feat: add delete command, make cancel abort for k8s (#1156)

Author: azzhipa

torchx/cli/cmd_delete.py

@@ -0,0 +1,30 @@
+#!/usr/bin/env python3
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+# pyre-strict
+
+import argparse
+import logging
+
+from torchx.cli.cmd_base import SubCommand
+from torchx.runner import get_runner
+
+logger: logging.Logger = logging.getLogger(__name__)
+
+
+class CmdDelete(SubCommand):
+    def add_arguments(self, subparser: argparse.ArgumentParser) -> None:
+        subparser.add_argument(
+            "app_handle",
+            type=str,
+            help="torchx app handle (e.g. local://session-name/app-id)",
+        )
+
+    def run(self, args: argparse.Namespace) -> None:
+        app_handle = args.app_handle
+        runner = get_runner()
+        runner.delete(app_handle)

torchx/cli/main.py

@@ -16,6 +16,7 @@
 from torchx.cli.cmd_base import SubCommand
 from torchx.cli.cmd_cancel import CmdCancel
 from torchx.cli.cmd_configure import CmdConfigure
+from torchx.cli.cmd_delete import CmdDelete
 from torchx.cli.cmd_describe import CmdDescribe
 from torchx.cli.cmd_list import CmdList
 from torchx.cli.cmd_log import CmdLog
@@ -37,6 +38,7 @@ def get_default_sub_cmds() -> Dict[str, SubCommand]:
         "builtins": CmdBuiltins(),
         "cancel": CmdCancel(),
         "configure": CmdConfigure(),
+        "delete": CmdDelete(),
         "describe": CmdDescribe(),
         "list": CmdList(),
         "log": CmdLog(),

torchx/cli/test/cmd_delete_test.py

@@ -0,0 +1,28 @@
+#!/usr/bin/env python3
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+# pyre-strict
+
+import argparse
+import unittest
+from unittest.mock import MagicMock, patch
+
+from torchx.cli.cmd_delete import CmdDelete
+
+
+class CmdDeleteTest(unittest.TestCase):
+    @patch("torchx.runner.api.Runner.delete")
+    def test_run(self, delete: MagicMock) -> None:
+        parser = argparse.ArgumentParser()
+        cmd_delete = CmdDelete()
+        cmd_delete.add_arguments(parser)
+
+        args = parser.parse_args(["foo://session/id"])
+        cmd_delete.run(args)
+
+        self.assertEqual(delete.call_count, 1)
+        delete.assert_called_with("foo://session/id")

torchx/runner/api.py

@@ -579,6 +579,16 @@ def cancel(self, app_handle: AppHandle) -> None:
             if status is not None and not status.is_terminal():
                 scheduler.cancel(app_id)
 
+    def delete(self, app_handle: AppHandle) -> None:
+        """
+        Deletes the application from the scheduler.
+        """
+        scheduler, scheduler_backend, app_id = self._scheduler_app_id(app_handle)
+        with log_event("delete", scheduler_backend, app_id):
+            status = self.status(app_handle)
+            if status is not None and not status.is_terminal():
+                scheduler.delete(app_id)
+
     def stop(self, app_handle: AppHandle) -> None:
         """
         See method ``cancel``.

torchx/schedulers/api.py

@@ -264,6 +264,13 @@ def cancel(self, app_id: str) -> None:
             # do nothing if the app does not exist
             return
 
+    def delete(self, app_id: str) -> None:
+        """
+        Deletes the application. By default, this calls cancel.
+        Schedulers can override to provide delete-specific behavior.
+        """
+        self.cancel(app_id)
+
     def log_iter(
         self,
         app_id: str,

torchx/schedulers/kubernetes_scheduler.py

@@ -622,6 +622,16 @@ class KubernetesScheduler(
         $ torchx status kubernetes://torchx_user/1234
         ...
 
+    **Cancellation**
+
+    Canceling a job aborts it while preserving the job spec for inspection
+    and cloning via kubectl apply. Use the delete command to remove the job entirely:
+
+    .. code-block:: bash
+
+        $ torchx cancel kubernetes://namespace/jobname  # abort, preserves spec
+        $ torchx delete kubernetes://namespace/jobname  # delete completely
+
     **Config Options**
 
     .. runopts::
@@ -818,6 +828,33 @@ def _validate(self, app: AppDef, scheduler: str, cfg: KubernetesOpts) -> None:
         pass
 
     def _cancel_existing(self, app_id: str) -> None:
+        """
+        Abort a Volcano job while preserving the spec for inspection.
+        """
+        namespace, name = app_id.split(":")
+        vcjob = self._custom_objects_api().get_namespaced_custom_object(
+            group="batch.volcano.sh",
+            version="v1alpha1",
+            namespace=namespace,
+            plural="jobs",
+            name=name,
+        )
+        vcjob["status"]["state"]["phase"] = "Aborted"
+        self._custom_objects_api().replace_namespaced_custom_object_status(
+            group="batch.volcano.sh",
+            version="v1alpha1",
+            namespace=namespace,
+            plural="jobs",
+            name=name,
+            body=vcjob,
+        )
+
+    def delete(self, app_id: str) -> None:
+        """
+        Delete a Volcano job completely from the cluster.
+        """
+        if not self.exists(app_id):
+            return
         namespace, name = app_id.split(":")
         self._custom_objects_api().delete_namespaced_custom_object(
             group="batch.volcano.sh",

torchx/schedulers/test/kubernetes_scheduler_test.py

@@ -800,11 +800,26 @@ def test_runopts(self) -> None:
             },
         )
 
-    @patch("kubernetes.client.CustomObjectsApi.delete_namespaced_custom_object")
-    def test_cancel_existing(self, delete_namespaced_custom_object: MagicMock) -> None:
+    @patch("kubernetes.client.CustomObjectsApi.get_namespaced_custom_object")
+    @patch("kubernetes.client.CustomObjectsApi.replace_namespaced_custom_object_status")
+    def test_cancel_existing(
+        self,
+        replace_namespaced_custom_object_status: MagicMock,
+        get_namespaced_custom_object: MagicMock,
+    ) -> None:
         scheduler = create_scheduler("test")
+        get_namespaced_custom_object.return_value = {
+            "status": {"state": {"phase": "Running"}}
+        }
         scheduler._cancel_existing("testnamespace:testjob")
-        call = delete_namespaced_custom_object.call_args
+        get_namespaced_custom_object.assert_called_once_with(
+            group="batch.volcano.sh",
+            version="v1alpha1",
+            namespace="testnamespace",
+            plural="jobs",
+            name="testjob",
+        )
+        call = replace_namespaced_custom_object_status.call_args
         args, kwargs = call
         self.assertEqual(
             kwargs,
@@ -814,9 +829,26 @@ def test_cancel_existing(self, delete_namespaced_custom_object: MagicMock) -> No
                 "namespace": "testnamespace",
                 "plural": "jobs",
                 "name": "testjob",
+                "body": {"status": {"state": {"phase": "Aborted"}}},
             },
         )
 
+    @patch("kubernetes.client.CustomObjectsApi.delete_namespaced_custom_object")
+    @patch("torchx.schedulers.kubernetes_scheduler.KubernetesScheduler.exists")
+    def test_delete(
+        self, exists: MagicMock, delete_namespaced_custom_object: MagicMock
+    ) -> None:
+        scheduler = create_scheduler("test")
+        exists.return_value = True
+        scheduler.delete("testnamespace:testjob")
+        delete_namespaced_custom_object.assert_called_once_with(
+            group="batch.volcano.sh",
+            version="v1alpha1",
+            namespace="testnamespace",
+            plural="jobs",
+            name="testjob",
+        )
+
     @patch("kubernetes.client.CustomObjectsApi.list_namespaced_custom_object")
     def test_list(self, list_namespaced_custom_object: MagicMock) -> None:
         with patch(