Version bump to v0.13.3

docs/trains.conf

@@ -38,8 +38,12 @@ agent {
     # currently supported pip and conda
     # poetry is used if pip selected and repository contains poetry.lock file
     package_manager: {
-        # supported options: pip, conda
+        # supported options: pip, conda, poetry
         type: pip,
+
+        # specify pip version to use (examples "<20", "==19.3.1", "", empty string will install the latest version)
+        # pip_version: "<20"
+
         # virtual environment inheres packages from system
         system_site_packages: false,
         # install with --upgrade
@@ -83,6 +87,17 @@ agent {
     # apt cache folder used mapped into docker, for ubuntu package caching
     docker_apt_cache = ~/.trains/apt-cache
 
+    # optional arguments to pass to docker image
+    # these are local for this agent and will not be updated in the experiment's docker_cmd section
+    # extra_docker_arguments: ["--ipc=host", ]
+
+    # optional shell script to run in docker when started before the experiment is started
+    # extra_docker_shell_script: ["apt-get install -y bindfs", ]
+
+    # set to true in order to force "docker pull" before running an experiment using a docker image.
+    # This makes sure the docker image is updated.
+    docker_force_pull: false
+
     default_docker: {
         # default docker image to use when running in docker mode
         image: "nvidia/cuda"
@@ -126,6 +141,9 @@ sdk {
             quality: 87
             subsampling: 0
         }
+
+        # Support plot-per-graph fully matching Tensorboard behavior (i.e. if this is set to True, each series should have its own graph)
+        tensorboard_single_series_per_graph: False
     }
 
     network {

examples/archive_experiments.py

@@ -0,0 +1,59 @@
+#!/usr/bin/python3
+"""
+An example script that cleans up failed experiments by moving them to the archive
+"""
+import argparse
+from datetime import datetime
+
+from trains_agent import APIClient
+
+parser = argparse.ArgumentParser(description=__doc__)
+parser.add_argument("--project", "-P", help="Project ID. Only clean up experiments from this project")
+parser.add_argument("--user", "-U", help="User ID. Only clean up experiments assigned to this user")
+parser.add_argument(
+    "--status", "-S", default="failed",
+    help="Experiment status. Only clean up experiments with this status (default %(default)s)"
+)
+parser.add_argument(
+    "--iterations", "-I", type=int,
+    help="Number of iterations. Only clean up experiments with less or equal number of iterations"
+)
+parser.add_argument(
+    "--sec-from-start", "-T", type=int,
+    help="Seconds from start time. "
+         "Only clean up experiments if less or equal number of seconds have elapsed since started"
+)
+
+args = parser.parse_args()
+
+client = APIClient()
+
+tasks = client.tasks.get_all(
+    project=[args.project] if args.project else None,
+    user=[args.user] if args.user else None,
+    status=[args.status] if args.status else None,
+    system_tags=["-archived"]
+)
+
+count = 0
+
+for task in tasks:
+    if args.iterations and (task.last_iteration or 0) > args.iterations:
+        continue
+    if args.sec_from_start:
+        if not task.started:
+            continue
+        if (datetime.utcnow() - task.started.replace(tzinfo=None)).total_seconds() > args.sec_from_start:
+            continue
+
+    try:
+        client.tasks.edit(
+            task=task.id,
+            system_tags=(task.system_tags or []) + ["archived"],
+            force=True
+        )
+        count += 1
+    except Exception as ex:
+        print("Failed editing experiment: {}".format(ex))
+
+print("Cleaned up {} experiments".format(count))

examples/dynamic_cloud_cluster.ipynb

@@ -4,10 +4,10 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "# Auto-Magically Spin AWS EC2 Instances on Demand \n",
-    "# And Create a Dynamic Cluster Running *Trains-Agent*\n",
+    "# Auto-Magically Spin AWS EC2 Instances On Demand \n",
+    "# and Create a Dynamic Cluster Running *Trains-Agent*\n",
     "\n",
-    "### Define your budget and execute the notebook, That's is it\n",
+    "### Define your budget and execute the notebook, that's it\n",
     "### You now have a fully managed cluster on AWS  🎉 🎊 "
    ]
   },
@@ -166,7 +166,11 @@
     "# echo \"This is the second line\"\n",
     "# \"\"\"\n",
     "EXTRA_BASH_SCRIPT = \"\"\"\n",
-    "\"\"\""
+    "\"\"\"\n",
+    "\n",
+    "# Default docker for trains-agent when running in docker mode (requires docker v19.03 and above). \n",
+    "# Leave empty to run trains-agent in non-docker mode.\n",
+    "DEFAULT_DOCKER_IMAGE = \"nvidia/cuda\""
    ]
   },
   {
@@ -180,7 +184,8 @@
     "# maximum idle time in minutes, after which the instance will be shutdown\n",
     "MAX_IDLE_TIME_MIN = 15\n",
     "# polling interval in minutes\n",
-    "POLLING_INTERVAL_MIN = 2.0"
+    "# make sure to increase in case bash commands were added in EXTRA_BASH_SCRIPT\n",
+    "POLLING_INTERVAL_MIN = 5.0"
    ]
   },
   {
@@ -264,17 +269,21 @@
     "    # user_data script will automatically run when the instance is started. \n",
     "    # It will install the required packages for trains-agent configure it using \n",
     "    # environment variables and run trains-agent on the required queue\n",
-    "    user_data = \"\"\"#!/bin/bash \n",
+    "    user_data = \"\"\"#!/bin/bash\n",
     "    sudo apt-get update\n",
     "    sudo apt-get install -y python3-dev\n",
     "    sudo apt-get install -y python3-pip\n",
     "    sudo apt-get install -y gcc\n",
     "    sudo apt-get install -y git\n",
-    "    sudo python3 -m pip install screen\n",
-    "    sudo python3 -m pip install trains-agent\n",
+    "    sudo apt-get install -y build-essential\n",
+    "    python3 -m pip install -U pip\n",
+    "    python3 -m pip install virtualenv\n",
+    "    python3 -m virtualenv trains_agent_venv\n",
+    "    source trains_agent_venv/bin/activate\n",
+    "    python -m pip install trains-agent\n",
     "    echo 'agent.git_user=\\\"{git_user}\\\"' >> /root/trains.conf\n",
     "    echo 'agent.git_pass=\\\"{git_pass}\\\"' >> /root/trains.conf\n",
-    "    echo {trains_conf} >> /root/trains.conf\n",
+    "    echo \"{trains_conf}\" >> /root/trains.conf\n",
     "    export TRAINS_API_HOST={api_server}\n",
     "    export TRAINS_WEB_HOST={web_server}\n",
     "    export TRAINS_FILES_HOST={files_server}\n",
@@ -282,9 +291,11 @@
     "    export TRAINS_WORKER_ID={worker_id}:$DYNAMIC_INSTANCE_ID\n",
     "    export TRAINS_API_ACCESS_KEY='{access_key}'\n",
     "    export TRAINS_API_SECRET_KEY='{secret_key}'\n",
-    "    screen\n",
     "    {bash_script}\n",
-    "    python3 -m trains_agent --config-file '/root/trains.conf' daemon --queue '{queue}' --docker\"\"\".format(\n",
+    "    source ~/.bashrc\n",
+    "    python -m trains_agent --config-file '/root/trains.conf' daemon --queue '{queue}' {docker}\n",
+    "    shutdown\n",
+    "    \"\"\".format(\n",
     "        api_server=TRAINS_SERVER_API_SERVER,\n",
     "        web_server=TRAINS_SERVER_WEB_SERVER,\n",
     "        files_server=TRAINS_SERVER_FILES_SERVER,\n",
@@ -295,7 +306,8 @@
     "        git_user=TRAINS_GIT_USER,\n",
     "        git_pass=TRAINS_GIT_PASS,\n",
     "        trains_conf=EXTRA_TRAINS_CONF_ENCODED,\n",
-    "        bash_script=EXTRA_BASH_SCRIPT\n",
+    "        bash_script=EXTRA_BASH_SCRIPT,\n",
+    "        docker=\"--docker '{}'\".format(DEFAULT_DOCKER_IMAGE) if DEFAULT_DOCKER_IMAGE else \"\"\n",
     "    )\n",
     "\n",
     "    ec2 = boto3.client(\n",
@@ -344,6 +356,7 @@
     "            MaxCount=1,\n",
     "            InstanceType=resource_conf[\"instance_type\"],\n",
     "            UserData=user_data,\n",
+    "            InstanceInitiatedShutdownBehavior='terminate',\n",
     "            BlockDeviceMappings=[\n",
     "                {\n",
     "                    \"DeviceName\": resource_conf[\"ebs_device_name\"],\n",
@@ -431,6 +444,12 @@
     "    os.environ[\"TRAINS_API_SECRET_KEY\"] = TRAINS_SECRET_KEY\n",
     "    api_client = APIClient()\n",
     "\n",
+    "    # Verify the requested queues exist and create those that doesn't exist\n",
+    "    all_queues = [q.name for q in list(api_client.queues.get_all())]\n",
+    "    missing_queues = [q for q in QUEUES if q not in all_queues]\n",
+    "    for q in missing_queues:\n",
+    "        api_client.queues.create(q)\n",
+    "\n",
     "    idle_workers = {}\n",
     "    while True:\n",
     "        queue_name_to_id = {\n",

requirements.txt

@@ -5,6 +5,7 @@ future>=0.16.0
 humanfriendly>=2.1
 jsonmodels>=2.2
 jsonschema>=2.6.0
+packaging>=16.0
 pathlib2>=2.3.0
 psutil>=3.4.2
 pyhocon>=0.3.38
@@ -15,9 +16,8 @@ PyYAML>=3.12
 requests-file>=1.4.2
 requests>=2.20.0
 requirements_parser>=0.2.0
-semantic_version>=2.6.0
 six>=1.11.0
 tqdm>=4.19.5
 typing>=3.6.4
 urllib3>=1.21.1
-virtualenv>=16
+virtualenv>=16,<20

tests/conftest.py

@@ -35,7 +35,7 @@ def trains_agentyaml(tmpdir):
     def _method(template_file):
         file = tmpdir.join("trains_agent.yaml")
         with (PROJECT_ROOT / "tests/templates" / template_file).open() as f:
-            code = yaml.load(f)
+            code = yaml.load(f, Loader=yaml.SafeLoader)
             yield Namespace(code=code, file=file.strpath)
         file.write(yaml.dump(code))
     return _method

trains_agent/backend_api/config/default/agent.conf

@@ -22,9 +22,12 @@
     # currently supported pip and conda
     # poetry is used if pip selected and repository contains poetry.lock file
     package_manager: {
-        # supported options: pip, conda
+        # supported options: pip, conda, poetry
         type: pip,
 
+        # specify pip version to use (examples "<20", "==19.3.1", "", empty string will install the latest version)
+        pip_version: "<20",
+
         # virtual environment inheres packages from system
         system_site_packages: false,
 
@@ -68,6 +71,17 @@
     # apt cache folder used mapped into docker, for ubuntu package caching
     docker_apt_cache = ~/.trains/apt-cache
 
+    # optional arguments to pass to docker image
+    # these are local for this agent and will not be updated in the experiment's docker_cmd section
+    # extra_docker_arguments: ["--ipc=host", ]
+
+    # optional shell script to run in docker when started before the experiment is started
+    # extra_docker_shell_script: ["apt-get install -y bindfs", ]
+
+    # set to true in order to force "docker pull" before running an experiment using a docker image.
+    # This makes sure the docker image is updated.
+    docker_force_pull: false
+
     default_docker: {
         # default docker image to use when running in docker mode
         image: "nvidia/cuda"

trains_agent/backend_api/services/__init__.py

@@ -4,6 +4,7 @@ from .v2_4 import queues
 from .v2_4 import tasks
 from .v2_4 import workers
 from .v2_4 import events
+from .v2_4 import models
 
 __all__ = [
     'auth',
@@ -12,4 +13,5 @@ __all__ = [
     'tasks',
     'workers',
     'events',
+    'models',
 ]

trains_agent/commands/worker.py

@@ -17,7 +17,7 @@ from datetime import datetime
 from distutils.spawn import find_executable
 from functools import partial
 from itertools import chain
-from tempfile import gettempdir, mkdtemp
+from tempfile import mkdtemp
 from time import sleep, time
 from typing import Text, Optional, Any, Tuple
 
@@ -28,9 +28,6 @@ from trains_agent.backend_api.services import queues as queues_api
 from trains_agent.backend_api.services import tasks as tasks_api
 from pathlib2 import Path
 from pyhocon import ConfigTree, ConfigFactory
-from requests import Session as HTTPSession
-from requests.adapters import HTTPAdapter
-from urllib3.util import Retry
 from six.moves.urllib.parse import quote
 
 from trains_agent.helper.check_update import start_check_update_daemon
@@ -40,7 +37,10 @@ from trains_agent.definitions import (
     ENVIRONMENT_SDK_PARAMS,
     INVALID_WORKER_ID,
     PROGRAM_NAME,
-    DEFAULT_VENV_UPDATE_URL)
+    DEFAULT_VENV_UPDATE_URL,
+    ENV_TASK_EXECUTE_AS_USER,
+    ENV_K8S_HOST_MOUNT
+)
 from trains_agent.definitions import WORKING_REPOSITORY_DIR, PIP_EXTRA_INDICES
 from trains_agent.errors import APIError, CommandFailedError, Sigterm
 from trains_agent.helper.base import (
@@ -59,12 +59,17 @@ from trains_agent.helper.base import (
     is_conda,
     named_temporary_file,
     ExecutionInfo,
-    HOCONEncoder, error, get_python_path)
-from trains_agent.helper.console import ensure_text
+    HOCONEncoder,
+    error,
+    get_python_path,
+    is_linux_platform,
+    rm_file
+)
+from trains_agent.helper.console import ensure_text, print_text, decode_binary_lines
 from trains_agent.helper.package.base import PackageManager
 from trains_agent.helper.package.conda_api import CondaAPI
 from trains_agent.helper.package.horovod_req import HorovodRequirement
-from trains_agent.helper.package.pip_api.system import SystemPip
+from trains_agent.helper.package.external_req import ExternalRequirements
 from trains_agent.helper.package.pip_api.venv import VirtualenvPip
 from trains_agent.helper.package.poetry_api import PoetryConfig, PoetryAPI
 from trains_agent.helper.package.pytorch import PytorchRequirement
@@ -72,7 +77,6 @@ from trains_agent.helper.package.requirements import RequirementsManager
 from trains_agent.helper.package.venv_update_api import VenvUpdateAPI
 from trains_agent.helper.process import (
     kill_all_child_processes,
-    check_if_command_exists,
     WorkerParams,
     ExitStatus,
     Argv,
@@ -228,6 +232,8 @@ class TaskStopSignal(object):
             return self._test()
         except Exception as ex:
             self.command.log_traceback(ex)
+            # make sure we break nothing
+            return TaskStopSignal.default
 
     def _test(self):
         # type: () -> TaskStopReason
@@ -287,6 +293,7 @@ class Worker(ServiceCommandSection):
         PytorchRequirement,
         CythonRequirement,
         HorovodRequirement,
+        ExternalRequirements,
     )
 
     # poll queues every _polling_interval seconds
@@ -350,12 +357,16 @@ class Worker(ServiceCommandSection):
 
         self.is_venv_update = self._session.config.agent.venv_update.enabled
         self.poetry = PoetryConfig(self._session)
-        self.poetry.initialize()
         self.docker_image_func = None
         self._docker_image = None
         self._docker_arguments = None
+        PackageManager.set_pip_version(self._session.config.get("agent.package_manager.pip_version", None))
+        self._extra_docker_arguments = self._session.config.get("agent.extra_docker_arguments", None)
+        self._extra_shell_script = self._session.config.get("agent.extra_docker_shell_script", None)
+        self._docker_force_pull = self._session.config.get("agent.docker_force_pull", False)
         self._daemon_foreground = None
         self._standalone_mode = None
+        self._force_current_version = None
 
     def _get_requirements_manager(self, os_override=None, base_interpreter=None):
         requirements_manager = RequirementsManager(
@@ -413,6 +424,7 @@ class Worker(ServiceCommandSection):
             )
         )
 
+        docker_image = None
         if self.docker_image_func:
             try:
                 response = get_task(self._session, task_id, only_fields=["execution.docker_cmd"])
@@ -467,6 +479,19 @@ class Worker(ServiceCommandSection):
         try:
             # set WORKER ID
             os.environ['TRAINS_WORKER_ID'] = self.worker_id
+
+            if self._docker_force_pull and docker_image:
+                full_pull_cmd = ['docker', 'pull', docker_image]
+                pull_cmd = Argv(*full_pull_cmd)
+                status, stop_signal_status = self._log_command_output(
+                    task_id=task_id,
+                    cmd=pull_cmd,
+                    stdout_path=temp_stdout_name,
+                    stderr_path=temp_stderr_name,
+                    daemon=True,
+                    stop_signal=stop_signal,
+                )
+
             status, stop_signal_status = self._log_command_output(
                 task_id=task_id,
                 cmd=cmd,
@@ -610,11 +635,13 @@ class Worker(ServiceCommandSection):
         self.log.debug("starting resource monitor thread")
         print("Worker \"{}\" - ".format(self.worker_id), end='')
 
-        if not queues:
+        if queues:
+            queues = return_list(queues)
+            queues = [self._resolve_name(q, "queues") for q in queues]
+        else:
             default_queue = self._session.send_api(queues_api.GetDefaultRequest())
             queues = [default_queue.id]
 
-        queues = return_list(queues)
         queues_info = [
             self._session.send_api(
                 queues_api.GetByIdRequest(queue)
@@ -635,6 +662,7 @@ class Worker(ServiceCommandSection):
 
         # print docker image
         if docker is not False and docker is not None:
+            self._force_current_version = kwargs.get('force_current_version', False)
             temp_config, docker_image_func = self.get_docker_config_cmd(docker)
             self.dump_config(temp_config)
             self.docker_image_func = docker_image_func
@@ -730,15 +758,17 @@ class Worker(ServiceCommandSection):
     ):
         # type: (...) -> Tuple[Optional[int], TaskStopReason]
         def _print_file(file_path, prev_line_count):
-            with open(file_path, "rt") as f:
+            with open(file_path, "rb") as f:
+                binary_text = f.read()
                 # skip the previously printed lines,
-                return f.readlines()[prev_line_count:]
+                blines = binary_text.split(b'\n')[prev_line_count:]
+                return decode_binary_lines(blines)
 
         stdout = open(stdout_path, "wt")
         stderr = open(stderr_path, "wt") if stderr_path else stdout
+        stdout_line_count, stdout_last_lines = 0, []
+        stderr_line_count, stderr_last_lines = 0, []
         try:
-            stdout_line_count, stdout_last_lines = 0, []
-            stderr_line_count, stderr_last_lines = 0, []
             status = None
             stopping = False
             _last_machine_update_ts = time()
@@ -761,7 +791,7 @@ class Worker(ServiceCommandSection):
                     if daemon:
                         self.send_logs(
                             task_id=task_id,
-                            lines=["User aborted: stopping task\n"],
+                            lines=["User aborted: stopping task ({})\n".format(str(stop_reason))],
                             level="ERROR",
                         )
                         kill_all_child_processes(process.pid)
@@ -786,6 +816,16 @@ class Worker(ServiceCommandSection):
             # non zero return code
             stop_reason = 'Exception occurred'
             status = ex.returncode
+        except KeyboardInterrupt:
+            # so someone else will catch us
+            raise
+        except Exception:
+            # we should not get here, but better safe than sorry
+            stdout_line_count += self.send_logs(task_id, _print_file(stdout_path, stdout_line_count))
+            if stderr_path:
+                stderr_line_count += self.send_logs(task_id, _print_file(stderr_path, stderr_line_count))
+            stop_reason = 'Exception occurred'
+            status = -1
 
         stdout.close()
         if stderr_path:
@@ -815,7 +855,8 @@ class Worker(ServiceCommandSection):
         """
         if not lines:
             return 0
-        print("".join(lines), end="")
+        print_text("".join(lines))
+
         # remove backspaces from the text log, they look bad.
         for i, l in enumerate(lines):
             lines[i] = l.replace('\x08', '')
@@ -876,8 +917,6 @@ class Worker(ServiceCommandSection):
     ):
         if not task_id:
             raise CommandFailedError("Worker build must have valid task id")
-        if not check_if_command_exists("virtualenv"):
-            raise CommandFailedError("Worker must have virtualenv installed")
 
         self._session.print_configuration()
 
@@ -893,7 +932,15 @@ class Worker(ServiceCommandSection):
         except AttributeError:
             requirements = None
 
-        # TODO: make sure we pass the correct python_version
+        if not python_version:
+            try:
+                python_version = current_task.script.binary
+                python_version = python_version.split('/')[-1].replace('python', '')
+                # if we can cast it, we are good
+                python_version = '{:.1f}'.format(float(python_version))
+            except:
+                python_version = None
+
         venv_folder, requirements_manager = self.install_virtualenv(venv_dir=target,
                                                                     requested_python_version=python_version)
 
@@ -907,8 +954,9 @@ class Worker(ServiceCommandSection):
             repo_info,
             requirements_manager=requirements_manager,
             cached_requirements=requirements,
+            cwd=vcs.location if vcs and vcs.location else directory,
         )
-        freeze = self.freeze_task_environment()
+        freeze = self.freeze_task_environment(requirements_manager=requirements_manager)
         script_dir = directory
 
         # Summary
@@ -1006,8 +1054,6 @@ class Worker(ServiceCommandSection):
     ):
         if not task_id:
             raise CommandFailedError("Worker execute must have valid task id")
-        if not check_if_command_exists("virtualenv"):
-            raise CommandFailedError("Worker must have virtualenv installed")
 
         try:
             current_task = self._session.api_client.tasks.get_by_id(task_id)
@@ -1059,7 +1105,16 @@ class Worker(ServiceCommandSection):
         except AttributeError:
             requirements = None
 
-        venv_folder, requirements_manager = self.install_virtualenv(standalone_mode=standalone_mode)
+        try:
+            python_ver = current_task.script.binary
+            python_ver = python_ver.split('/')[-1].replace('python', '')
+            # if we can cast it, we are good
+            python_ver = '{:.1f}'.format(float(python_ver))
+        except:
+            python_ver = None
+
+        venv_folder, requirements_manager = self.install_virtualenv(standalone_mode=standalone_mode,
+                                                                    requested_python_version=python_ver)
 
         if not standalone_mode:
             if self._default_pip:
@@ -1079,11 +1134,13 @@ class Worker(ServiceCommandSection):
                 repo_info,
                 requirements_manager=requirements_manager,
                 cached_requirements=requirements,
+                cwd=vcs.location if vcs and vcs.location else directory,
             )
 
         # do not update the task packages if we are using conda,
         # it will most likely make the task environment unreproducible
-        freeze = self.freeze_task_environment(current_task.id if not self.is_conda else None)
+        freeze = self.freeze_task_environment(current_task.id if not self.is_conda else None,
+                                              requirements_manager=requirements_manager)
         script_dir = (directory if isinstance(directory, Path) else Path(directory)).absolute().as_posix()
 
         # run code
@@ -1124,10 +1181,21 @@ class Worker(ServiceCommandSection):
             self._update_commit_id(task_id, execution, repo_info)
 
         # Add the script CWD to the python path
-        python_path = get_python_path(script_dir, execution.entry_point, self.package_api)
+        python_path = get_python_path(script_dir, execution.entry_point, self.package_api) \
+            if not self.is_conda else None
         if python_path:
             os.environ['PYTHONPATH'] = python_path
 
+        # check if we want to run as another user, only supported on linux
+        if os.environ.get(ENV_TASK_EXECUTE_AS_USER, None) and is_linux_platform():
+            command, script_dir = self._run_as_user_patch(
+                command, script_dir, venv_folder,
+                self._session.config.get('sdk.storage.cache.default_base_dir'),
+                os.environ.get(ENV_TASK_EXECUTE_AS_USER))
+            use_execv = False
+        else:
+            use_execv = is_linux_platform() and not isinstance(self.package_api, (PoetryAPI, CondaAPI))
+
         print("Starting Task Execution:\n".format(task_id))
         exit_code = -1
         try:
@@ -1136,7 +1204,7 @@ class Worker(ServiceCommandSection):
                     sys.stdout.flush()
                     sys.stderr.flush()
                     os.chdir(script_dir)
-                    if not is_windows_platform():
+                    if use_execv:
                         os.execv(command.argv[0].as_posix(), tuple([command.argv[0].as_posix()])+command.argv[1:])
                     else:
                         exit_code = command.check_call(cwd=script_dir)
@@ -1144,10 +1212,10 @@ class Worker(ServiceCommandSection):
                 except subprocess.CalledProcessError as ex:
                     # non zero return code
                     exit_code = ex.returncode
-                    if is_windows_platform():
+                    if not use_execv:
                         exit(exit_code)
                 except Exception as ex:
-                    if is_windows_platform():
+                    if not use_execv:
                         exit(-1)
                     raise ex
             else:
@@ -1346,13 +1414,17 @@ class Worker(ServiceCommandSection):
                 status_message=self._task_status_change_message,
             )
 
-    def freeze_task_environment(self, task_id=None):
+    def freeze_task_environment(self, task_id=None, requirements_manager=None):
         try:
             freeze = self.package_api.freeze()
         except Exception as e:
             print("Could not freeze installed packages")
             self.log_traceback(e)
             return None
+
+        if requirements_manager:
+            freeze = requirements_manager.replace_back(freeze)
+
         if not task_id:
             return freeze
 
@@ -1377,8 +1449,12 @@ class Worker(ServiceCommandSection):
         if not repo_info:
             return None
         try:
+            if not self.poetry.enabled:
+                return None
+            self.poetry.initialize(cwd=repo_info.root)
             api = self.poetry.get_api(repo_info.root)
             if api.enabled:
+                print('Poetry Enabled: Ignoring requested python packages, using repository poetry lock file!')
                 api.install()
                 return api
         except Exception:
@@ -1386,7 +1462,7 @@ class Worker(ServiceCommandSection):
         return None
 
     def install_requirements(
-        self, execution, repo_info, requirements_manager, cached_requirements=None
+        self, execution, repo_info, requirements_manager, cached_requirements=None, cwd=None,
     ):
         # type: (ExecutionInfo, RepoInfo, RequirementsManager, Optional[dict]) -> None
         """
@@ -1399,6 +1475,8 @@ class Worker(ServiceCommandSection):
         :param requirements_manager: requirements manager for task
         :param cached_requirements: cached requirements from previous run
          """
+        if self.package_api:
+            self.package_api.cwd = cwd
         api = self._install_poetry_requirements(repo_info)
         if api:
             self.package_api = api
@@ -1420,7 +1498,7 @@ class Worker(ServiceCommandSection):
             except Exception as e:
                 self.log_traceback(e)
                 cached_requirements_failed = True
-                raise ValueError("Could not install task requirements!")
+                raise ValueError("Could not install task requirements!\n{}".format(e))
             else:
                 self.log("task requirements installation passed")
                 return
@@ -1458,10 +1536,7 @@ class Worker(ServiceCommandSection):
                 raise e
             finally:
                 if self._session.debug_mode and temp_file:
-                    try:
-                        Path(temp_file.name).unlink()
-                    except OSError:
-                        pass
+                    rm_file(temp_file.name)
             # call post installation callback
             requirements_manager.post_install()
             # mark as successful installation
@@ -1587,7 +1662,7 @@ class Worker(ServiceCommandSection):
         )
 
     def install_virtualenv(self, venv_dir=None, requested_python_version=None, standalone_mode=False):
-        # type: (str, str) -> Tuple[Path, RequirementsManager]
+        # type: (str, str, bool) -> Tuple[Path, RequirementsManager]
         """
         Install a new python virtual environment, removing the old one if exists
         :return: virtualenv directory and requirements manager to use with task
@@ -1595,14 +1670,27 @@ class Worker(ServiceCommandSection):
         requested_python_version = requested_python_version or \
                                    Text(self._session.config.get("agent.python_binary", None)) or \
                                    Text(self._session.config.get("agent.default_python", None))
-        executable_version, executable_version_suffix, executable_name = self.find_python_executable_for_version(
-            requested_python_version
-        )
+        if self.is_conda:
+            executable_version_suffix = \
+                requested_python_version[max(requested_python_version.find('python'), 0):].replace('python', '')
+            executable_name = 'python'
+        else:
+            try:
+                executable_version, executable_version_suffix, executable_name = \
+                    self.find_python_executable_for_version(requested_python_version)
+            except Exception:
+                def_python_version = Text(self._session.config.get("agent.python_binary", None)) or \
+                                     Text(self._session.config.get("agent.default_python", None))
+                print('Warning: could not locate requested Python version {}, reverting to version {}'.format(
+                    requested_python_version, def_python_version))
+                executable_version, executable_version_suffix, executable_name = \
+                    self.find_python_executable_for_version(def_python_version)
+            self._session.config.put("agent.default_python", executable_version)
+            self._session.config.put("agent.python_binary", executable_name)
+
         venv_dir = Path(venv_dir) if venv_dir else \
             Path(self._session.config["agent.venvs_dir"], executable_version_suffix)
 
-        self._session.config.put("agent.default_python", executable_version)
-        self._session.config.put("agent.python_binary", executable_name)
         first_time = not standalone_mode and (
             is_windows_platform()
             or self.is_conda
@@ -1745,9 +1833,19 @@ class Worker(ServiceCommandSection):
         # store docker arguments
         self._docker_image = docker_image
         self._docker_arguments = docker_arguments
+
+        extra_shell_script_str = ""
+        if self._extra_shell_script:
+            cmds = self._extra_shell_script
+            if not isinstance(cmds, (list, tuple)):
+                cmds = [cmds]
+            extra_shell_script_str = " ; ".join(map(str, cmds)) + " ; "
+
         docker_cmd = dict(worker_id=self.worker_id,
                           # docker_image=docker_image,
                           # docker_arguments=docker_arguments,
+                          extra_docker_arguments=self._extra_docker_arguments,
+                          extra_shell_script=extra_shell_script_str,
                           python_version=python_version, conf_file=self.temp_config_path,
                           host_apt_cache=host_apt_cache,
                           host_pip_cache=host_pip_cache,
@@ -1755,7 +1853,7 @@ class Worker(ServiceCommandSection):
                           host_cache=host_cache, mounted_cache=mounted_cache_dir,
                           host_pip_dl=host_pip_dl, mounted_pip_dl=mounted_pip_dl_dir,
                           host_vcs_cache=host_vcs_cache, mounted_vcs_cache=mounted_vcs_cache,
-                          standalone_mode=self._standalone_mode)
+                          standalone_mode=self._standalone_mode, force_current_version=self._force_current_version)
         return temp_config, partial(docker_cmd_functor, docker_cmd)
 
     @staticmethod
@@ -1766,15 +1864,26 @@ class Worker(ServiceCommandSection):
                         host_ssh_cache,
                         host_cache, mounted_cache,
                         host_pip_dl, mounted_pip_dl,
-                        host_vcs_cache, mounted_vcs_cache, standalone_mode=False):
+                        host_vcs_cache, mounted_vcs_cache,
+                        standalone_mode=False, extra_docker_arguments=None, extra_shell_script=None,
+                        force_current_version=None):
         docker = 'docker'
 
         base_cmd = [docker, 'run', '-t']
+        update_scheme = ""
+        dockers_nvidia_visible_devices = 'all'
         gpu_devices = os.environ.get('NVIDIA_VISIBLE_DEVICES', None)
         if gpu_devices is None or gpu_devices.lower().strip() == 'all':
-            base_cmd += ['--gpus', 'all', ]
+            if os.environ.get('TRAINS_DOCKER_SKIP_GPUS_FLAG', None):
+                dockers_nvidia_visible_devices = os.environ.get('NVIDIA_VISIBLE_DEVICES') or \
+                                                 dockers_nvidia_visible_devices
+            else:
+                base_cmd += ['--gpus', 'all', ]
         elif gpu_devices.strip() and gpu_devices.strip() != 'none':
-            base_cmd += ['--gpus', 'device='+gpu_devices, ]
+            if os.environ.get('TRAINS_DOCKER_SKIP_GPUS_FLAG', None):
+                dockers_nvidia_visible_devices = gpu_devices
+            else:
+                base_cmd += ['--gpus', 'device='+gpu_devices, ]
             # We are using --gpu, so we should not pass NVIDIA_VISIBLE_DEVICES, I think.
             # base_cmd += ['-e', 'NVIDIA_VISIBLE_DEVICES=' + gpu_devices, ]
 
@@ -1783,10 +1892,56 @@ class Worker(ServiceCommandSection):
                 if isinstance(docker_arguments, (list, tuple)) else [docker_arguments]
             base_cmd += [a for a in docker_arguments if a]
 
-        base_cmd += ['-e', 'TRAINS_WORKER_ID='+worker_id, ]
+        if extra_docker_arguments:
+            extra_docker_arguments = [extra_docker_arguments] \
+                if isinstance(extra_docker_arguments, six.string_types) else extra_docker_arguments
+            base_cmd += [str(a) for a in extra_docker_arguments if a]
 
-        if host_ssh_cache:
-            base_cmd += ['-v', host_ssh_cache+':/root/.ssh', ]
+        # check if running inside a kubernetes
+        if os.environ.get('KUBERNETES_SERVICE_HOST') and os.environ.get('KUBERNETES_PORT'):
+            # map network to sibling docker, unless we have other network argument
+            if not any(a.strip().startswith('--network') for a in base_cmd):
+                try:
+                    network_mode = get_bash_output(
+                        'docker inspect --format=\'{{.HostConfig.NetworkMode}}\' $(basename $(cat /proc/1/cpuset))')
+                    base_cmd += ['--network', network_mode]
+                except:
+                    pass
+            base_cmd += ['-e', 'NVIDIA_VISIBLE_DEVICES={}'.format(dockers_nvidia_visible_devices)]
+
+        # check if we need to map host folders
+        if os.environ.get(ENV_K8S_HOST_MOUNT):
+            # expect TRAINS_AGENT_K8S_HOST_MOUNT = '/mnt/host/data:/root/.trains'
+            k8s_node_mnt, _, k8s_pod_mnt = os.environ.get(ENV_K8S_HOST_MOUNT).partition(':')
+            # search and replace all the host folders with the k8s
+            host_mounts = [host_apt_cache, host_pip_cache, host_pip_dl, host_cache, host_vcs_cache]
+            for i, m in enumerate(host_mounts):
+                if k8s_pod_mnt not in m:
+                    print('Warning: K8S mount missing, ignoring cached folder {}'.format(m))
+                    host_mounts[i] = None
+                else:
+                    host_mounts[i] = m.replace(k8s_pod_mnt, k8s_node_mnt)
+            host_apt_cache, host_pip_cache, host_pip_dl, host_cache, host_vcs_cache = host_mounts
+
+            # copy the configuration file into the mounted folder
+            new_conf_file = os.path.join(k8s_pod_mnt, '.trains_agent.{}.cfg'.format(quote(worker_id, safe="")))
+            try:
+                rm_file(new_conf_file)
+                shutil.copy(conf_file, new_conf_file)
+                conf_file = new_conf_file.replace(k8s_pod_mnt, k8s_node_mnt)
+            except Exception:
+                raise ValueError('Error: could not copy configuration file into: {}'.format(new_conf_file))
+
+            if host_ssh_cache:
+                new_ssh_cache = os.path.join(k8s_pod_mnt, '.trains_agent.{}.ssh'.format(quote(worker_id, safe="")))
+                try:
+                    rm_tree(new_ssh_cache)
+                    shutil.copytree(host_ssh_cache, new_ssh_cache)
+                    host_ssh_cache = new_ssh_cache.replace(k8s_pod_mnt, k8s_node_mnt)
+                except Exception:
+                    raise ValueError('Error: could not copy .ssh directory into: {}'.format(new_ssh_cache))
+
+        base_cmd += ['-e', 'TRAINS_WORKER_ID='+worker_id, ]
 
         # if we are running a RC version, install the same version in the docker
         # because the default latest, will be a release version (not RC)
@@ -1794,38 +1949,103 @@ class Worker(ServiceCommandSection):
         try:
             from trains_agent.version import __version__
             _version_parts = __version__.split('.')
-            if 'rc' in _version_parts[-1].lower() or 'rc' in _version_parts[-2].lower():
+            if force_current_version or 'rc' in _version_parts[-1].lower() or 'rc' in _version_parts[-2].lower():
                 specify_version = '=={}'.format(__version__)
         except:
             pass
 
-        if standalone_mode:
-            update_scheme = ""
-        else:
-            update_scheme = \
+        if not standalone_mode:
+            update_scheme += \
                 "echo 'Binary::apt::APT::Keep-Downloaded-Packages \"true\";' > /etc/apt/apt.conf.d/docker-clean ; " \
                 "chown -R root /root/.cache/pip ; " \
                 "apt-get update ; " \
                 "apt-get install -y git libsm6 libxext6 libxrender-dev libglib2.0-0 {python_single_digit}-pip ; " \
-                "{python} -m pip install -U pip ; " \
+                "{python} -m pip install -U \"pip{pip_version}\" ; " \
                 "{python} -m pip install -U trains-agent{specify_version} ; ".format(
                     python_single_digit=python_version.split('.')[0],
-                    python=python_version, specify_version=specify_version)
+                    python=python_version, pip_version=PackageManager.get_pip_version(),
+                    specify_version=specify_version)
 
-        base_cmd += [
-            '-v', conf_file+':/root/trains.conf',
-            '-v', host_apt_cache+':/var/cache/apt/archives',
-            '-v', host_pip_cache+':/root/.cache/pip',
-            '-v', host_pip_dl+':'+mounted_pip_dl,
-            '-v', host_cache+':'+mounted_cache,
-            '-v', host_vcs_cache+':'+mounted_vcs_cache,
-            '--rm', docker_image, 'bash', '-c',
-            update_scheme +
-            "NVIDIA_VISIBLE_DEVICES=all {python} -u -m trains_agent ".format(python=python_version)
-        ]
+        base_cmd += (
+            ['-v', conf_file+':/root/trains.conf'] +
+            (['-v', host_ssh_cache+':/root/.ssh'] if host_ssh_cache else []) +
+            (['-v', host_apt_cache+':/var/cache/apt/archives'] if host_apt_cache else []) +
+            (['-v', host_pip_cache+':/root/.cache/pip'] if host_pip_cache else []) +
+            (['-v', host_pip_dl+':'+mounted_pip_dl] if host_pip_dl else []) +
+            (['-v', host_cache+':'+mounted_cache] if host_cache else []) +
+            (['-v', host_vcs_cache+':'+mounted_vcs_cache] if host_vcs_cache else []) +
+            ['--rm', docker_image, 'bash', '-c',
+                update_scheme +
+                extra_shell_script +
+                "NVIDIA_VISIBLE_DEVICES={nv_visible} {python} -u -m trains_agent ".format(
+                    nv_visible=dockers_nvidia_visible_devices, python=python_version)
+             ])
 
         return base_cmd
 
+    def _run_as_user_patch(self, command, script_dir, venv_folder, sdk_cache_folder, user_uid):
+        class RunasArgv(Argv):
+            def __init__(self, *args):
+                super(RunasArgv, self).__init__(*args)
+                self.uid = 0
+                self.gid = 0
+
+            def call_subprocess(self, func, censor_password=False, *args, **kwargs):
+                self._log.debug("running: %s: %s", func.__name__, list(self))
+                with self.normalize_exception(censor_password):
+                    return func(list(self), *args, preexec_fn=self._change_uid, **kwargs)
+
+            def set_uid(self, user_uid, user_gid):
+                from pwd import getpwnam
+                self.uid = getpwnam(user_uid).pw_uid
+                self.gid = getpwnam(user_gid).pw_gid
+
+            def _change_uid(self):
+                os.setgid(self.gid)
+                os.setuid(self.uid)
+
+        # create a home folder for our user
+        try:
+            home_folder = '/trains_agent_home'
+            rm_tree(home_folder)
+            Path(home_folder).mkdir(parents=True, exist_ok=True)
+        except:
+            home_folder = '/home/trains_agent_home'
+            rm_tree(home_folder)
+            Path(home_folder).mkdir(parents=True, exist_ok=True)
+
+        # move our entire venv into the new home
+        venv_folder = venv_folder.as_posix()
+        if not venv_folder.endswith(os.path.sep):
+            venv_folder += os.path.sep
+        new_venv_folder = os.path.join(home_folder, 'venv/')
+        shutil.move(venv_folder, new_venv_folder)
+        # allow everyone to access it
+        for f in Path(new_venv_folder).rglob('*'):
+            try:
+                f.chmod(0o0777)
+            except:
+                pass
+        # make sure we will be able to access the cache folder (we assume we have the ability change mod)
+        if sdk_cache_folder:
+            sdk_cache_folder = Path(os.path.expandvars(sdk_cache_folder)).expanduser().absolute()
+            for f in sdk_cache_folder.rglob('*'):
+                try:
+                    f.chmod(0o0777)
+                except:
+                    pass
+
+        # patch venv folder to new location
+        script_dir = script_dir.replace(venv_folder, new_venv_folder)
+        # New command line execution
+        command = RunasArgv('bash', '-c', 'HOME=\"{}\" PATH=\"{}\" {}'.format(
+            home_folder,
+            os.environ.get('PATH', '').replace(venv_folder, new_venv_folder),
+            command.serialize().replace(venv_folder, new_venv_folder)))
+        command.set_uid(user_uid=user_uid, user_gid=user_uid)
+
+        return command, script_dir
+
     def _singleton(self):
         # ensure singleton
         worker_id = self._session.config["agent.worker_id"]
@@ -1839,7 +2059,8 @@ class Worker(ServiceCommandSection):
             else:
                 worker_name = '{}:cpu'.format(worker_name)
 
-        self.worker_id, worker_slot = Singleton.register_instance(unique_worker_id=worker_id, worker_name=worker_name)
+        self.worker_id, worker_slot = Singleton.register_instance(unique_worker_id=worker_id, worker_name=worker_name,
+                                                                  api_client=self._session.api_client)
         if self.worker_id is None:
             error('Instance with the same WORKER_ID [{}] is already running'.format(worker_id))
             exit(1)

trains_agent/definitions.py

@@ -73,6 +73,12 @@ ENVIRONMENT_CONFIG = {
     "agent.cpu_only": EnvironmentConfig(
         "TRAINS_CPU_ONLY", "ALG_CPU_ONLY", "CPU_ONLY", type=bool
     ),
+    "sdk.aws.s3.key": EnvironmentConfig("AWS_ACCESS_KEY_ID"),
+    "sdk.aws.s3.secret": EnvironmentConfig("AWS_SECRET_ACCESS_KEY"),
+    "sdk.aws.s3.region": EnvironmentConfig("AWS_DEFAULT_REGION"),
+    "sdk.azure.storage.containers.0": {'account_name': EnvironmentConfig("AZURE_STORAGE_ACCOUNT"),
+                                       'account_key': EnvironmentConfig("AZURE_STORAGE_KEY")},
+    "sdk.google.storage.credentials_json": EnvironmentConfig("GOOGLE_APPLICATION_CREDENTIALS"),
 }
 
 CONFIG_FILE_ENV = EnvironmentConfig("ALG_CONFIG_FILE")
@@ -114,6 +120,8 @@ DEFAULT_VCS_CACHE = normalize_path(CONFIG_DIR, "vcs-cache")
 PIP_EXTRA_INDICES = [
 ]
 DEFAULT_PIP_DOWNLOAD_CACHE = normalize_path(CONFIG_DIR, "pip-download-cache")
+ENV_TASK_EXECUTE_AS_USER = 'TRAINS_AGENT_EXEC_USER'
+ENV_K8S_HOST_MOUNT = 'TRAINS_AGENT_K8S_HOST_MOUNT'
 
 
 class FileBuffering(IntEnum):

trains_agent/glue/k8s.py

@@ -0,0 +1,169 @@
+from __future__ import print_function, division, unicode_literals
+
+import logging
+import os
+import subprocess
+from time import sleep
+from typing import Text, List
+
+from pyhocon import HOCONConverter
+
+from trains_agent.commands.events import Events
+from trains_agent.commands.worker import Worker
+from trains_agent.helper.process import get_bash_output
+from trains_agent.helper.resource_monitor import ResourceMonitor
+
+
+class K8sIntegration(Worker):
+    K8S_PENDING_QUEUE = "k8s_scheduler"
+
+    KUBECTL_RUN_CMD = "kubectl run trains_id_{task_id} " \
+                      "--image {docker_image} " \
+                      "--restart=Never --replicas=1 " \
+                      "--generator=run-pod/v1"
+
+    KUBECTL_DELETE_CMD = "kubectl delete pods " \
+                         "--selector=TRAINS=agent " \
+                         "--field-selector=status.phase!=Pending,status.phase!=Running"
+
+    CONTAINER_BASH_SCRIPT = "apt-get install -y git python-pip && " \
+                            "pip install trains-agent && " \
+                            "python -u -m trains_agent execute --full-monitoring --require-queue --id {}"
+
+    def __init__(self, k8s_pending_queue_name=None, kubectl_cmd=None, container_bash_script=None, debug=False):
+        """
+        Initialize the k8s integration glue layer daemon
+
+        :param str k8s_pending_queue_name: queue name to use when task is pending in the k8s scheduler
+        :param str|callable kubectl_cmd: kubectl command line str, supports formating (default: KUBECTL_RUN_CMD)
+            example: "task={task_id} image={docker_image} queue_id={queue_id}"
+            or a callable function: kubectl_cmd(task_id, docker_image, queue_id, task_data)
+        :param str container_bash_script: container bash script to be executed in k8s (default: CONTAINER_BASH_SCRIPT)
+        :param bool debug: Switch logging on
+        """
+        super(K8sIntegration, self).__init__()
+        self.k8s_pending_queue_name = k8s_pending_queue_name or self.K8S_PENDING_QUEUE
+        self.kubectl_cmd = kubectl_cmd or self.KUBECTL_RUN_CMD
+        self.container_bash_script = container_bash_script or self.CONTAINER_BASH_SCRIPT
+        # Always do system packages, because by we will be running inside a docker
+        self._session.config.put("agent.package_manager.system_site_packages", True)
+        # Add debug logging
+        if debug:
+            self.log.logger.disabled = False
+            self.log.logger.setLevel(logging.INFO)
+
+    def run_one_task(self, queue: Text, task_id: Text, worker_args=None):
+        task_data = self._session.api_client.tasks.get_all(id=[task_id])[0]
+
+        # push task into the k8s queue, so we have visibility on pending tasks in the k8s scheduler
+        try:
+            self._session.api_client.tasks.enqueue(task_id, queue=self.k8s_pending_queue_name,
+                                                   status_reason='k8s pending scheduler')
+        except Exception as e:
+            self.log.error("ERROR: Could not push back task [{}] to k8s pending queue [{}], error: {}".format(
+                task_id, self.k8s_pending_queue_name, e))
+            return
+
+        if task_data.execution.docker_cmd:
+            docker_image = task_data.execution.docker_cmd
+        else:
+            docker_image = str(os.environ.get("TRAINS_DOCKER_IMAGE") or
+                               self._session.config.get("agent.default_docker.image", "nvidia/cuda"))
+
+        # take the first part, this is the docker image name (not arguments)
+        docker_image = docker_image.split()[0]
+
+        create_trains_conf = "echo '{}' >> ~/trains.conf && ".format(
+            HOCONConverter.to_hocon(self._session.config._config))
+
+        if callable(self.kubectl_cmd):
+            kubectl_cmd = self.kubectl_cmd(task_id, docker_image, queue, task_data)
+        else:
+            kubectl_cmd = self.kubectl_cmd.format(task_id=task_id, docker_image=docker_image, queue_id=queue)
+
+        # make sure we gave a list
+        if isinstance(kubectl_cmd, str):
+            kubectl_cmd = kubectl_cmd.split()
+
+        kubectl_cmd += ["--labels=TRAINS=agent", "--command", "--", "/bin/sh", "-c",
+                        create_trains_conf + self.container_bash_script.format(task_id)]
+        process = subprocess.Popen(kubectl_cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
+        output, error = process.communicate()
+        self.log.info("K8s scheduling experiment task id={}".format(task_id))
+        if error:
+            self.log.error("Running kubectl encountered an error: {}".format(
+                error if isinstance(error, str) else error.decode()))
+
+    def run_tasks_loop(self, queues: List[Text], worker_params):
+        """
+        :summary: Pull and run tasks from queues.
+        :description: 1. Go through ``queues`` by order.
+                      2. Try getting the next task for each and run the first one that returns.
+                      3. Go to step 1
+        :param queues: IDs of queues to pull tasks from
+        :type queues: list of ``Text``
+        :param worker_params: Worker command line arguments
+        :type worker_params: ``trains_agent.helper.process.WorkerParams``
+        """
+        events_service = self.get_service(Events)
+
+        # make sure we have a k8s pending queue
+        try:
+            self._session.api_client.queues.create(self.k8s_pending_queue_name)
+        except Exception:
+            pass
+        # get queue id
+        self.k8s_pending_queue_name = self._resolve_name(self.k8s_pending_queue_name, "queues")
+
+        _last_machine_update_ts = 0
+        while True:
+            # iterate over queues (priority style, queues[0] is highest)
+            for queue in queues:
+                # delete old completed /failed pods
+                get_bash_output(self.KUBECTL_DELETE_CMD)
+
+                # get next task in queue
+                try:
+                    response = self._session.api_client.queues.get_next_task(queue=queue)
+                except Exception as e:
+                    print("Warning: Could not access task queue [{}], error: {}".format(queue, e))
+                    continue
+                else:
+                    try:
+                        task_id = response.entry.task
+                    except AttributeError:
+                        print("No tasks in queue {}".format(queue))
+                        continue
+                    events_service.send_log_events(
+                        self.worker_id,
+                        task_id=task_id,
+                        lines="task {} pulled from {} by worker {}".format(
+                            task_id, queue, self.worker_id
+                        ),
+                        level="INFO",
+                    )
+
+                    self.report_monitor(ResourceMonitor.StatusReport(queues=queues, queue=queue, task=task_id))
+                    self.run_one_task(queue, task_id, worker_params)
+                    self.report_monitor(ResourceMonitor.StatusReport(queues=self.queues))
+                    break
+            else:
+                # sleep and retry polling
+                print("No tasks in Queues, sleeping for {:.1f} seconds".format(self._polling_interval))
+                sleep(self._polling_interval)
+
+            if self._session.config["agent.reload_config"]:
+                self.reload_config()
+
+    def k8s_daemon(self, queues):
+        """
+        Start the k8s Glue service.
+        This service will be pulling tasks from *queues* and scheduling them for execution using kubectl.
+        Notice all scheduled tasks are pushed back into K8S_PENDING_QUEUE,
+        and popped when execution actually starts. This creates full visibility into the k8s scheduler.
+        Manually popping a task from the K8S_PENDING_QUEUE,
+        will cause the k8s scheduler to skip the execution once the scheduled tasks needs to be executed
+
+        :param list(str) queues: List of queue names to pull from
+        """
+        return self.daemon(queues=queues, log_level=logging.INFO, foreground=True, docker=False)

trains_agent/helper/base.py

@@ -157,6 +157,10 @@ def is_windows_platform():
     return any(platform.win32_ver())
 
 
+def is_linux_platform():
+    return 'linux' in platform.system().lower()
+
+
 def normalize_path(*paths):
     """
     normalize_path
@@ -459,6 +463,17 @@ def rm_tree(root):  # type: (Union[Path, Text]) -> None
     return shutil.rmtree(os.path.expanduser(os.path.expandvars(Text(root))), onerror=on_error)
 
 
+def rm_file(filename):  # type: (Union[Path, Text]) -> None
+    """
+    A version of os.unlink that will not raise error
+    """
+    try:
+        os.unlink(os.path.expanduser(os.path.expandvars(Text(filename))))
+    except:
+        return False
+    return True
+
+
 def is_conda(config):
     return config['agent.package_manager.type'].lower() == 'conda'
 

trains_agent/helper/check_update.py

@@ -4,7 +4,7 @@ from time import sleep
 import requests
 import json
 from threading import Thread
-from semantic_version import Version
+from packaging import version as packaging_version
 from ..version import __version__
 
 __check_update_thread = None
@@ -30,8 +30,8 @@ def _check_new_version_available():
         return None
     trains_answer = update_server_releases.get("trains-agent", {})
     latest_version = trains_answer.get("version")
-    cur_version = Version(cur_version)
-    latest_version = Version(latest_version)
+    cur_version = packaging_version.parse(cur_version)
+    latest_version = packaging_version.parse(latest_version or '')
     if cur_version >= latest_version:
         return None
     patch_upgrade = latest_version.major == cur_version.major and latest_version.minor == cur_version.minor

trains_agent/helper/console.py

@@ -22,6 +22,18 @@ def print_text(text, newline=True):
         sys.stdout.write(data)
 
 
+def decode_binary_lines(binary_lines, encoding='utf-8'):
+    # decode per line, if we failed decoding skip the line
+    lines = []
+    for b in binary_lines:
+        try:
+            l = b.decode(encoding=encoding, errors='replace').replace('\r', '\n')
+        except:
+            l = ''
+        lines.append(l + '\n' if l and l[-1] != '\n' else l)
+    return lines
+
+
 def ensure_text(s, encoding='utf-8', errors='strict'):
     """Coerce *s* to six.text_type.
     For Python 2:

trains_agent/helper/package/base.py

@@ -16,6 +16,8 @@ class PackageManager(object):
     """
 
     _selected_manager = None
+    _cwd = None
+    _pip_version = None
 
     @abc.abstractproperty
     def bin(self):
@@ -64,7 +66,7 @@ class PackageManager(object):
         pass
 
     def upgrade_pip(self):
-        return self._install("pip", "--upgrade")
+        return self._install("pip"+self.get_pip_version(), "--upgrade")
 
     def get_python_command(self, extra=()):
         # type: (...) -> Executable
@@ -97,11 +99,42 @@ class PackageManager(object):
         # this is helpful when we want out of context requirement installations
         PackageManager._selected_manager = self
 
+    @property
+    def cwd(self):
+        return self._cwd
+
+    @cwd.setter
+    def cwd(self, value):
+        self._cwd = value
+
     @classmethod
-    def out_of_scope_install_package(cls, package_name):
+    def out_of_scope_install_package(cls, package_name, *args):
         if PackageManager._selected_manager is not None:
             try:
-                return PackageManager._selected_manager._install(package_name)
+                return PackageManager._selected_manager._install(package_name, *args)
             except Exception:
                 pass
-        return
+        return
+
+    @classmethod
+    def out_of_scope_freeze(cls):
+        if PackageManager._selected_manager is not None:
+            try:
+                return PackageManager._selected_manager.freeze()
+            except Exception:
+                pass
+        return []
+
+    @classmethod
+    def set_pip_version(cls, version):
+        if not version:
+            return
+        version = version.replace(' ', '')
+        if ('=' in version) or ('~' in version) or ('<' in version) or ('>' in version):
+            cls._pip_version = version
+        else:
+            cls._pip_version = "=="+version
+
+    @classmethod
+    def get_pip_version(cls):
+        return cls._pip_version or ''

trains_agent/helper/package/conda_api.py

@@ -14,7 +14,7 @@ import yaml
 from time import time
 from attr import attrs, attrib, Factory
 from pathlib2 import Path
-from semantic_version import Version
+from packaging import version as packaging_version
 from requirements import parse
 from requirements.requirement import Requirement
 
@@ -59,7 +59,7 @@ class CondaAPI(PackageManager):
     A programmatic interface for controlling conda
     """
 
-    MINIMUM_VERSION = Version("4.3.30", partial=True)
+    MINIMUM_VERSION = packaging_version.parse("4.3.30")
 
     def __init__(self, session, path, python, requirements_manager):
         # type: (Session, PathLike, float, RequirementsManager) -> None
@@ -93,7 +93,7 @@ class CondaAPI(PackageManager):
                 )
             )
         self.conda_version = self.get_conda_version(output)
-        if Version(self.conda_version, partial=True) < self.MINIMUM_VERSION:
+        if packaging_version.parse(self.conda_version) < self.MINIMUM_VERSION:
             raise CommandFailedError(
                 "conda version '{}' is smaller than minimum supported conda version '{}'".format(
                     self.conda_version, self.MINIMUM_VERSION
@@ -112,7 +112,7 @@ class CondaAPI(PackageManager):
         return self.pip.bin
 
     def upgrade_pip(self):
-        return self.pip.upgrade_pip()
+        return self._install("pip" + self.pip.get_pip_version())
 
     def create(self):
         """
@@ -227,20 +227,20 @@ class CondaAPI(PackageManager):
             self.pip.install_from_file(reqs)
 
     def freeze(self):
-        # result = yaml.load(
-        #     self._run_command((self.conda, "env", "export", "-p", self.path), raw=True)
-        # )
-        # for key in "name", "prefix":
-        #     result.pop(key, None)
-        # freeze = {"conda": result}
-        # try:
-        #     freeze["pip"] = result["dependencies"][-1]["pip"]
-        # except (TypeError, KeyError):
-        #     freeze["pip"] = []
-        # else:
-        #     del result["dependencies"][-1]
-        # return freeze
-        return self.pip.freeze()
+        requirements = self.pip.freeze()
+        try:
+            conda_packages = json.loads(self._run_command((self.conda, "list", "--json", "-p", self.path), raw=True))
+            conda_packages_txt = []
+            requirements_pip = [r.split('==')[0].strip().lower() for r in requirements['pip']]
+            for pkg in conda_packages:
+                # skip if this is a pypi package or it is not a python package at all
+                if pkg['channel'] == 'pypi' or pkg['name'].lower() not in requirements_pip:
+                    continue
+                conda_packages_txt.append('{0}{1}{2}'.format(pkg['name'], '==', pkg['version']))
+            requirements['conda'] = conda_packages_txt
+        except:
+            pass
+        return requirements
 
     def load_requirements(self, requirements):
         # create new environment file
@@ -249,6 +249,8 @@ class CondaAPI(PackageManager):
         reqs = []
         if isinstance(requirements['pip'], six.string_types):
             requirements['pip'] = requirements['pip'].split('\n')
+        if isinstance(requirements.get('conda'), six.string_types):
+            requirements['conda'] = requirements['conda'].split('\n')
         has_torch = False
         has_matplotlib = False
         try:
@@ -256,35 +258,86 @@ class CondaAPI(PackageManager):
         except:
             cuda_version = 0
 
-        for r in requirements['pip']:
-            marker = list(parse(r))
-            if marker:
-                m = MarkerRequirement(marker[0])
-                if m.req.name.lower() == 'matplotlib':
-                    has_matplotlib = True
-                elif m.req.name.lower().startswith('torch'):
-                    has_torch = True
+        # notice 'conda' entry with empty string is a valid conda requirements list, it means pip only
+        # this should happen if experiment was executed on non-conda machine or old trains client
+        conda_supported_req = requirements['pip'] if requirements.get('conda', None) is None else requirements['conda']
+        conda_supported_req_names = []
+        for r in conda_supported_req:
+            try:
+                marker = list(parse(r))
+            except:
+                marker = None
+            if not marker:
+                continue
 
-                if m.req.name.lower() in ('torch', 'pytorch'):
-                    has_torch = True
-                    m.req.name = 'pytorch'
+            m = MarkerRequirement(marker[0])
+            conda_supported_req_names.append(m.name.lower())
+            if m.req.name.lower() == 'matplotlib':
+                has_matplotlib = True
+            elif m.req.name.lower().startswith('torch'):
+                has_torch = True
 
-                if m.req.name.lower() in ('tensorflow_gpu', 'tensorflow-gpu', 'tensorflow'):
-                    has_torch = True
-                    m.req.name = 'tensorflow-gpu' if cuda_version > 0 else 'tensorflow'
+            if m.req.name.lower() in ('torch', 'pytorch'):
+                has_torch = True
+                m.req.name = 'pytorch'
+
+            if m.req.name.lower() in ('tensorflow_gpu', 'tensorflow-gpu', 'tensorflow'):
+                has_torch = True
+                m.req.name = 'tensorflow-gpu' if cuda_version > 0 else 'tensorflow'
+
+            reqs.append(m)
 
-                reqs.append(m)
         pip_requirements = []
+        # if we have a conda list, the rest should be installed with pip,
+        if requirements.get('conda', None) is not None:
+            for r in requirements['pip']:
+                try:
+                    marker = list(parse(r))
+                except:
+                    marker = None
+                if not marker:
+                    continue
+
+                m = MarkerRequirement(marker[0])
+                m_name = m.name.lower()
+                if m_name in conda_supported_req_names:
+                    # this package is in the conda list,
+                    # make sure that if we changed version and we match it in conda
+                    conda_supported_req_names.remove(m_name)
+                    for cr in reqs:
+                        if m_name == cr.name.lower():
+                            # match versions
+                            cr.specs = m.specs
+                            break
+                else:
+                    # not in conda, it is a pip package
+                    pip_requirements.append(m)
+                    if m_name == 'matplotlib':
+                        has_matplotlib = True
+
+            # remove any leftover conda packages (they were removed from the pip list)
+            if conda_supported_req_names:
+                reqs = [r for r in reqs if r.name.lower() not in conda_supported_req_names]
 
         # Conda requirements Hacks:
         if has_matplotlib:
             reqs.append(MarkerRequirement(Requirement.parse('graphviz')))
+            reqs.append(MarkerRequirement(Requirement.parse('python-graphviz')))
             reqs.append(MarkerRequirement(Requirement.parse('kiwisolver')))
         if has_torch and cuda_version == 0:
             reqs.append(MarkerRequirement(Requirement.parse('cpuonly')))
 
+        # conform conda packages (version/name)
+        for r in reqs:
+            # remove .post from version numbers, it fails ~= version, and change == to ~=
+            if r.specs and r.specs[0]:
+                r.specs = [(r.specs[0][0].replace('==', '~='), r.specs[0][1].split('.post')[0])]
+            # conda always likes "-" not "_"
+            r.req.name = r.req.name.replace('_', '-')
+
         while reqs:
-            conda_env['dependencies'] = [r.tostr().replace('==', '=') for r in reqs]
+            # notice, we give conda more freedom in version selection, to help it choose best combination
+            conda_env['dependencies'] = [r.tostr() for r in reqs]
             with self.temp_file("conda_env", yaml.dump(conda_env), suffix=".yml") as name:
                 print('Conda: Trying to install requirements:\n{}'.format(conda_env['dependencies']))
                 result = self._run_command(
@@ -297,7 +350,7 @@ class CondaAPI(PackageManager):
 
             solved = False
             for bad_r in bad_req:
-                name = bad_r.split('[')[0].split('=')[0]
+                name = bad_r.split('[')[0].split('=')[0].split('~')[0].split('<')[0].split('>')[0]
                 # look for name in requirements
                 for r in reqs:
                     if r.name.lower() == name.lower():
@@ -338,7 +391,7 @@ class CondaAPI(PackageManager):
                 if len(empty_lines) >= 2:
                     deps = error_lines[empty_lines[0]+1:empty_lines[1]]
                     try:
-                        return yaml.load('\n'.join(deps))
+                        return yaml.load('\n'.join(deps), Loader=yaml.SafeLoader)
                     except:
                         return None
         return None
@@ -412,4 +465,4 @@ class PackageNotFoundError(CondaException):
                           as a singleton YAML list.
     """
 
-    pkg = attrib(default="", converter=lambda val: yaml.load(val)[0].replace(" ", ""))
+    pkg = attrib(default="", converter=lambda val: yaml.load(val, Loader=yaml.SafeLoader)[0].replace(" ", ""))

trains_agent/helper/package/cython_req.py

@@ -6,14 +6,14 @@ from .requirements import SimpleSubstitution
 
 class CythonRequirement(SimpleSubstitution):
 
-    name = "cython"
+    name = ("cython", "numpy", )
 
     def __init__(self, *args, **kwargs):
         super(CythonRequirement, self).__init__(*args, **kwargs)
 
     def match(self, req):
         # match both Cython & cython
-        return self.name == req.name.lower()
+        return req.name and req.name.lower() in self.name
 
     def replace(self, req):
         """

trains_agent/helper/package/external_req.py

@@ -0,0 +1,60 @@
+from collections import OrderedDict
+from typing import Text
+
+from .base import PackageManager
+from .requirements import SimpleSubstitution
+
+
+class ExternalRequirements(SimpleSubstitution):
+
+    name = "external_link"
+
+    def __init__(self, *args, **kwargs):
+        super(ExternalRequirements, self).__init__(*args, **kwargs)
+        self.post_install_req = []
+        self.post_install_req_lookup = OrderedDict()
+
+    def match(self, req):
+        # match both editable or code or unparsed
+        if not (not req.name or req.req and (req.req.editable or req.req.vcs)):
+            return False
+        if not req.req or not req.req.line or not req.req.line.strip() or req.req.line.strip().startswith('#'):
+            return False
+        return True
+
+    def post_install(self):
+        post_install_req = self.post_install_req
+        self.post_install_req = []
+        for req in post_install_req:
+            try:
+                freeze_base = PackageManager.out_of_scope_freeze() or ''
+            except:
+                freeze_base = ''
+            PackageManager.out_of_scope_install_package(req.tostr(markers=False), "--no-deps")
+            try:
+                freeze_post = PackageManager.out_of_scope_freeze() or ''
+                package_name = list(set(freeze_post['pip']) - set(freeze_base['pip']))
+                if package_name and package_name[0] not in self.post_install_req_lookup:
+                    self.post_install_req_lookup[package_name[0]] = req.req.line
+            except:
+                pass
+            PackageManager.out_of_scope_install_package(req.tostr(markers=False), "--ignore-installed")
+
+    def replace(self, req):
+        """
+        Replace a requirement
+        :raises: ValueError if version is pre-release
+        """
+        # Store in post req install, and return nothing
+        self.post_install_req.append(req)
+        # mark skip package, we will install it in post install hook
+        return Text('')
+
+    def replace_back(self, list_of_requirements):
+        if 'pip' in list_of_requirements:
+            original_requirements = list_of_requirements['pip']
+            list_of_requirements['pip'] = [r for r in original_requirements
+                                           if r not in self.post_install_req_lookup]
+            list_of_requirements['pip'] += [self.post_install_req_lookup.get(r, '')
+                                            for r in self.post_install_req_lookup.keys() if r in original_requirements]
+        return list_of_requirements

trains_agent/helper/package/horovod_req.py

@@ -14,7 +14,7 @@ class HorovodRequirement(SimpleSubstitution):
 
     def match(self, req):
         # match both horovod
-        return self.name == req.name.lower()
+        return req.name and self.name == req.name.lower()
 
     def post_install(self):
         if self.post_install_req:

trains_agent/helper/package/pip_api/system.py

@@ -29,13 +29,13 @@ class SystemPip(PackageManager):
         pass
 
     def install_from_file(self, path):
-        self.run_with_env(('install', '-r', path) + self.install_flags())
+        self.run_with_env(('install', '-r', path) + self.install_flags(), cwd=self.cwd)
 
     def install_packages(self, *packages):
         self._install(*(packages + self.install_flags()))
 
     def _install(self, *args):
-        self.run_with_env(('install',) + args)
+        self.run_with_env(('install',) + args, cwd=self.cwd)
 
     def uninstall_packages(self, *packages):
         self.run_with_env(('uninstall', '-y') + packages)
@@ -82,7 +82,7 @@ class SystemPip(PackageManager):
         return (command.get_output if output else command.check_call)(stdin=DEVNULL, **kwargs)
 
     def _make_command(self, command):
-        return Argv(self.bin, '-m', 'pip', *command)
+        return Argv(self.bin, '-m', 'pip', '--disable-pip-version-check', *command)
 
     def install_flags(self):
         if self.indices_args is None:

trains_agent/helper/package/pip_api/venv.py

@@ -33,7 +33,7 @@ class VirtualenvPip(SystemPip, PackageManager):
         self.python = python
 
     def _make_command(self, command):
-        return self.session.command(self.bin, "-m", "pip", *command)
+        return self.session.command(self.bin, "-m", "pip", "--disable-pip-version-check", *command)
 
     def load_requirements(self, requirements):
         if isinstance(requirements, dict) and requirements.get("pip"):

trains_agent/helper/package/poetry_api.py

@@ -1,8 +1,11 @@
+from copy import deepcopy
 from functools import wraps
 
 import attr
+import sys
+import os
 from pathlib2 import Path
-from trains_agent.helper.process import Argv, DEVNULL
+from trains_agent.helper.process import Argv, DEVNULL, check_if_command_exists
 from trains_agent.session import Session, POETRY
 
 
@@ -35,10 +38,12 @@ def prop_guard(prop, log_prop=None):
 
 class PoetryConfig:
 
-    def __init__(self, session):
-        # type: (Session) -> ()
+    def __init__(self, session, interpreter=None):
+        # type: (Session, str) -> ()
         self.session = session
         self._log = session.get_logger(__name__)
+        self._python = interpreter or sys.executable
+        self._initialized = False
 
     @property
     def log(self):
@@ -53,7 +58,20 @@ class PoetryConfig:
     def run(self, *args, **kwargs):
         func = kwargs.pop("func", Argv.get_output)
         kwargs.setdefault("stdin", DEVNULL)
-        argv = Argv("poetry", "-n", *args)
+        kwargs['env'] = deepcopy(os.environ)
+        if 'VIRTUAL_ENV' in kwargs['env'] or 'CONDA_PREFIX' in kwargs['env']:
+            kwargs['env'].pop('VIRTUAL_ENV', None)
+            kwargs['env'].pop('CONDA_PREFIX', None)
+            kwargs['env'].pop('PYTHONPATH', None)
+            if hasattr(sys, "real_prefix") and hasattr(sys, "base_prefix"):
+                path = ':'+kwargs['env']['PATH']
+                path = path.replace(':'+sys.base_prefix, ':'+sys.real_prefix, 1)
+                kwargs['env']['PATH'] = path
+
+        if check_if_command_exists("poetry"):
+            argv = Argv("poetry", *args)
+        else:
+            argv = Argv(self._python, "-m", "poetry", *args)
         self.log.debug("running: %s", argv)
         return func(argv, **kwargs)
 
@@ -61,10 +79,16 @@ class PoetryConfig:
         return self.run("config", *args, **kwargs)
 
     @_guard_enabled
-    def initialize(self):
-        self._config("settings.virtualenvs.in-project",  "true")
-        # self._config("repositories.{}".format(self.REPO_NAME), PYTHON_INDEX)
-        # self._config("http-basic.{}".format(self.REPO_NAME), *PYTHON_INDEX_CREDENTIALS)
+    def initialize(self, cwd=None):
+        if not self._initialized:
+            self._initialized = True
+            try:
+                self._config("--local", "virtualenvs.in-project",  "true", cwd=cwd)
+                # self._config("repositories.{}".format(self.REPO_NAME), PYTHON_INDEX)
+                # self._config("http-basic.{}".format(self.REPO_NAME), *PYTHON_INDEX_CREDENTIALS)
+            except Exception as ex:
+                print("Exception: {}\nError: Failed configuring Poetry virtualenvs.in-project".format(ex))
+                raise
 
     def get_api(self, path):
         # type: (Path) -> PoetryAPI
@@ -81,7 +105,7 @@ class PoetryAPI(object):
     def install(self):
         # type: () -> bool
         if self.enabled:
-            self.config.run("install", cwd=str(self.path), func=Argv.check_call)
+            self.config.run("install", "-n", cwd=str(self.path), func=Argv.check_call)
             return True
         return False
 
@@ -92,10 +116,15 @@ class PoetryAPI(object):
         )
 
     def freeze(self):
-        return {"poetry": self.config.run("show", cwd=str(self.path)).splitlines()}
+        lines = self.config.run("show", cwd=str(self.path)).splitlines()
+        lines = [[p for p in line.split(' ') if p] for line in lines]
+        return {"pip": [parts[0]+'=='+parts[1]+' # '+' '.join(parts[2:]) for parts in lines]}
 
     def get_python_command(self, extra):
-        return Argv("poetry", "run", "python", *extra)
+        if check_if_command_exists("poetry"):
+            return Argv("poetry", "run", "python", *extra)
+        else:
+            return Argv(self.config._python, "-m", "poetry", "run", "python", *extra)
 
     def upgrade_pip(self, *args, **kwargs):
         pass

trains_agent/helper/package/pytorch.py

@@ -10,7 +10,8 @@ from typing import Text
 
 import attr
 import requests
-from semantic_version import Version, Spec
+from packaging import version as packaging_version
+from packaging.specifiers import SpecifierSet
 
 import six
 from .requirements import SimpleSubstitution, FatalSpecsResolutionError
@@ -155,10 +156,16 @@ class PytorchRequirement(SimpleSubstitution):
         self.os = os_name or self.get_platform()
         self.cuda = "cuda{}".format(self.cuda_version).lower()
         self.python_version_string = str(self.config["agent.default_python"])
-        self.python_semantic_version = Version.coerce(
-            self.python_version_string, partial=True
-        )
-        self.python = "python{}.{}".format(self.python_semantic_version.major, self.python_semantic_version.minor)
+        self.python_major_minor_str = '.'.join(packaging_version.parse(
+            self.python_version_string).base_version.split('.')[:2])
+        if '.' not in self.python_major_minor_str:
+            raise PytorchResolutionError(
+                "invalid python version {!r} defined in configuration file, key 'agent.default_python': "
+                "must have both major and minor parts of the version (for example: '3.7')".format(
+                    self.python_version_string
+                )
+            )
+        self.python = "python{}".format(self.python_major_minor_str)
 
         self.exceptions = [
             PytorchResolutionError(message)
@@ -188,9 +195,7 @@ class PytorchRequirement(SimpleSubstitution):
         """
         Make sure python version has both major and minor versions as required for choosing pytorch wheel
         """
-        if self.is_pip and not (
-            self.python_semantic_version.major and self.python_semantic_version.minor
-        ):
+        if self.is_pip and not self.python_major_minor_str:
             raise PytorchResolutionError(
                 "invalid python version {!r} defined in configuration file, key 'agent.default_python': "
                 "must have both major and minor parts of the version (for example: '3.7')".format(
@@ -215,8 +220,10 @@ class PytorchRequirement(SimpleSubstitution):
         links_parser = LinksHTMLParser()
         links_parser.feed(requests.get(torch_url, timeout=10).text)
         platform_wheel = "win" if self.get_platform() == "windows" else self.get_platform()
-        py_ver = "{0.major}{0.minor}".format(self.python_semantic_version)
+        py_ver = self.python_major_minor_str.replace('.', '')
         url = None
+        spec = SpecifierSet(req.format_specs())
+        last_v = None
         # search for our package
         for l in links_parser.links:
             parts = l.split('/')[-1].split('-')
@@ -225,21 +232,23 @@ class PytorchRequirement(SimpleSubstitution):
             if parts[0] != req.name:
                 continue
             # version (ignore +cpu +cu92 etc. + is %2B in the file link)
-            if parts[1].split('%')[0].split('+')[0] != req.specs[0][1]:
+            # version ignore .postX suffix (treat as regular version)
+            try:
+                v = packaging_version.parse(parts[1].split('%')[0].split('+')[0])
+            except Exception:
+                continue
+            if v not in spec or (last_v and last_v > v):
                 continue
             if not parts[2].endswith(py_ver):
                 continue
             if platform_wheel not in parts[4]:
                 continue
             url = '/'.join(torch_url.split('/')[:-1] + l.split('/'))
-            break
+            last_v = v
 
         return url
 
     def get_url_for_platform(self, req):
-        assert self.package_manager == "pip"
-        assert self.os != "mac"
-
         # check if package is already installed with system packages
         try:
             if self.config.get("agent.package_manager.system_site_packages"):
@@ -254,7 +263,8 @@ class PytorchRequirement(SimpleSubstitution):
             pass
 
         # make sure we have a specific version to retrieve
-        assert req.specs
+        if not req.specs:
+            req.specs = [('>', '0')]
 
         try:
             req.specs[0] = (req.specs[0][0], req.specs[0][1].split('+')[0])
@@ -282,7 +292,7 @@ class PytorchRequirement(SimpleSubstitution):
         if not url:
             url = PytorchWheel(
                 torch_version=fix_version(version),
-                python="{0.major}{0.minor}".format(self.python_semantic_version),
+                python=self.python_major_minor_str.replace('.', ''),
                 os_name=self.os,
                 cuda_version=self.cuda_version,
             ).make_url()
@@ -296,13 +306,13 @@ class PytorchRequirement(SimpleSubstitution):
     @staticmethod
     def match_version(req, options):
         versioned_options = sorted(
-            ((Version(fix_version(key)), value) for key, value in options.items()),
+            ((packaging_version.parse(fix_version(key)), value) for key, value in options.items()),
             key=itemgetter(0),
             reverse=True,
         )
         req.specs = [(op, fix_version(version)) for op, version in req.specs]
         if req.specs:
-            specs = Spec(req.format_specs())
+            specs = SpecifierSet(req.format_specs())
         else:
             specs = None
         try:

trains_agent/helper/package/requirements.py

@@ -8,9 +8,9 @@ from copy import deepcopy
 from itertools import chain, starmap
 from operator import itemgetter
 from os import path
-from typing import Text, List, Type, Optional, Tuple
+from typing import Text, List, Type, Optional, Tuple, Dict
 
-import semantic_version
+from packaging import version as packaging_version
 from pathlib2 import Path
 from pyhocon import ConfigTree
 from requirements import parse
@@ -48,7 +48,7 @@ class MarkerRequirement(object):
 
     def tostr(self, markers=True):
         if not self.uri:
-            parts = [self.name]
+            parts = [self.name or self.line]
 
             if self.extras:
                 parts.append('[{0}]'.format(','.join(sorted(self.extras))))
@@ -177,13 +177,20 @@ class SimpleSubstitution(RequirementSubstitution):
 
         if req.specs:
             _, version_number = req.specs[0]
-            assert semantic_version.Version(version_number, partial=True)
+            assert packaging_version.parse(version_number)
         else:
             version_number = self.get_pip_version(self.name)
 
         req.specs = [('==', version_number + self.suffix)]
         return Text(req)
 
+    def replace_back(self, list_of_requirements):  # type: (Dict) -> Dict
+        """
+        :param list_of_requirements: {'pip': ['a==1.0', ]}
+        :return: {'pip': ['a==1.0', ]}
+        """
+        return list_of_requirements
+
 
 @six.add_metaclass(ABCMeta)
 class CudaSensitiveSubstitution(SimpleSubstitution):
@@ -235,15 +242,17 @@ class RequirementsManager(object):
         return None
 
     def replace(self, requirements):  # type: (Text) -> Text
+        def safe_parse(req_str):
+            try:
+                return next(parse(req_str))
+            except Exception as ex:
+                return Requirement(req_str)
+
         parsed_requirements = tuple(
             map(
                 MarkerRequirement,
-                filter(
-                    None,
-                    parse(requirements)
-                    if isinstance(requirements, six.text_type)
-                    else (next(parse(line), None) for line in requirements)
-                )
+                [safe_parse(line) for line in (requirements.splitlines()
+                                               if isinstance(requirements, six.text_type) else requirements)]
             )
         )
         if not parsed_requirements:
@@ -258,7 +267,7 @@ class RequirementsManager(object):
                 warning('could not resolve python wheel replacement for {}'.format(req))
                 raise
             except Exception:
-                warning('could not resolve python wheel replacement for {}, '
+                warning('could not resolve python wheel replacement for \"{}\", '
                         'using original requirements line: {}'.format(req, i))
                 return None
 
@@ -280,6 +289,14 @@ class RequirementsManager(object):
             except Exception as ex:
                 print('RequirementsManager handler {} raised exception: {}'.format(h, ex))
 
+    def replace_back(self, requirements):
+        for h in self.handlers:
+            try:
+                requirements = h.replace_back(requirements)
+            except Exception:
+                pass
+        return requirements
+
     @staticmethod
     def get_cuda_version(config):  # type: (ConfigTree) -> (Text, Text)
         # we assume os.environ already updated the config['agent.cuda_version'] & config['agent.cudnn_version']

trains_agent/helper/repo.py

@@ -42,7 +42,9 @@ class VcsFactory(object):
         :param location: (desired) clone location
         """
         url = execution_info.repository
-        is_git = url.endswith(cls.GIT_SUFFIX)
+        # We only support git, hg is deprecated
+        is_git = True
+        # is_git = url.endswith(cls.GIT_SUFFIX)
         vcs_cls = Git if is_git else Hg
         revision = (
             execution_info.version_num
@@ -454,7 +456,7 @@ class Git(VCS):
         )
 
     def pull(self):
-        self.call("fetch", "--all", cwd=self.location)
+        self.call("fetch", "--all", "--recurse-submodules", cwd=self.location)
 
     info_commands = dict(
         url=Argv(executable_name, "ls-remote", "--get-url", "origin"),

trains_agent/helper/singleton.py

@@ -4,11 +4,12 @@ from time import sleep
 from glob import glob
 from tempfile import gettempdir, NamedTemporaryFile
 
+from trains_agent.definitions import ENV_K8S_HOST_MOUNT
 from trains_agent.helper.base import warning
 
 
 class Singleton(object):
-    prefix = 'trainsagent'
+    prefix = '.trainsagent'
     sep = '_'
     ext = '.tmp'
     worker_id = None
@@ -19,7 +20,7 @@ class Singleton(object):
     _lock_timeout = 10
 
     @classmethod
-    def register_instance(cls, unique_worker_id=None, worker_name=None):
+    def register_instance(cls, unique_worker_id=None, worker_name=None, api_client=None):
         """
         # Exit the process if another instance of us is using the same worker_id
 
@@ -28,7 +29,7 @@ class Singleton(object):
         :return: (str worker_id, int slot_number) Return None value on instance already running
         """
         # try to lock file
-        lock_file = os.path.join(gettempdir(), cls._lock_file_name)
+        lock_file = os.path.join(cls._get_temp_folder(), cls._lock_file_name)
         timeout = 0
         while os.path.exists(lock_file):
             if timeout > cls._lock_timeout:
@@ -46,7 +47,8 @@ class Singleton(object):
             f.write(bytes(os.getpid()))
             f.flush()
             try:
-                ret = cls._register_instance(unique_worker_id=unique_worker_id, worker_name=worker_name)
+                ret = cls._register_instance(unique_worker_id=unique_worker_id, worker_name=worker_name,
+                                             api_client=api_client)
             except:
                 ret = None, None
 
@@ -58,12 +60,12 @@ class Singleton(object):
         return ret
 
     @classmethod
-    def _register_instance(cls, unique_worker_id=None, worker_name=None):
+    def _register_instance(cls, unique_worker_id=None, worker_name=None, api_client=None):
         if cls.worker_id:
             return cls.worker_id, cls.instance_slot
         # make sure we have a unique name
         instance_num = 0
-        temp_folder = gettempdir()
+        temp_folder = cls._get_temp_folder()
         files = glob(os.path.join(temp_folder, cls.prefix + cls.sep + '*' + cls.ext))
         slots = {}
         for file in files:
@@ -73,8 +75,24 @@ class Singleton(object):
             except Exception:
                 # something is wrong, use non existing pid and delete the file
                 pid = -1
+
+            uid, slot = None, None
+            try:
+                with open(file, 'r') as f:
+                    uid, slot = str(f.read()).split('\n')
+                    slot = int(slot)
+            except Exception:
+                pass
+
+            worker = None
+            if api_client and os.environ.get(ENV_K8S_HOST_MOUNT) and uid:
+                try:
+                    worker = [w for w in api_client.workers.get_all() if w.id == uid]
+                except Exception:
+                    worker = None
+
             # count active instances and delete dead files
-            if not psutil.pid_exists(pid):
+            if not worker and not psutil.pid_exists(pid):
                 # delete the file
                 try:
                     os.remove(os.path.join(file))
@@ -83,11 +101,7 @@ class Singleton(object):
                 continue
 
             instance_num += 1
-            try:
-                with open(file, 'r') as f:
-                    uid, slot = str(f.read()).split('\n')
-                    slot = int(slot)
-            except Exception:
+            if slot is None:
                 continue
 
             if uid == unique_worker_id:
@@ -110,10 +124,16 @@ class Singleton(object):
             unique_worker_id = worker_name + cls.worker_name_sep + str(cls.instance_slot)
 
         # create lock
-        cls._pid_file = NamedTemporaryFile(dir=gettempdir(), prefix=cls.prefix + cls.sep + str(os.getpid()) + cls.sep,
-                                           suffix=cls.ext)
+        cls._pid_file = NamedTemporaryFile(dir=cls._get_temp_folder(),
+                                           prefix=cls.prefix + cls.sep + str(os.getpid()) + cls.sep, suffix=cls.ext)
         cls._pid_file.write(('{}\n{}'.format(unique_worker_id, cls.instance_slot)).encode())
         cls._pid_file.flush()
         cls.worker_id = unique_worker_id
 
         return cls.worker_id, cls.instance_slot
+
+    @classmethod
+    def _get_temp_folder(cls):
+        if os.environ.get(ENV_K8S_HOST_MOUNT):
+            return os.environ.get(ENV_K8S_HOST_MOUNT).split(':')[-1]
+        return gettempdir()

trains_agent/interface/__init__.py

@@ -1,3 +1,4 @@
+import itertools
 from functools import partial
 from importlib import import_module
 import argparse
@@ -24,8 +25,16 @@ def get_parser():
     from .worker import COMMANDS
     subparsers = top_parser.add_subparsers(dest='command')
     for c in COMMANDS:
-        parser = subparsers.add_parser(name=c, help=COMMANDS[c]['help'])
-        for a in COMMANDS[c].get('args', {}).keys():
-            parser.add_argument(a, **COMMANDS[c]['args'][a])
+        parser = subparsers.add_parser(name=c, help=COMMANDS[c]["help"])
+        groups = itertools.groupby(
+            sorted(
+                COMMANDS[c].get("args", {}).items(), key=lambda x: x[1].get("group", "")
+            ),
+            key=lambda x: x[1].pop("group", ""),
+        )
+        for group_name, group in groups:
+            p = parser if not group_name else parser.add_argument_group(group_name)
+            for key, value in group:
+                p.add_argument(key, **value)
 
     return top_parser

trains_agent/interface/worker.py

@@ -37,21 +37,29 @@ DAEMON_ARGS = dict({
         'help': 'Pipe full log to stdout/stderr, should not be used if running in background',
         'action': 'store_true',
     },
-    '--gpus': {
-        'help': 'Specify active GPUs for the daemon to use (docker / virtual environment), '
-                'Equivalent to setting NVIDIA_VISIBLE_DEVICES '
-                'Examples: --gpus 0 or --gpu 0,1,2 or --gpus all',
-    },
-    '--cpu-only': {
-        'help': 'Disable GPU access for the daemon, only use CPU in either docker or virtual environment',
-        'action': 'store_true',
-    },
     '--docker': {
         'help': 'Run execution task inside a docker (v19.03 and above). Optional args <image> <arguments> or '
                 'specify default docker image in agent.default_docker.image / agent.default_docker.arguments'
                 'use --gpus/--cpu-only (or set NVIDIA_VISIBLE_DEVICES) to limit gpu visibility for docker',
         'nargs': '*',
         'default': False,
+        'group': 'Docker support',
+    },
+    '--gpus': {
+        'help': 'Specify active GPUs for the daemon to use (docker / virtual environment), '
+                'Equivalent to setting NVIDIA_VISIBLE_DEVICES '
+                'Examples: --gpus 0 or --gpu 0,1,2 or --gpus all',
+        'group': 'Docker support',
+    },
+    '--cpu-only': {
+        'help': 'Disable GPU access for the daemon, only use CPU in either docker or virtual environment',
+        'action': 'store_true',
+        'group': 'Docker support',
+    },
+    '--force-current-version': {
+        'help': 'Force trains-agent to use the current trains-agent version when running in the docker',
+        'action': 'store_true',
+        'group': 'Docker support',
     },
     '--queue': {
         'help': 'Queue ID(s)/Name(s) to pull tasks from (\'default\' queue)',

trains_agent/session.py

@@ -15,7 +15,7 @@ from pyhocon import ConfigFactory, HOCONConverter, ConfigTree
 from trains_agent.backend_api.session import Session as _Session, Request
 from trains_agent.backend_api.session.client import APIClient
 from trains_agent.backend_config.defs import LOCAL_CONFIG_FILE_OVERRIDE_VAR, LOCAL_CONFIG_FILES
-from trains_agent.definitions import ENVIRONMENT_CONFIG
+from trains_agent.definitions import ENVIRONMENT_CONFIG, ENV_TASK_EXECUTE_AS_USER
 from trains_agent.errors import APIError
 from trains_agent.helper.base import HOCONEncoder
 from trains_agent.helper.process import Argv
@@ -75,7 +75,8 @@ class Session(_Session):
         cpu_only = kwargs.get('cpu_only')
         if cpu_only:
             os.environ['CUDA_VISIBLE_DEVICES'] = os.environ['NVIDIA_VISIBLE_DEVICES'] = 'none'
-        if kwargs.get('gpus'):
+        if kwargs.get('gpus') and not os.environ.get('KUBERNETES_SERVICE_HOST') \
+                and not os.environ.get('KUBERNETES_PORT'):
             os.environ['CUDA_VISIBLE_DEVICES'] = os.environ['NVIDIA_VISIBLE_DEVICES'] = kwargs.get('gpus')
         if kwargs.get('only_load_config'):
             from trains_agent.backend_api.config import load
@@ -111,6 +112,17 @@ class Session(_Session):
         # override with environment variables
         # cuda_version & cudnn_version are overridden with os.environ here, and normalized in the next section
         for config_key, env_config in ENVIRONMENT_CONFIG.items():
+            # check if the propery is of a list:
+            if config_key.endswith('.0'):
+                if all(not i.get() for i in env_config.values()):
+                    continue
+                parent = config_key.partition('.0')[0]
+                if not self.config[parent]:
+                    self.config.put(parent, [])
+
+                self.config.put(parent, self.config[parent] + [ConfigTree((k, v.get()) for k, v in env_config.items())])
+                continue
+
             value = env_config.get()
             if not value:
                 continue
@@ -165,7 +177,11 @@ class Session(_Session):
         folder_keys = ('agent.venvs_dir', 'agent.vcs_cache.path',
                        'agent.pip_download_cache.path',
                        'agent.docker_pip_cache', 'agent.docker_apt_cache')
-        singleton_folders = ('agent.venvs_dir', 'agent.vcs_cache.path',)
+        singleton_folders = ('agent.venvs_dir', 'agent.vcs_cache.path', 'agent.docker_apt_cache')
+
+        if os.environ.get(ENV_TASK_EXECUTE_AS_USER):
+            folder_keys = tuple(list(folder_keys) + ['sdk.storage.cache.default_base_dir'])
+            singleton_folders = tuple(list(singleton_folders) + ['sdk.storage.cache.default_base_dir'])
 
         for key in folder_keys:
             folder_key = ConfigValue(self.config, key)

trains_agent/version.py

@@ -1 +1 @@
-__version__ = '0.13.0'
+__version__ = '0.13.3'