skypilot-nightly 1.0.0.dev20251009__py3-none-any.whl → 1.0.0.dev20251107__py3-none-any.whl

sky/jobs/file_content_utils.py

@@ -0,0 +1,80 @@
+"""Utilities for managing managed job file content.
+
+The helpers in this module fetch job file content (DAG YAML/env files) from the
+database-first storage added for managed jobs, transparently falling back to
+legacy on-disk paths when needed. Consumers should prefer the string-based
+helpers so controllers never have to rely on local disk state.
+"""
+
+import os
+from typing import Optional
+
+from sky import sky_logging
+from sky.jobs import state as managed_job_state
+
+logger = sky_logging.init_logger(__name__)
+
+
+def get_job_dag_content(job_id: int) -> Optional[str]:
+    """Get DAG YAML content for a job from database or disk.
+
+    Args:
+        job_id: The job ID
+
+    Returns:
+        DAG YAML content as string, or None if not found
+    """
+    file_info = managed_job_state.get_job_file_contents(job_id)
+
+    # Prefer content stored in the database
+    if file_info['dag_yaml_content'] is not None:
+        return file_info['dag_yaml_content']
+
+    # Fallback to disk path for backward compatibility
+    dag_yaml_path = file_info.get('dag_yaml_path')
+    if dag_yaml_path and os.path.exists(dag_yaml_path):
+        try:
+            with open(dag_yaml_path, 'r', encoding='utf-8') as f:
+                content = f.read()
+                logger.debug('Loaded DAG YAML from disk for job %s: %s', job_id,
+                             dag_yaml_path)
+                return content
+        except (FileNotFoundError, IOError, OSError) as e:
+            logger.warning(
+                f'Failed to read DAG YAML from disk {dag_yaml_path}: {e}')
+
+    logger.warning(f'DAG YAML content not found for job {job_id}')
+    return None
+
+
+def get_job_env_content(job_id: int) -> Optional[str]:
+    """Get environment file content for a job from database or disk.
+
+    Args:
+        job_id: The job ID
+
+    Returns:
+        Environment file content as string, or None if not found
+    """
+    file_info = managed_job_state.get_job_file_contents(job_id)
+
+    # Prefer content stored in the database
+    if file_info['env_file_content'] is not None:
+        return file_info['env_file_content']
+
+    # Fallback to disk path for backward compatibility
+    env_file_path = file_info.get('env_file_path')
+    if env_file_path and os.path.exists(env_file_path):
+        try:
+            with open(env_file_path, 'r', encoding='utf-8') as f:
+                content = f.read()
+                logger.debug('Loaded environment file from disk for job %s: %s',
+                             job_id, env_file_path)
+                return content
+        except (FileNotFoundError, IOError, OSError) as e:
+            logger.warning(
+                f'Failed to read environment file from disk {env_file_path}: '
+                f'{e}')
+
+    # Environment file is optional, so don't warn if not found
+    return None

sky/jobs/log_gc.py

@@ -0,0 +1,201 @@
+"""Log garbage collection for managed jobs."""
+
+import asyncio
+from datetime import datetime
+import os
+import pathlib
+import shutil
+import time
+
+import anyio
+import filelock
+
+from sky import sky_logging
+from sky import skypilot_config
+from sky.jobs import constants as managed_job_constants
+from sky.jobs import state as managed_job_state
+from sky.jobs import utils as managed_job_utils
+from sky.utils import context
+from sky.utils import context_utils
+
+logger = sky_logging.init_logger(__name__)
+
+# Filelock for garbage collector leader election.
+_JOB_CONTROLLER_GC_LOCK_PATH = os.path.expanduser(
+    '~/.sky/locks/job_controller_gc.lock')
+
+_DEFAULT_TASK_LOGS_GC_RETENTION_HOURS = 24 * 7
+_DEFAULT_CONTROLLER_LOGS_GC_RETENTION_HOURS = 24 * 7
+
+_LEAST_FREQUENT_GC_INTERVAL_SECONDS = 3600
+_MOST_FREQUENT_GC_INTERVAL_SECONDS = 30
+
+
+def _next_gc_interval(retention_seconds: int) -> int:
+    """Get the next GC interval."""
+    # Run the GC at least per hour to ensure hourly accuracy and
+    # at most per 30 seconds (when retention_seconds is small) to
+    # avoid too frequent cleanup.
+    return max(min(retention_seconds, _LEAST_FREQUENT_GC_INTERVAL_SECONDS),
+               _MOST_FREQUENT_GC_INTERVAL_SECONDS)
+
+
+async def gc_controller_logs_for_job():
+    """Garbage collect job and controller logs."""
+    while True:
+        skypilot_config.reload_config()
+        controller_logs_retention = skypilot_config.get_nested(
+            ('jobs', 'controller', 'controller_logs_gc_retention_hours'),
+            _DEFAULT_CONTROLLER_LOGS_GC_RETENTION_HOURS) * 3600
+        # Negative value disables the GC
+        if controller_logs_retention >= 0:
+            logger.info(f'GC controller logs for job: retention '
+                        f'{controller_logs_retention} seconds')
+            try:
+                finished = False
+                while not finished:
+                    finished = await _clean_controller_logs_with_retention(
+                        controller_logs_retention)
+            except asyncio.CancelledError:
+                logger.info('Managed jobs logs GC task cancelled')
+                break
+            except Exception as e:  # pylint: disable=broad-except
+                logger.error(f'Error GC controller logs for job: {e}',
+                             exc_info=True)
+        else:
+            logger.info('Controller logs GC is disabled')
+
+        interval = _next_gc_interval(controller_logs_retention)
+        logger.info('Next controller logs GC is scheduled after '
+                    f'{interval} seconds')
+        await asyncio.sleep(interval)
+
+
+async def gc_task_logs_for_job():
+    """Garbage collect task logs for job."""
+    while True:
+        skypilot_config.reload_config()
+        task_logs_retention = skypilot_config.get_nested(
+            ('jobs', 'controller', 'task_logs_gc_retention_hours'),
+            _DEFAULT_TASK_LOGS_GC_RETENTION_HOURS) * 3600
+        # Negative value disables the GC
+        if task_logs_retention >= 0:
+            logger.info('GC task logs for job: '
+                        f'retention {task_logs_retention} seconds')
+            try:
+                finished = False
+                while not finished:
+                    finished = await _clean_task_logs_with_retention(
+                        task_logs_retention)
+            except asyncio.CancelledError:
+                logger.info('Task logs GC task cancelled')
+                break
+            except Exception as e:  # pylint: disable=broad-except
+                logger.error(f'Error GC task logs for job: {e}', exc_info=True)
+        else:
+            logger.info('Controller logs GC is disabled')
+
+        interval = _next_gc_interval(task_logs_retention)
+        logger.info(f'Next task logs GC is scheduled after {interval} seconds')
+        await asyncio.sleep(_next_gc_interval(task_logs_retention))
+
+
+async def _clean_controller_logs_with_retention(retention_seconds: int,
+                                                batch_size: int = 100):
+    """Clean controller logs with retention.
+
+    Returns:
+        Whether the GC of this round has finished, False means there might
+        still be more controller logs to clean.
+    """
+    assert batch_size > 0, 'Batch size must be positive'
+    jobs = await managed_job_state.get_controller_logs_to_clean_async(
+        retention_seconds, batch_size=batch_size)
+    job_ids_to_update = []
+    for job in jobs:
+        job_ids_to_update.append(job['job_id'])
+        log_file = managed_job_utils.controller_log_file_for_job(job['job_id'])
+        cleaned_at = time.time()
+        if await anyio.Path(log_file).exists():
+            ts_str = datetime.fromtimestamp(cleaned_at).strftime(
+                '%Y-%m-%d %H:%M:%S')
+            msg = f'Controller log has been cleaned at {ts_str}.'
+            # Sync down logs will reference to this file directly, so we
+            # keep the file and delete the content.
+            # TODO(aylei): refactor sync down logs if the inode usage
+            # becomes an issue.
+            async with await anyio.open_file(log_file, 'w',
+                                             encoding='utf-8') as f:
+                await f.write(msg + '\n')
+    # Batch the update, the timestamp will be not accurate but it's okay.
+    await managed_job_state.set_controller_logs_cleaned_async(
+        job_ids=job_ids_to_update, logs_cleaned_at=time.time())
+    complete = len(jobs) < batch_size
+    logger.info(f'Cleaned {len(jobs)} controller logs with retention '
+                f'{retention_seconds} seconds, complete: {complete}')
+    return complete
+
+
+async def _clean_task_logs_with_retention(retention_seconds: int,
+                                          batch_size: int = 100):
+    """Clean task logs with retention.
+
+    Returns:
+        Whether the GC of this round has finished, False means there might
+        still be more task logs to clean.
+    """
+    assert batch_size > 0, 'Batch size must be positive'
+    tasks = await managed_job_state.get_task_logs_to_clean_async(
+        retention_seconds, batch_size=batch_size)
+    tasks_to_update = []
+    for task in tasks:
+        local_log_file = anyio.Path(task['local_log_file'])
+        # We assume the log directory has the following layout:
+        # task-id/
+        #   - run.log
+        #   - tasks/
+        #     - run.log
+        # and also remove the tasks directory on cleanup.
+        task_log_dir = local_log_file.parent.joinpath('tasks')
+        await local_log_file.unlink(missing_ok=True)
+        await context_utils.to_thread(shutil.rmtree,
+                                      str(task_log_dir),
+                                      ignore_errors=True)
+        # We have at least once semantic guarantee for the cleanup here.
+        tasks_to_update.append((task['job_id'], task['task_id']))
+    await managed_job_state.set_task_logs_cleaned_async(
+        tasks=list(tasks_to_update), logs_cleaned_at=time.time())
+    complete = len(tasks) < batch_size
+    logger.info(f'Cleaned {len(tasks)} task logs with retention '
+                f'{retention_seconds} seconds, complete: {complete}')
+    return complete
+
+
+@context.contextual_async
+async def run_log_gc():
+    """Run the log garbage collector."""
+    log_dir = os.path.expanduser(managed_job_constants.JOBS_CONTROLLER_LOGS_DIR)
+    os.makedirs(log_dir, exist_ok=True)
+    log_path = os.path.join(log_dir, 'garbage_collector.log')
+    # Remove previous log file
+    await anyio.Path(log_path).unlink(missing_ok=True)
+    ctx = context.get()
+    assert ctx is not None, 'Context is not initialized'
+    ctx.redirect_log(pathlib.Path(log_path))
+    gc_controller_logs_for_job_task = asyncio.create_task(
+        gc_controller_logs_for_job())
+    gc_task_logs_for_job_task = asyncio.create_task(gc_task_logs_for_job())
+    await asyncio.gather(gc_controller_logs_for_job_task,
+                         gc_task_logs_for_job_task)
+
+
+def elect_for_log_gc():
+    """Use filelock to elect for the log garbage collector.
+
+    The log garbage collector runs in the controller process to avoid the
+    overhead of launching a new process and the lifecycle management, the
+    threads that does not elected as the log garbage collector just wait.
+    on the filelock and bring trivial overhead.
+    """
+    with filelock.FileLock(_JOB_CONTROLLER_GC_LOCK_PATH):
+        asyncio.run(run_log_gc())

sky/jobs/recovery_strategy.py

@@ -70,7 +70,6 @@ class StrategyExecutor:
         max_restarts_on_errors: int,
         job_id: int,
         task_id: int,
-        job_logger: logging.Logger,
         pool: Optional[str],
         starting: Set[int],
         starting_lock: asyncio.Lock,
@@ -85,7 +84,6 @@ class StrategyExecutor:
             max_restarts_on_errors: Maximum number of restarts on errors.
             job_id: The ID of the job.
             task_id: The ID of the task.
-            job_logger: Logger instance for this specific job.
             starting: Set of job IDs that are currently starting.
             starting_lock: Lock to synchronize starting jobs.
             starting_signal: Condition to signal when a job can start.
@@ -105,7 +103,6 @@ class StrategyExecutor:
         self.task_id = task_id
         self.pool = pool
         self.restart_cnt_on_failure = 0
-        self._logger = job_logger
         self.job_id_on_pool_cluster: Optional[int] = None
         self.starting = starting
         self.starting_lock = starting_lock
@@ -119,7 +116,6 @@ class StrategyExecutor:
         task: 'task_lib.Task',
         job_id: int,
         task_id: int,
-        job_logger: logging.Logger,
         pool: Optional[str],
         starting: Set[int],
         starting_lock: asyncio.Lock,
@@ -156,7 +152,7 @@ class StrategyExecutor:
         assert job_recovery_strategy is not None, job_recovery_name
         return job_recovery_strategy(cluster_name, backend, task,
                                      max_restarts_on_errors, job_id, task_id,
-                                     job_logger, pool, starting, starting_lock,
+                                     pool, starting, starting_lock,
                                      starting_signal)
 
     async def launch(self) -> float:
@@ -224,7 +220,7 @@ class StrategyExecutor:
                 **kwargs,
                 _try_cancel_if_cluster_is_init=True,
             )
-            self._logger.debug(f'sdk.cancel request ID: {request_id}')
+            logger.debug(f'sdk.cancel request ID: {request_id}')
             await context_utils.to_thread(
                 sdk.get,
                 request_id,
@@ -261,16 +257,15 @@ class StrategyExecutor:
                 # loop.
                 # TODO(zhwu): log the unexpected error to usage collection
                 # for future debugging.
-                self._logger.info(
-                    f'Unexpected exception: {e}\nFailed to get the '
-                    'refresh the cluster status. Retrying.')
+                logger.info(f'Unexpected exception: {e}\nFailed to get the '
+                            'refresh the cluster status. Retrying.')
                 continue
             if cluster_status != status_lib.ClusterStatus.UP:
                 # The cluster can be preempted before the job is
                 # launched.
                 # Break to let the retry launch kick in.
-                self._logger.info('The cluster is preempted before the job '
-                                  'is submitted.')
+                logger.info('The cluster is preempted before the job '
+                            'is submitted.')
                 # TODO(zhwu): we should recover the preemption with the
                 # recovery strategy instead of the current while loop.
                 break
@@ -279,7 +274,6 @@ class StrategyExecutor:
                 status = await managed_job_utils.get_job_status(
                     self.backend,
                     self.cluster_name,
-                    job_logger=self._logger,
                     job_id=self.job_id_on_pool_cluster)
             except Exception as e:  # pylint: disable=broad-except
                 # If any unexpected error happens, retry the job checking
@@ -288,9 +282,8 @@ class StrategyExecutor:
                 # get_job_status, so it should not happen here.
                 # TODO(zhwu): log the unexpected error to usage collection
                 # for future debugging.
-                self._logger.info(
-                    f'Unexpected exception: {e}\nFailed to get the '
-                    'job status. Retrying.')
+                logger.info(f'Unexpected exception: {e}\nFailed to get the '
+                            'job status. Retrying.')
                 continue
 
             # Check the job status until it is not in initialized status
@@ -306,9 +299,8 @@ class StrategyExecutor:
                 except Exception as e:  # pylint: disable=broad-except
                     # If we failed to get the job timestamp, we will retry
                     # job checking loop.
-                    self._logger.info(
-                        f'Unexpected Exception: {e}\nFailed to get '
-                        'the job start timestamp. Retrying.')
+                    logger.info(f'Unexpected Exception: {e}\nFailed to get '
+                                'the job start timestamp. Retrying.')
                     continue
             # Wait for the job to be started
             await asyncio.sleep(
@@ -370,7 +362,6 @@ class StrategyExecutor:
                         self.starting,
                         self.starting_lock,
                         self.starting_signal,
-                        self._logger,
                 ):
                     # The job state may have been PENDING during backoff -
                     # update to STARTING or RECOVERING.
@@ -394,21 +385,19 @@ class StrategyExecutor:
                                 for env_var in ENV_VARS_TO_CLEAR:
                                     vars_to_restore[env_var] = os.environ.pop(
                                         env_var, None)
-                                    self._logger.debug('Cleared env var: '
-                                                       f'{env_var}')
-                                self._logger.debug('Env vars for api_start: '
-                                                   f'{os.environ}')
+                                    logger.debug('Cleared env var: '
+                                                 f'{env_var}')
+                                logger.debug('Env vars for api_start: '
+                                             f'{os.environ}')
                                 await context_utils.to_thread(sdk.api_start)
-                                self._logger.info('API server started.')
+                                logger.info('API server started.')
                             finally:
                                 for env_var, value in vars_to_restore.items():
                                     if value is not None:
-                                        self._logger.debug(
-                                            'Restored env var: '
-                                            f'{env_var}: {value}')
+                                        logger.debug('Restored env var: '
+                                                     f'{env_var}: {value}')
                                         os.environ[env_var] = value
 
-                            log_file = _get_logger_file(self._logger)
                             request_id = None
                             try:
                                 request_id = await context_utils.to_thread(
@@ -429,31 +418,27 @@ class StrategyExecutor:
                                     # down=True,
                                     _is_launched_by_jobs_controller=True,
                                 )
-                                self._logger.debug('sdk.launch request ID: '
-                                                   f'{request_id}')
-                                if log_file is None:
-                                    raise OSError('Log file is None')
-                                with open(log_file, 'a', encoding='utf-8') as f:
-                                    await context_utils.to_thread(
-                                        sdk.stream_and_get,
-                                        request_id,
-                                        output_stream=f,
-                                    )
+                                logger.debug('sdk.launch request ID: '
+                                             f'{request_id}')
+                                await context_utils.to_thread(
+                                    sdk.stream_and_get,
+                                    request_id,
+                                )
                             except asyncio.CancelledError:
                                 if request_id:
                                     req = await context_utils.to_thread(
                                         sdk.api_cancel, request_id)
-                                    self._logger.debug('sdk.api_cancel request '
-                                                       f'ID: {req}')
+                                    logger.debug('sdk.api_cancel request '
+                                                 f'ID: {req}')
                                     try:
                                         await context_utils.to_thread(
                                             sdk.get, req)
                                     except Exception as e:  # pylint: disable=broad-except
                                         # we must still return a CancelledError
-                                        self._logger.error(
+                                        logger.error(
                                             f'Failed to cancel the job: {e}')
                                 raise
-                            self._logger.info('Managed job cluster launched.')
+                            logger.info('Managed job cluster launched.')
                         else:
                             self.cluster_name = await (context_utils.to_thread(
                                 serve_utils.get_next_cluster_name, self.pool,
@@ -468,8 +453,8 @@ class StrategyExecutor:
                                     self.dag,
                                     cluster_name=self.cluster_name,
                                 )
-                                self._logger.debug('sdk.exec request ID: '
-                                                   f'{request_id}')
+                                logger.debug('sdk.exec request ID: '
+                                             f'{request_id}')
                                 job_id_on_pool_cluster, _ = (
                                     await context_utils.to_thread(
                                         sdk.get, request_id))
@@ -477,14 +462,14 @@ class StrategyExecutor:
                                 if request_id:
                                     req = await context_utils.to_thread(
                                         sdk.api_cancel, request_id)
-                                    self._logger.debug('sdk.api_cancel request '
-                                                       f'ID: {req}')
+                                    logger.debug('sdk.api_cancel request '
+                                                 f'ID: {req}')
                                     try:
                                         await context_utils.to_thread(
                                             sdk.get, req)
                                     except Exception as e:  # pylint: disable=broad-except
                                         # we must still return a CancelledError
-                                        self._logger.error(
+                                        logger.error(
                                             f'Failed to cancel the job: {e}')
                                 raise
                             assert job_id_on_pool_cluster is not None, (
@@ -492,15 +477,14 @@ class StrategyExecutor:
                             self.job_id_on_pool_cluster = job_id_on_pool_cluster
                             await state.set_job_id_on_pool_cluster_async(
                                 self.job_id, job_id_on_pool_cluster)
-                        self._logger.info('Managed job cluster launched.')
+                        logger.info('Managed job cluster launched.')
                     except (exceptions.InvalidClusterNameError,
                             exceptions.NoCloudAccessError,
                             exceptions.ResourcesMismatchError,
                             exceptions.StorageSpecError,
                             exceptions.StorageError) as e:
-                        self._logger.error(
-                            'Failure happened before provisioning. '
-                            f'{common_utils.format_exception(e)}')
+                        logger.error('Failure happened before provisioning. '
+                                     f'{common_utils.format_exception(e)}')
                         if raise_on_failure:
                             raise exceptions.ProvisionPrechecksError(
                                 reasons=[e])
@@ -528,24 +512,22 @@ class StrategyExecutor:
                             reasons_str = '; '.join(
                                 common_utils.format_exception(err)
                                 for err in reasons)
-                            self._logger.error(
+                            logger.error(
                                 'Failure happened before provisioning. '
                                 f'Failover reasons: {reasons_str}')
                             if raise_on_failure:
                                 raise exceptions.ProvisionPrechecksError(
                                     reasons)
                             return None
-                        self._logger.info(
-                            'Failed to launch a cluster with error: '
-                            f'{common_utils.format_exception(e)})')
+                        logger.info('Failed to launch a cluster with error: '
+                                    f'{common_utils.format_exception(e)})')
                     except Exception as e:  # pylint: disable=broad-except
                         # If the launch fails, it will be recovered by the
                         # following code.
-                        self._logger.info(
-                            'Failed to launch a cluster with error: '
-                            f'{common_utils.format_exception(e)})')
+                        logger.info('Failed to launch a cluster with error: '
+                                    f'{common_utils.format_exception(e)})')
                         with ux_utils.enable_traceback():
-                            self._logger.info(
+                            logger.info(
                                 f'  Traceback: {traceback.format_exc()}')
                     else:  # No exception, the launch succeeds.
                         # At this point, a sky.launch() has succeeded. Cluster
@@ -559,7 +541,7 @@ class StrategyExecutor:
                         # launch.
                         # TODO(zhwu): log the unexpected error to usage
                         # collection for future debugging.
-                        self._logger.info(
+                        logger.info(
                             'Failed to successfully submit the job to the '
                             'launched cluster, due to unexpected submission '
                             'errors or the cluster being preempted during '
@@ -594,8 +576,8 @@ class StrategyExecutor:
                 # Calculate the backoff time and sleep.
                 gap_seconds = (backoff.current_backoff()
                                if self.pool is None else 1)
-                self._logger.info('Retrying to launch the cluster in '
-                                  f'{gap_seconds:.1f} seconds.')
+                logger.info('Retrying to launch the cluster in '
+                            f'{gap_seconds:.1f} seconds.')
                 await asyncio.sleep(gap_seconds)
                 continue
             else:
@@ -630,15 +612,14 @@ class FailoverStrategyExecutor(StrategyExecutor):
         max_restarts_on_errors: int,
         job_id: int,
         task_id: int,
-        job_logger: logging.Logger,
         pool: Optional[str],
         starting: Set[int],
         starting_lock: asyncio.Lock,
         starting_signal: asyncio.Condition,
     ) -> None:
         super().__init__(cluster_name, backend, task, max_restarts_on_errors,
-                         job_id, task_id, job_logger, pool, starting,
-                         starting_lock, starting_signal)
+                         job_id, task_id, pool, starting, starting_lock,
+                         starting_signal)
         # Note down the cloud/region of the launched cluster, so that we can
         # first retry in the same cloud/region. (Inside recover() we may not
         # rely on cluster handle, as it can be None if the cluster is
@@ -694,14 +675,13 @@ class FailoverStrategyExecutor(StrategyExecutor):
                     return job_submitted_at
 
             # Step 2
-            self._logger.debug('Terminating unhealthy cluster and reset cloud '
-                               'region.')
+            logger.debug('Terminating unhealthy cluster and reset cloud '
+                         'region.')
             await context_utils.to_thread(self._cleanup_cluster)
 
             # Step 3
-            self._logger.debug(
-                'Relaunch the cluster  without constraining to prior '
-                'cloud/region.')
+            logger.debug('Relaunch the cluster  without constraining to prior '
+                         'cloud/region.')
             # Not using self.launch to avoid the retry until up logic.
             job_submitted_at = await self._launch(max_retry=self._MAX_RETRY_CNT,
                                                   raise_on_failure=False,
@@ -709,8 +689,8 @@ class FailoverStrategyExecutor(StrategyExecutor):
             if job_submitted_at is None:
                 # Failed to launch the cluster.
                 gap_seconds = self.RETRY_INIT_GAP_SECONDS
-                self._logger.info('Retrying to recover the cluster in '
-                                  f'{gap_seconds:.1f} seconds.')
+                logger.info('Retrying to recover the cluster in '
+                            f'{gap_seconds:.1f} seconds.')
                 await asyncio.sleep(gap_seconds)
                 continue
 
@@ -755,14 +735,12 @@ class EagerFailoverStrategyExecutor(FailoverStrategyExecutor):
         # task.resources.
 
         # Step 1
-        self._logger.debug(
-            'Terminating unhealthy cluster and reset cloud region.')
+        logger.debug('Terminating unhealthy cluster and reset cloud region.')
         await context_utils.to_thread(self._cleanup_cluster)
 
         # Step 2
-        self._logger.debug(
-            'Relaunch the cluster skipping the previously launched '
-            'cloud/region.')
+        logger.debug('Relaunch the cluster skipping the previously launched '
+                     'cloud/region.')
         if self._launched_resources is not None:
             task = self.dag.tasks[0]
             requested_resources = self._launched_resources
@@ -787,9 +765,8 @@ class EagerFailoverStrategyExecutor(FailoverStrategyExecutor):
 
         while True:
             # Step 3
-            self._logger.debug(
-                'Relaunch the cluster without constraining to prior '
-                'cloud/region.')
+            logger.debug('Relaunch the cluster without constraining to prior '
+                         'cloud/region.')
             # Not using self.launch to avoid the retry until up logic.
             job_submitted_at = await self._launch(max_retry=self._MAX_RETRY_CNT,
                                                   raise_on_failure=False,
@@ -797,8 +774,8 @@ class EagerFailoverStrategyExecutor(FailoverStrategyExecutor):
             if job_submitted_at is None:
                 # Failed to launch the cluster.
                 gap_seconds = self.RETRY_INIT_GAP_SECONDS
-                self._logger.info('Retrying to recover the cluster in '
-                                  f'{gap_seconds:.1f} seconds.')
+                logger.info('Retrying to recover the cluster in '
+                            f'{gap_seconds:.1f} seconds.')
                 await asyncio.sleep(gap_seconds)
                 continue