JobController:
                 logger.info(
                     f'Cluster is preempted or failed{cluster_status_str}. '
                     'Recovering...')
+                if ExternalFailureSource.is_registered():
+                    cluster_failures = await context_utils.to_thread(
+                        ExternalFailureSource.get, cluster_name=cluster_name)
+                    if cluster_failures:
+                        logger.info(
+                            f'Detected cluster failures: {cluster_failures}')
+                        external_failures = (
+                            ExternalClusterFailure.from_failure_list(
+                                cluster_failures))
             else:
                 if job_status is not None and not job_status.is_terminal():
                     # The multi-node job is still running, continue monitoring.
@@ -612,18 +716,37 @@ class JobController:
                             'can be caused by the job taking too much memory '
                             'or other resources. Try adding more memory, CPU, '
                             f'or disk in your job definition. {failure_reason}')
+
+                    # Retrieve exit codes from the failed job
+                    exit_codes = await self._get_job_exit_codes(
+                        job_id_on_pool_cluster, handle)
+
                     should_restart_on_failure = (
-                        self._strategy_executor.should_restart_on_failure())
+                        self._strategy_executor.should_restart_on_failure(
+                            exit_codes=exit_codes))
                     if should_restart_on_failure:
                         max_restarts = (
                             self._strategy_executor.max_restarts_on_errors)
-                        logger.info(
-                            f'User program crashed '
-                            f'({managed_job_status.value}). '
-                            f'Retry the job as max_restarts_on_errors is '
-                            f'set to {max_restarts}. '
+                        exit_code_msg = (
+                            '(Retry the job as '
+                            f'max_restarts_on_errors is set to {max_restarts}. '
                             f'[{self._strategy_executor.restart_cnt_on_failure}'
-                            f'/{max_restarts}]')
+                            f'/{max_restarts}])')
+                        if (exit_codes and
+                                self._strategy_executor.recover_on_exit_codes):
+                            recover_codes = (
+                                self._strategy_executor.recover_on_exit_codes)
+                            matching_codes = [
+                                c for c in exit_codes if c in recover_codes
+                            ]
+                            if matching_codes:
+                                exit_code_msg = (
+                                    f'(Exit code(s) {matching_codes} matched '
+                                    'recover_on_exit_codes '
+                                    f'[{recover_codes}])')
+                        logger.info(
+                            'User program crashed '
+                            f'({managed_job_status.value}). {exit_code_msg}')
                     else:
                         logger.info(
                             f'Task {task_id} failed and will not be retried')
@@ -655,9 +778,42 @@ class JobController:
                     # job status. Try to recover the job (will not restart the
                     # cluster, if the cluster is healthy).
                     assert job_status is None, job_status
-                    logger.info('Failed to fetch the job status while the '
-                                'cluster is healthy. Try to recover the job '
-                                '(the cluster will not be restarted).')
+                    if transient_job_check_error_reason is not None:
+                        assert (transient_job_check_error_start_time
+                                is not None), (
+                                    transient_job_check_error_start_time,
+                                    transient_job_check_error_reason)
+                        assert job_check_backoff is not None, (
+                            job_check_backoff, transient_job_check_error_reason)
+                        elapsed = time.time(
+                        ) - transient_job_check_error_start_time
+                        if (elapsed < managed_job_utils.
+                                JOB_STATUS_FETCH_TOTAL_TIMEOUT_SECONDS):
+                            remaining_timeout = (
+                                managed_job_utils.
+                                JOB_STATUS_FETCH_TOTAL_TIMEOUT_SECONDS -
+                                elapsed)
+                            backoff_time = min(
+                                job_check_backoff.current_backoff(),
+                                remaining_timeout)
+                            logger.info(
+                                'Failed to fetch the job status while the '
+                                'cluster is healthy. Retrying to avoid false'
+                                'alarm for job failure. Retrying in '
+                                f'{backoff_time:.1f} seconds...')
+                            await asyncio.sleep(backoff_time)
+                            continue
+                        else:
+                            logger.info(
+                                'Failed to fetch the job status after retrying '
+                                f'for {elapsed:.1f} seconds. Try to recover '
+                                'the job by restarting the job/cluster.')
+                    else:
+                        logger.info(
+                            'Failed to fetch the job status due to '
+                            'unrecoverable error. Try to recover the job by'
+                            ' restarting the job/cluster.')
+
             # When the handle is None, the cluster should be cleaned up already.
             if handle is not None:
                 resources = handle.launched_resources
@@ -688,7 +844,9 @@ class JobController:
                 job_id=self._job_id,
                 task_id=task_id,
                 force_transit_to_recovering=force_transit_to_recovering,
-                callback_func=callback_func)
+                callback_func=callback_func,
+                external_failures=external_failures,
+            )
 
             recovered_time = await self._strategy_executor.recover()
 
@@ -1183,6 +1341,8 @@ async def main(controller_uuid: str):
 
     context_utils.hijack_sys_attrs()
 
+    plugins.load_plugins(plugins.ExtensionContext())
+
     controller = ControllerManager(controller_uuid)
 
     # Will happen multiple times, who cares though