Add command line arguments for k8s_glue_example.py (#196)

Co-authored-by: Meshcheryakov Ilya <i.meshcheryakov@mts.ai>
Co-authored-by: Jake Henning <59198928+jkhenning@users.noreply.github.com>

clearml_agent/commands/worker.py

@@ -79,7 +79,11 @@ from clearml_agent.definitions import (
     ENV_AGENT_FORCE_EXEC_SCRIPT,
     ENV_TEMP_STDOUT_FILE_DIR,
     ENV_AGENT_FORCE_TASK_INIT,
-    ENV_AGENT_DEBUG_GET_NEXT_TASK, ENV_ABORT_CALLBACK_CMD, ENV_ABORT_CALLBACK_CMD_TIMEOUT,
+    ENV_AGENT_DEBUG_GET_NEXT_TASK,
+    ENV_ABORT_CALLBACK_CMD,
+    ENV_ABORT_CALLBACK_CMD_TIMEOUT,
+    ENV_QUEUE_POLL_FREQ_SEC,
+    ENV_STATUS_REPORT_FREQ_SEC,
 )
 from clearml_agent.definitions import WORKING_REPOSITORY_DIR, PIP_EXTRA_INDICES
 from clearml_agent.errors import (
@@ -865,10 +869,10 @@ class Worker(ServiceCommandSection):
     )
 
     # default poll queues every _polling_interval seconds
-    _polling_interval = 5.0
+    _polling_interval = ENV_QUEUE_POLL_FREQ_SEC.get() or 5.0
 
     # machine status update intervals, seconds
-    _machine_update_interval = 30.0
+    _machine_update_interval = ENV_STATUS_REPORT_FREQ_SEC.get() or 30.0
 
     # message printed before starting task logging,
     # it will be parsed by services_mode, to identify internal docker logging start
@@ -1391,6 +1395,10 @@ class Worker(ServiceCommandSection):
                 headers=headers
             )
             if not (result.ok() and result.response):
+                try:
+                    self.log.debug("Failed creating task session: %s", result.response)
+                except:
+                    pass
                 return
             new_session = copy(session)
             new_session.config = deepcopy(session.config)
@@ -1722,16 +1730,24 @@ class Worker(ServiceCommandSection):
     def _dynamic_gpu_get_available(self, gpu_indexes):
         # key: cast to string, value: 1 (i.e. gull GPU)
         gpu_indexes = {str(g): 1 for g in gpu_indexes}
+        worker_name = self._session.config.get("agent.worker_name", "") + ':gpu'
+
+        # only return "Our" workers (requires server API +2, otherwise the selecort pattern is ignored)
         # noinspection PyBroadException
         try:
-            response = self._session.send_api(workers_api.GetAllRequest(last_seen=600))
+            response = self._session.send_api(workers_api.GetAllRequest(
+                last_seen=600,
+                worker_pattern="{}*".format(worker_name),
+                _allow_extra_fields_=True
+            ))
         except Exception:
             return None
 
-        worker_name = self._session.config.get("agent.worker_name", "") + ':gpu'
+        # filter only our workers, in case the selector pattern above was ignored due to lower version API server
         our_workers = [
             w.id for w in response.workers
-            if w.id.startswith(worker_name) and w.id != self.worker_id]
+            if w.id.startswith(worker_name) and w.id != self.worker_id
+        ]
         gpus = {}
         allocated_gpus = {}
         gpu_pattern = re.compile(r"\d+[.]?\d*[a-z]?")
@@ -2017,7 +2033,7 @@ class Worker(ServiceCommandSection):
         columns = ("id", "name", "tags")
         print("Listening to queues:")
         if dynamic_gpus:
-            columns = ("id", "name", "tags", "gpus")
+            columns = ("id", "name", "tags", "gpus (min, max)")
             for q in queues_info:
                 q['gpus'] = str(dict(dynamic_gpus).get(q['id']) or '')
         print_table(queues_info, columns=columns, titles=columns)
@@ -3477,16 +3493,24 @@ class Worker(ServiceCommandSection):
         session = session or self._session
         try:
             if stop_reason == TaskStopReason.stopped:
-                self.log("Stopping - tasks.stop was called for task")
-                self.send_logs(task_id, ["Process aborted by user"], session=session)
-                session.send_api(
-                    tasks_api.StoppedRequest(
-                        task=task_id,
-                        status_reason="task was stopped by tasks.stop",
-                        status_message=self._task_status_change_message,
-                        force=False
+                # do not change the status to stopped if the Task status is already failed
+                task_status = get_task(
+                    session, task_id, only_fields=["status"]
+                ).status
+                if str(task_status) == "failed":
+                    self.send_logs(task_id, ["Process aborted by user - Task status was Failed"], session=session)
+                    self.log("Stopping - task was already marked as failed")
+                else:
+                    self.send_logs(task_id, ["Process aborted by user"], session=session)
+                    self.log("Stopping - tasks.stop was called for task")
+                    session.send_api(
+                        tasks_api.StoppedRequest(
+                            task=task_id,
+                            status_reason="task was stopped by tasks.stop",
+                            status_message=self._task_status_change_message,
+                            force=False
+                        )
                     )
-                )
 
             elif stop_reason == TaskStopReason.status_changed:
                 try:
@@ -3736,7 +3760,7 @@ class Worker(ServiceCommandSection):
                     # revert to venv that we used inside UV
                     api = None
                     self.package_api = package_api = package_api.get_venv_manager()
-            elif not api:
+            elif self._session.config.get("agent.package_manager.type", None) == "uv" and not api:
                 # this means `agent.package_manager.uv_replace_pip` is set to true
                 print("INFO: using UV as pip drop-in replacement")
 
@@ -4894,7 +4918,7 @@ class Worker(ServiceCommandSection):
                     "declare LOCAL_PYTHON",
                     "[ ! -z $LOCAL_PYTHON ] || for i in {{20..5}}; do (which {python_single_digit}.$i 2> /dev/null || command -v {python_single_digit}.$i) && " +
                     "{python_single_digit}.$i -m pip --version && " +
-                    "export LOCAL_PYTHON=$(which {python_single_digit}.$i 2> /dev/null || command -v git) && break ; done",
+                    "export LOCAL_PYTHON=$(which {python_single_digit}.$i 2> /dev/null || command -v {python_single_digit}.$i) && break ; done",
                     "[ ! -z $LOCAL_PYTHON ] || export CLEARML_APT_INSTALL=\"$CLEARML_APT_INSTALL {python_single_digit}-pip\"",  # noqa
                     "[ -z \"$CLEARML_APT_INSTALL\" ] || "
                     "(apt-get update -y ; apt-get install -y $CLEARML_APT_INSTALL) || "
@@ -4908,11 +4932,13 @@ class Worker(ServiceCommandSection):
             docker_bash_script = " ; ".join([line for line in bash_script if line]) \
                 if not isinstance(bash_script, str) else bash_script
 
-            # make sure that if we do not have $LOCAL_PYTHON defined
-            # we set it to python3
+            # make sure that if we do not have $LOCAL_PYTHON defined, we set it to python3
+            # notice that if $LOCAL_PYTHON -m pip fails, that means we might have a broken python in the path
+            # so we set to default path and try to set global python
             update_scheme += (
                     docker_bash_script + " ; " +
                     "[ ! -z $LOCAL_PYTHON ] || export LOCAL_PYTHON={python} ; " +
+                    "$LOCAL_PYTHON -m pip --version > /dev/null || export LOCAL_PYTHON=$(PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin command -v python3) ; " +
                     "$LOCAL_PYTHON -m pip install -U {pip_version} ; " +
                     "$LOCAL_PYTHON -m pip install -U {clearml_agent_wheel} ; ").format(
                 python_single_digit=python_version.split('.')[0],

clearml_agent/definitions.py

@@ -256,6 +256,10 @@ ENV_GIT_CLONE_VERBOSE = EnvironmentConfig("CLEARML_AGENT_GIT_CLONE_VERBOSE", typ
 
 ENV_GPU_FRACTIONS = EnvironmentConfig("CLEARML_AGENT_GPU_FRACTIONS")
 
+ENV_QUEUE_POLL_FREQ_SEC = EnvironmentConfig("CLEARML_AGENT_QUEUE_POLL_FREQ_SEC", type=float)
+
+ENV_STATUS_REPORT_FREQ_SEC = EnvironmentConfig("CLEARML_AGENT_STATUS_REPORT_FREQ_SEC", type=float)
+
 
 class FileBuffering(IntEnum):
     """

clearml_agent/glue/k8s.py

@@ -90,6 +90,7 @@ class K8sIntegration(Worker):
         '[ ! -z "$CLEARML_AGENT_SKIP_CONTAINER_APT" ] || [ ! -z "$LOCAL_PYTHON" ] || '
         'apt-get install -y python3-pip || dnf install -y python3-pip',
         "[ ! -z $LOCAL_PYTHON ] || export LOCAL_PYTHON=python3",
+        "[ ! -z $CLEARML_AGENT_NO_UPDATE ] || $LOCAL_PYTHON -m pip --version > /dev/null || export LOCAL_PYTHON=$(PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin command -v python3)",
         "rm -f /usr/lib/python3.*/EXTERNALLY-MANAGED",  # remove PEP 668
         "{extra_bash_init_cmd}",
         "[ ! -z $CLEARML_AGENT_NO_UPDATE ] || $LOCAL_PYTHON -m pip install clearml-agent{agent_install_args}",
@@ -147,6 +148,10 @@ class K8sIntegration(Worker):
         :param str extra_bash_init_script: Additional bash script to run before starting the Task inside the container
         :param str namespace: K8S namespace to be used when creating the new pods (default: clearml)
         :param int max_pods_limit: Maximum number of pods that K8S glue can run at the same time
+        :param str pod_name_prefix: Define pod name prefix for k8s (default: clearml-id-)
+        :param str limit_pod_label: Define limit pod label for k8s (default: ai.allegro.agent.serial=pod-{pod_number})
+        :param bool force_system_packages: true when running tasks in containers (i.e. docker mode or k8s glue).
+            (default: true)
         """
         super(K8sIntegration, self).__init__()
         self.kind = os.environ.get("CLEARML_K8S_GLUE_KIND", "pod").strip().lower()
@@ -839,8 +844,11 @@ class K8sIntegration(Worker):
     def get_task_worker_id(self, template, task_id, pod_name, namespace, queue):
         return f"{self.worker_id}:{task_id}"
 
+    def use_image_entrypoint(self, queue: str, task_id: str, docker_image: str) -> bool:
+        return ENV_POD_USE_IMAGE_ENTRYPOINT.get()
+
     def _create_template_container(
-        self, pod_name: str, task_id: str, docker_image: str, docker_args: List[str],
+        self, pod_name: str, task_id: str, docker_image: str, docker_args: List[str], queue: str,
         docker_bash: str, clearml_conf_create_script: List[str], task_worker_id: str, task_token: str = None
     ) -> dict:
         container = self._get_docker_args(
@@ -862,7 +870,7 @@ class K8sIntegration(Worker):
         # Set worker ID
         add_or_update_env_var('CLEARML_WORKER_ID', task_worker_id)
 
-        if ENV_POD_USE_IMAGE_ENTRYPOINT.get():
+        if self.use_image_entrypoint(queue=queue, task_id=task_id, docker_image=docker_image):
             # Don't add a cmd and args, just the image
 
             # Add the task ID and token since we need it (it's usually in the init script passed to us
@@ -979,6 +987,7 @@ class K8sIntegration(Worker):
             clearml_conf_create_script=clearml_conf_create_script,
             task_worker_id=task_worker_id,
             task_token=task_token,
+            queue=queue,
         )
 
         if containers:
@@ -1292,7 +1301,7 @@ class K8sIntegration(Worker):
 
                     self.report_monitor(ResourceMonitor.StatusReport(queues=queues, queue=queue, task=task_id))
                     self.run_one_task(queue, task_id, worker_params, task_session)
-                    self.report_monitor(ResourceMonitor.StatusReport(queues=self.queues))
+                    self.report_monitor(ResourceMonitor.StatusReport(queues=queues))
                     break
             else:
                 # sleep and retry polling

clearml_agent/helper/package/pip_api/venv.py

@@ -1,3 +1,4 @@
+import sys
 from typing import Any
 
 from ...._vendor.pathlib2 import Path
@@ -72,11 +73,20 @@ class VirtualenvPip(SystemPip, PackageManager):
                 self.python, "-m", "virtualenv", self.path, *self.create_flags()
             ).check_call()
         except Exception as ex:
-            # let's try with std library instead
-            print("WARNING: virtualenv call failed: {}\n INFO: Creating virtual environment with venv".format(ex))
-            self.session.command(
-                self.python, "-m", "venv", self.path, *self.create_flags()
-            ).check_call()
+            try:
+                # let's try with std library instead
+                print("WARNING: virtualenv call failed: {}\n INFO: Creating virtual environment with venv".format(ex))
+                self.session.command(
+                    self.python, "-m", "venv", self.path, *self.create_flags()
+                ).check_call()
+            except Exception as ex:
+                # let's try with std library instead
+                print("WARNING: virtualenv and venv failed with [{}] trying virtualenv with python [{}]".format(
+                    self.python, sys.executable))
+                self.python = str(sys.executable)
+                self.session.command(
+                    self.python, "-m", "virtualenv", self.path, *self.create_flags()
+                ).check_call()
 
         return self
 

clearml_agent/helper/resource_monitor.py

@@ -2,6 +2,7 @@ from __future__ import unicode_literals, division
 
 import logging
 import re
+import os
 import shlex
 from collections import deque
 from itertools import starmap
@@ -112,7 +113,15 @@ class ResourceMonitor(object):
                 active_gpus = Session.get_nvidia_visible_env()
                 # None means no filtering, report all gpus
                 if active_gpus and active_gpus != "all":
-                    self._active_gpus = [g.strip() for g in str(active_gpus).split(',')]
+                    if os.path.isdir(active_gpus):
+                        try:
+                            self._active_gpus = os.listdir(active_gpus)
+                        except OSError as e:
+                            log.warning(
+                                "Failed listing {}: {}".format(active_gpus, e)
+                            )
+                    else:
+                        self._active_gpus = [g.strip() for g in active_gpus.split(",")]
             except Exception:
                 pass
         self._cluster_report_interval_sec = int(session.config.get(

clearml_agent/interface/worker.py

@@ -141,7 +141,9 @@ DAEMON_ARGS = dict({
         'action': 'store_true',
     },
     '--use-owner-token': {
-        'help': 'Generate and use task owner token for the execution of the task',
+        'help': 'Run tasks under the identity of each task\'s owner: all calls made by the task code during execution '
+                'will use the owner\'s credentials instead of the agent\'s. This feature requires the agent to use a '
+                'ClearML Enterprise Server.',
         'action': 'store_true',
     }
 }, **WORKER_ARGS)

examples/k8s_glue_example.py

@@ -13,9 +13,19 @@ def parse_args():
     group = parser.add_mutually_exclusive_group()
 
     parser.add_argument(
-        "--queue",
-        type=str,
-        help="Queues to pull tasks from. If multiple queues, use comma separated list, e.g. 'queue1,queue2'",
+        "--k8s-pending-queue-name", type=str,
+        help="Queue name to use when task is pending in the k8s scheduler (default: %(default)s)", default="k8s_scheduler"
+    )
+    parser.add_argument(
+        "--container-bash-script", type=str,
+        help="Path to the file with container bash script to be executed in k8s", default=None
+    )
+    parser.add_argument(
+        "--debug", action="store_true", default=False,
+        help="Switch logging on (default: %(default)s)"
+    )
+    parser.add_argument(
+        "--queue", type=str, help="Queues to pull tasks from. If multiple queues, use comma separated list, e.g. 'queue1,queue2'",
     )
     group.add_argument(
         "--ports-mode",
@@ -82,11 +92,25 @@ def parse_args():
         help="Limit the maximum number of pods that this service can run at the same time."
         "Should not be used with ports-mode",
     )
+    parser.add_argument(
+        "--pod-name-prefix", type=str,
+        help="Define pod name prefix for k8s (default: %(default)s)", default="clearml-id-"
+    )
+    parser.add_argument(
+        "--limit-pod-label", type=str,
+        help="Define limit pod label for k8s (default: %(default)s)", default="ai.allegro.agent.serial=pod-{pod_number}"
+    )
+    parser.add_argument(
+        "--no-system-packages", action="store_true", default=False,
+        help="False when running tasks in containers (default: %(default)s)"
+    )
     parser.add_argument(
         "--use-owner-token",
         action="store_true",
         default=False,
-        help="Generate and use task owner token for the execution of each task",
+        help="Run tasks under the identity of each task's owner: all calls made by the task code during execution will "
+        "use the owner's credentials instead of the agent's. This features requires the agent to use a ClearML "
+        "Enterprise Server.",
     )
     parser.add_argument(
         "--create-queue",
@@ -111,7 +135,15 @@ def main():
 
         user_props_cb = k8s_user_props_cb
 
+    if args.container_bash_script:
+        with open(args.container_bash_script, "r") as file:
+            container_bash_script = file.read().splitlines()
+    else:
+        container_bash_script = None
+
     k8s = K8sIntegration(
+        k8s_pending_queue_name=args.k8s_pending_queue_name,
+        container_bash_script=container_bash_script,
         ports_mode=args.ports_mode,
         num_of_services=args.num_of_services,
         base_pod_num=args.base_pod_num,
@@ -124,6 +156,10 @@ def main():
         else None,
         namespace=args.namespace,
         max_pods_limit=args.max_pods or None,
+        pod_name_prefix=args.pod_name_prefix,
+        limit_pod_label=args.limit_pod_label,
+        force_system_packages=not args.no_system_packages,
+        debug=args.debug,
     )
     queue = [q.strip() for q in args.queue.split(",") if q.strip()] if args.queue else None