konduktor-nightly 0.1.0.dev20250811105223__tar.gz → 0.1.0.dev20250813105033__tar.gz

{konduktor_nightly-0.1.0.dev20250811105223 → konduktor_nightly-0.1.0.dev20250813105033}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: konduktor-nightly
-Version: 0.1.0.dev20250811105223
+Version: 0.1.0.dev20250813105033
 Summary: GPU Cluster Health Management
 Author: Andrew Aikawa
 Author-email: asai@berkeley.edu

{konduktor_nightly-0.1.0.dev20250811105223 → konduktor_nightly-0.1.0.dev20250813105033}/konduktor/__init__.py

@@ -11,7 +11,7 @@ from konduktor.task import Task
 __all__ = ['launch', 'Resources', 'Task', 'Serving']
 
 # Replaced with the current commit when building the wheels.
-_KONDUKTOR_COMMIT_SHA = '92fe69bd3f29e7b191de663c598dfcf10738f87a'
+_KONDUKTOR_COMMIT_SHA = 'f4ba2084fac1c1030245b475323f4f3a57fd3fa3'
 os.makedirs(os.path.expanduser('~/.konduktor'), exist_ok=True)
 
 
@@ -45,5 +45,5 @@ def _get_git_commit():
 
 
 __commit__ = _get_git_commit()
-__version__ = '1.0.0.dev0.1.0.dev20250811105223'
+__version__ = '1.0.0.dev0.1.0.dev20250813105033'
 __root_dir__ = os.path.dirname(os.path.abspath(__file__))

{konduktor_nightly-0.1.0.dev20250811105223 → konduktor_nightly-0.1.0.dev20250813105033}/konduktor/backends/constants.py

@@ -8,6 +8,7 @@ USERID_LABEL = 'trainy.ai/user-id'
 USER_LABEL = 'trainy.ai/username'
 ACCELERATOR_LABEL = 'trainy.ai/accelerator'
 NUM_ACCELERATORS_LABEL = 'trainy.ai/num-accelerators'
+MAX_EXECUTION_TIME_LABEL = 'kueue.x-k8s.io/max-exec-time-seconds'
 
 # Start/stop/status related labels
 STOP_USERID_LABEL = 'trainy.ai/stop-userid'

{konduktor_nightly-0.1.0.dev20250811105223 → konduktor_nightly-0.1.0.dev20250813105033}/konduktor/backends/jobset.py

@@ -176,7 +176,7 @@ class JobsetBackend(backend.Backend):
         context = kubernetes_utils.get_current_kube_config_context_name()
         namespace = kubernetes_utils.get_kube_config_context_namespace(context)
         # TODO(asaiacai): need to set env variables in pod
-        jobset_utils.create_jobset(
+        jobset_response = jobset_utils.create_jobset(
             namespace,
             task,
             pod_spec['kubernetes']['pod_config'],
@@ -192,9 +192,10 @@ class JobsetBackend(backend.Backend):
                 ):
                     _wait_for_jobset_start(namespace, task.name)
                 try:
+                    assert jobset_response is not None
                     log_thread = threading.Thread(
                         target=log_utils.tail_logs,
-                        args=(task.name,),
+                        args=(jobset_response,),
                         daemon=True,
                     )
                     logger.info('streaming logs...')

{konduktor_nightly-0.1.0.dev20250811105223 → konduktor_nightly-0.1.0.dev20250813105033}/konduktor/backends/jobset_utils.py

@@ -39,6 +39,7 @@ JOBSET_USERID_LABEL = backend_constants.USERID_LABEL
 JOBSET_USER_LABEL = backend_constants.USER_LABEL
 JOBSET_ACCELERATOR_LABEL = backend_constants.ACCELERATOR_LABEL
 JOBSET_NUM_ACCELERATORS_LABEL = backend_constants.NUM_ACCELERATORS_LABEL
+JOBSET_MAX_EXECUTION_TIME_LABEL = backend_constants.MAX_EXECUTION_TIME_LABEL
 
 SECRET_BASENAME_LABEL = backend_constants.SECRET_BASENAME_LABEL
 
@@ -48,6 +49,7 @@ _JOBSET_METADATA_LABELS = {
     'jobset_user_label': JOBSET_USER_LABEL,
     'jobset_accelerator_label': JOBSET_ACCELERATOR_LABEL,
     'jobset_num_accelerators_label': JOBSET_NUM_ACCELERATORS_LABEL,
+    'jobset_max_execution_time_label': JOBSET_MAX_EXECUTION_TIME_LABEL,
 }
 
 
@@ -79,6 +81,7 @@ def create_jobset(
     assert task.resources is not None, 'Task resources are undefined'
     accelerator_type = task.resources.get_accelerator_type() or 'None'
     num_accelerators = task.resources.get_accelerator_count() or 0
+    labels = task.resources.labels if task.resources.labels else {}
     with tempfile.NamedTemporaryFile() as temp:
         common_utils.fill_template(
             'jobset.yaml.j2',
@@ -91,6 +94,7 @@ def create_jobset(
                 'num_accelerators': num_accelerators,
                 'completions': task.resources.get_completions(),
                 'max_restarts': task.resources.get_max_restarts(),
+                'max_execution_time': labels.get('maxRunDurationSeconds', None),
                 **_JOBSET_METADATA_LABELS,
             },
             temp.name,
@@ -430,6 +434,36 @@ def _parse_timestamp_filter(timestamp_str: str) -> datetime:
     )
 
 
+def _format_timestamp(timestamp: str) -> str:
+    """Format timestamp as MM/DD/YY HH:MMAM/PM in local timezone"""
+    # Parse UTC timestamp and convert to local time
+    dt_utc = datetime.strptime(timestamp, '%Y-%m-%dT%H:%M:%SZ').replace(
+        tzinfo=timezone.utc
+    )
+    dt_local = dt_utc.astimezone()  # Convert to local timezone
+    return dt_local.strftime('%m/%d/%y %I:%M%p')
+
+
+def _get_job_start_time(job: Dict[str, Any]) -> str:
+    for condition in job['status']['conditions']:
+        if condition['reason'] == 'ResumeJobs':
+            return condition.get('lastTransitionTime', '')
+    return '-'
+
+
+def _get_end_time_from_conditions(job: Dict[str, Any]) -> str:
+    """Extract end time from JobSet conditions (Completed or Failed)"""
+    conditions = job.get('status', {}).get('conditions', [])
+    for condition in conditions:
+        # Look for terminal conditions with status=True
+        if (
+            condition.get('type') in ['Completed', 'Failed']
+            and condition.get('status') == 'True'
+        ):
+            return condition.get('lastTransitionTime', '')
+    return '-'
+
+
 def show_status_table(
     namespace: str,
     all_users: bool,
@@ -523,15 +557,6 @@ def show_status_table(
         result = f'{days_str}{hours_str}{minutes_str}{seconds_str}'
         return result if result else '<1 minute', delta
 
-    def _format_timestamp(timestamp: str) -> str:
-        """Format timestamp as MM/DD/YY HH:MMAM/PM in local timezone"""
-        # Parse UTC timestamp and convert to local time
-        dt_utc = datetime.strptime(timestamp, '%Y-%m-%dT%H:%M:%SZ').replace(
-            tzinfo=timezone.utc
-        )
-        dt_local = dt_utc.astimezone()  # Convert to local timezone
-        return dt_local.strftime('%m/%d/%y %I:%M%p')
-
     def _get_resources(job: Dict[str, Any]) -> str:
         num_pods = int(
             job['spec']['replicatedJobs'][0]['template']['spec']['parallelism']
@@ -591,25 +616,17 @@ def show_status_table(
             if before_dt and job_creation_time >= before_dt:
                 continue
         # Get start time
-        start_time = _format_timestamp(job['metadata']['creationTimestamp'])
+        start_time = _get_job_start_time(job)
+        if start_time != '-':
+            start_time = _format_timestamp(start_time)
 
         # Get submitted time (how long ago)
         submitted_time, _ = _get_time_delta(job['metadata']['creationTimestamp'])
 
         # Get end time (from JobSet conditions)
-        def _get_end_time_from_conditions(job: Dict[str, Any]) -> str:
-            """Extract end time from JobSet conditions (Completed or Failed)"""
-            conditions = job.get('status', {}).get('conditions', [])
-            for condition in conditions:
-                # Look for terminal conditions with status=True
-                if (
-                    condition.get('type') in ['Completed', 'Failed']
-                    and condition.get('status') == 'True'
-                ):
-                    return _format_timestamp(condition.get('lastTransitionTime', ''))
-            return '-'
-
         end_time = _get_end_time_from_conditions(job)
+        if end_time != '-':
+            end_time = _format_timestamp(end_time)
 
         if all_users:
             rows.append(

{konduktor_nightly-0.1.0.dev20250811105223 → konduktor_nightly-0.1.0.dev20250813105033}/konduktor/cli.py

@@ -732,7 +732,7 @@ def logs(
     # Verify the job exists before attempting to tail logs
     # TODO(asaiacai): unify the 404 logic under jobset_utils
     try:
-        jobset_utils.get_jobset(namespace, job_id)
+        jobset_response = jobset_utils.get_jobset(namespace, job_id)
     except jobset_utils.JobNotFoundError:
         raise click.UsageError(
             f"Job '{job_id}' not found in namespace "
@@ -741,12 +741,9 @@ def logs(
             f'{colorama.Style.RESET_ALL}.'
         )
 
-    click.secho(
-        'Logs are tailed from 1 hour ago, ' 'to see more logs, check Grafana.',
-        fg='yellow',
-    )
+    assert isinstance(jobset_response, dict), f'jobset_response: {jobset_response}'
     log_utils.tail_logs(
-        job_id,
+        jobset_response,
         worker_id=node_rank,
         follow=follow,
         num_logs=num_lines,

{konduktor_nightly-0.1.0.dev20250811105223 → konduktor_nightly-0.1.0.dev20250813105033}/konduktor/templates/jobset.yaml.j2

@@ -11,6 +11,9 @@ jobset:
       {{ jobset_accelerator_label }}: "{{ accelerator_type }}"
       {{ jobset_num_accelerators_label }}: "{{ num_accelerators }}"
       {% endif %}
+      {% if max_execution_time %}
+      {{ jobset_max_execution_time_label }}: "{{ max_execution_time }}"
+      {% endif %}
       trainy.ai/konduktor-managed: "true"
       parent: "trainy"
     annotations: {}

{konduktor_nightly-0.1.0.dev20250811105223 → konduktor_nightly-0.1.0.dev20250813105033}/konduktor/utils/log_utils.py

@@ -337,45 +337,51 @@ def tail_loki_logs_ws(
 
 
 def tail_vicky_logs(
-    job_name: str,
+    jobset_response: Dict[str, Any],
     worker_id: int = 0,
-    num_logs: int = 1000,
+    num_logs: int = -1,
     follow: bool = True,
 ):
+    job_name = jobset_response['metadata']['name']
     context = kubernetes_utils.get_current_kube_config_context_name()
     namespace = kubernetes_utils.get_kube_config_context_namespace(context)
     query: Dict[str, Any] = {}
-    if num_logs > 5000:
-        # TODO(asaiacai): we should not have a limit on the number of logs, but rather
-        # let the user specify any number of lines, and we can print the last N lines.
-        # this can be done in chunks. Potentially, we can query range
-        # until we reach the end of the log and then invoke tail again.
-        # Also include checks that the job is running/ever ran.
-        raise ValueError('num_logs must be less than or equal to 5000')
-    logger.info('ignoring num_logs argument for VictoriaLogs')
     vicky_svc = kr8s.objects.Service.get(
         'vls-victoria-logs-single-server', namespace='victoria-logs'
     )
+
+    if num_logs == -1:
+        query = {}
+    else:
+        assert num_logs > 0, f'num_logs must be greater than 0, got {num_logs}'
+        query = {'limit': num_logs}
+    if follow:
+        logger.info(
+            'No end time found, tailing logs from 1 hour ago. '
+            'If logs come up empty, there might be logs just earlier '
+            'than the past hour, check Grafana or use:\n'
+            f'{colorama.Style.BRIGHT}{colorama.Fore.YELLOW}'
+            f'`konduktor tail --no-follow {job_name}`'
+            f'{colorama.Style.RESET_ALL}'
+        )
+        query['start_offset'] = '1h'
+    query['query'] = (
+        f'k8s.namespace.name: "{namespace}" AND '
+        f'batch.kubernetes.io/job-name: "{job_name}-workers-0" AND '
+        f'batch.kubernetes.io/job-completion-index: "{worker_id}"'
+    )
+
     with kr8s.portforward.PortForward(
         vicky_svc, VICKY_REMOTE_PORT, local_port='auto'
     ) as port:
         if follow:
             timeout = INFINITY
             vicky_url = f'http://localhost:{port}/select/logsql/tail'
-            query = {}
         else:
             vicky_url = f'http://localhost:{port}/select/logsql/query'
-            query = {'limit': num_logs}
             timeout = 1
         logger.debug(f'Vicky URL: {vicky_url}')
 
-        query['query'] = (
-            f'k8s.namespace.name: "{namespace}" AND '
-            f'batch.kubernetes.io/job-name: "{job_name}-workers-0" AND '
-            f'batch.kubernetes.io/job-completion-index: "{worker_id}"'
-        )
-        query['start_offset'] = '1h'
-
         try:
             logger.debug(f'Making request to {vicky_url} with query: {query}')
             with requests.post(
@@ -412,16 +418,17 @@ def tail_vicky_logs(
 
 
 def tail_logs(
-    job_name: str,
+    jobset_response: Dict[str, Any],
     worker_id: int = 0,
     num_logs: int = 1000,
     follow: bool = True,
 ):
+    job_name = jobset_response['metadata']['name']
     logs_backend = config.get_nested(('logs', 'backend'), None)
     if logs_backend == LogBackend.VICTORIA:
-        tail_vicky_logs(job_name, worker_id, num_logs, follow)
+        tail_vicky_logs(jobset_response, worker_id, num_logs, follow)
     elif logs_backend == LogBackend.LOKI:
         tail_loki_logs_ws(job_name, worker_id, num_logs, follow)
     else:
         logger.info('Defaulting to VictoriaLogs')
-        tail_vicky_logs(job_name, worker_id, num_logs, follow)
+        tail_vicky_logs(jobset_response, worker_id, num_logs, follow)

{konduktor_nightly-0.1.0.dev20250811105223 → konduktor_nightly-0.1.0.dev20250813105033}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "konduktor-nightly"
-version = "0.1.0.dev20250811105223"
+version = "0.1.0.dev20250813105033"
 description = "GPU Cluster Health Management"
 packages = [
     {include = "konduktor"}