ob-metaflow-extensions 1.1.160rc0__py2.py3-none-any.whl → 1.1.161rc1__py2.py3-none-any.whl

metaflow_extensions/outerbounds/plugins/kubernetes/kubernetes_client.py

@@ -5,6 +5,7 @@ import time
 
 from metaflow.exception import MetaflowException
 from metaflow.metaflow_config import KUBERNETES_NAMESPACE
+from .pod_killer import PodKiller
 
 
 CLIENT_REFRESH_INTERVAL_SECONDS = 300
@@ -105,50 +106,23 @@ class KubernetesClient(object):
         return list(results)
 
     def kill_pods(self, flow_name, run_id, user, echo):
-        from kubernetes.stream import stream
-
-        api_instance = self._client.CoreV1Api()
-        job_api = self._client.BatchV1Api()
-        pods = self._find_active_pods(flow_name, run_id, user)
-
-        def _kill_pod(pod):
-            echo("Killing Kubernetes pod %s\n" % pod.metadata.name)
-            try:
-                stream(
-                    api_instance.connect_get_namespaced_pod_exec,
-                    name=pod.metadata.name,
-                    namespace=pod.metadata.namespace,
-                    command=[
-                        "/bin/sh",
-                        "-c",
-                        "/sbin/killall5",
-                    ],
-                    stderr=True,
-                    stdin=False,
-                    stdout=True,
-                    tty=False,
-                )
-            except Exception:
-                # best effort kill for pod can fail.
-                try:
-                    job_name = pod.metadata.labels.get("job-name", None)
-                    if job_name is None:
-                        raise Exception("Could not determine job name")
-
-                    job_api.patch_namespaced_job(
-                        name=job_name,
-                        namespace=pod.metadata.namespace,
-                        field_manager="metaflow",
-                        body={"spec": {"parallelism": 0}},
-                    )
-                except Exception as e:
-                    echo("failed to kill pod %s - %s" % (pod.metadata.name, str(e)))
-
-        with ThreadPoolExecutor() as executor:
-            operated_pods = list(executor.map(_kill_pod, pods))
-
-            if not operated_pods:
-                echo("No active Kubernetes pods found for run *%s*" % run_id)
+        # Create PodKiller instance
+        killer = PodKiller(self._client, echo, self._namespace)
+
+        # Process all matching jobs and jobsets based on their outcomes
+        (
+            job_jobset_results,
+            num_jobs,
+            num_jobsets,
+        ) = killer.process_matching_jobs_and_jobsets(flow_name, run_id, user)
+
+        if job_jobset_results:
+            successful_operations = sum(1 for result in job_jobset_results if result)
+            echo(
+                f"Found and processed {num_jobs} jobs and {num_jobsets} jobsets, {successful_operations} operations successful\n"
+            )
+        else:
+            echo("No matching jobs or jobsets found for run *%s*" % run_id)
 
     def job(self, **kwargs):
         from metaflow.plugins.kubernetes.kubernetes_job import KubernetesJob

metaflow_extensions/outerbounds/plugins/kubernetes/pod_killer.py

@@ -0,0 +1,298 @@
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from kubernetes.client.models.v1_job import V1Job
+    from kubernetes.client.models.v1_job_status import V1JobStatus
+
+
+def _is_jobset_child(job: "V1Job"):
+    if job.metadata.owner_references:
+        for owner_ref in job.metadata.owner_references:
+            if owner_ref.kind == "JobSet":
+                return owner_ref
+    return None
+
+
+class JobOutcomes:
+    KILL = "kill"
+    DELETE = "delete"
+    LEAVE_UNCHANGED = "leave_unchanged"
+
+
+def derive_jobset_outcome(jobset_status):
+    return (
+        JobOutcomes.LEAVE_UNCHANGED
+        if jobset_status.get("terminalState", None)
+        else JobOutcomes.DELETE
+    )
+
+
+def derive_job_outcome(job_status: "V1JobStatus"):
+    if job_status.start_time is None:
+        # If the job has not started even then just wipe it!
+        return JobOutcomes.DELETE
+    if job_status.succeeded or job_status.failed:
+        return JobOutcomes.LEAVE_UNCHANGED
+
+    if job_status.completion_time is not None:
+        return JobOutcomes.LEAVE_UNCHANGED
+
+    # This means that the job has neither finished or succedded.
+    if job_status.active:
+        return JobOutcomes.KILL
+
+    # This means that the job is not active. Had started. There is not succedded/fail.
+    # This is a weird state. Better to just kill the job
+    return JobOutcomes.DELETE
+
+
+class PodKiller:
+    def __init__(self, kubernetes_client, echo_func, namespace):
+        self.client = kubernetes_client
+        self.echo = echo_func
+        self.api_instance = self.client.CoreV1Api()
+        self.job_api = self.client.BatchV1Api()
+        self._namespace = namespace
+        self.jobset_api = None
+        self.jobset_api = self.client.CustomObjectsApi()
+
+    def _delete_jobset(self, owner_ref, namespace):
+        """Delete a JobSet if it's the owner of a job."""
+        if not self.jobset_api:
+            self.echo("JobSet API not available, cannot delete JobSet\n")
+            return False
+
+        try:
+            jobset_name = owner_ref.name
+            self.echo(f"Deleting JobSet: {jobset_name}\n")
+
+            self.jobset_api.delete_namespaced_custom_object(
+                group="jobset.x-k8s.io",
+                version="v1alpha2",
+                namespace=namespace,
+                plural="jobsets",
+                name=jobset_name,
+            )
+            return True
+        except Exception as e:
+            self.echo(f"Failed to delete JobSet {owner_ref.name}: {str(e)}\n")
+            return False
+
+    def _delete_job(self, job_name, namespace):
+        """Delete a Batch Job and check for JobSet owner reference."""
+        try:
+            # First get the job to check for owner references
+            job = self.job_api.read_namespaced_job(name=job_name, namespace=namespace)
+            # Check for JobSet owner reference
+            jobset_ref = _is_jobset_child(job)
+            if jobset_ref:
+                if self._delete_jobset(jobset_ref, namespace):
+                    return True
+
+            # If no JobSet owner or JobSet deletion failed, delete the job
+            self.echo(f"Deleting Batch Job: {job_name}")
+            self.job_api.delete_namespaced_job(
+                name=job_name, namespace=namespace, propagation_policy="Background"
+            )
+            return True
+
+        except Exception as e:
+            self.echo(f"Failed to delete job {job_name}: {str(e)}")
+            return False
+
+    def _kill_pod_process(self, pod):
+        """Attempt to kill processes inside a pod."""
+        from kubernetes.stream import stream
+
+        try:
+            stream(
+                self.api_instance.connect_get_namespaced_pod_exec,
+                name=pod.metadata.name,
+                namespace=pod.metadata.namespace,
+                command=["/bin/sh", "-c", "/sbin/killall5"],
+                stderr=True,
+                stdin=False,
+                stdout=True,
+                tty=False,
+            )
+            return True
+        except Exception as e:
+            self.echo(
+                f"Failed to kill processes in pod {pod.metadata.name}: {str(e)}\n"
+            )
+            return False
+
+    @staticmethod
+    def _metaflow_matching_spec(run_id, user, flow_name, annotations, labels):
+        # Handle argo prefixes in run_id like in _find_active_pods
+        _argo_run_id = None
+        if run_id is not None:
+            _argo_run_id = run_id[run_id.startswith("argo-") and len("argo-") :]
+        return (
+            annotations
+            and (
+                run_id is None
+                or (annotations.get("metaflow/run_id") == run_id)
+                # we want to also match jobsets launched by argo-workflows
+                # This line has no real value since the We already avoid any
+                # argo-workflows related terminations.
+                or (
+                    labels.get("workflows.argoproj.io/workflow") is not None
+                    and labels.get("workflows.argoproj.io/workflow") == _argo_run_id
+                )
+            )
+            and (user is None or annotations.get("metaflow/user") == user)
+            and (annotations.get("metaflow/flow_name") == flow_name)
+        )
+
+    def _find_matching_jobs(self, flow_name, run_id=None, user=None):
+        """Find jobs that match the flow_name, run_id, and user criteria using similar logic to _find_active_pods"""
+        try:
+            jobs = self.job_api.list_namespaced_job(namespace=self._namespace)
+            matching_jobs = []
+            for _job in jobs.items:
+                job = _job.to_dict()
+                _match = self._metaflow_matching_spec(
+                    run_id=run_id,
+                    user=user,
+                    flow_name=flow_name,
+                    annotations=job.get("metadata", {}).get("annotations", {}),
+                    labels=job.get("metadata", {}).get("labels", {}),
+                )
+                if _match:
+                    matching_jobs.append(_job)
+            return matching_jobs
+        except Exception as e:
+            self.echo(f"Error finding jobs: {str(e)}\n")
+            return []
+
+    def _find_matching_jobsets(self, flow_name, run_id=None, user=None):
+        """Find jobsets that match the flow_name, run_id, and user criteria using similar logic to _find_active_pods"""
+        if not self.jobset_api:
+            return []
+
+        try:
+            jobsets = self.jobset_api.list_namespaced_custom_object(
+                group="jobset.x-k8s.io",
+                version="v1alpha2",
+                namespace=self._namespace,
+                plural="jobsets",
+            )
+            matching_jobsets = []
+
+            for jobset in jobsets.get("items", []):
+                _match = self._metaflow_matching_spec(
+                    run_id=run_id,
+                    user=user,
+                    flow_name=flow_name,
+                    annotations=jobset.get("metadata", {}).get("annotations", {}),
+                    labels=jobset.get("metadata", {}).get("labels", {}),
+                )
+                if _match:
+                    matching_jobsets.append(jobset)
+
+            return matching_jobsets
+        except Exception as e:
+            self.echo(f"Error finding jobsets: {str(e)}\n")
+            return []
+
+    def _kill_pods_for_job(self, job):
+        """Find and kill pods associated with a specific job"""
+        job_name = job.metadata.name
+        namespace = job.metadata.namespace
+
+        try:
+            # Find pods with the job-name label matching this job
+            pods = self.api_instance.list_namespaced_pod(
+                namespace=namespace, label_selector=f"job-name={job_name}"
+            )
+
+            killed_pods = 0
+            for pod in pods.items:
+                if pod.status.phase in ["Running"]:
+                    self.echo(
+                        f"Killing processes in pod {pod.metadata.name} for job {job_name}"
+                    )
+                    if self._kill_pod_process(pod):
+                        killed_pods += 1
+
+            return killed_pods > 0
+        except Exception as e:
+            self.echo(f"Failed to find/kill pods for job {job_name}: {str(e)}")
+            return False
+
+    def _handle_job_outcome(self, job, outcome):
+        """Handle a job based on the derived outcome"""
+        job_name = job.metadata.name
+        namespace = job.metadata.namespace
+
+        if outcome == JobOutcomes.LEAVE_UNCHANGED:
+            # self.echo(f"Job {job_name} is in terminal state, leaving unchanged")
+            return None
+        elif outcome == JobOutcomes.DELETE:
+            self.echo(f"Deleting Job {job_name}")
+            return self._delete_job(job_name, namespace)
+        elif outcome == JobOutcomes.KILL:
+            self.echo(f"Killing Job {job_name}")
+            # First try to kill the pod processes
+            pods_killed = self._kill_pods_for_job(job)
+            if pods_killed > 0:
+                return True
+            # Worst case if we are not able to delete any pod, then delete the Job.
+            return self._delete_job(job_name, namespace)
+        else:
+            self.echo(f"Unknown outcome {outcome} for job {job_name}\n")
+            return False
+
+    def _handle_jobset_outcome(self, jobset, outcome):
+        """Handle a jobset based on the derived outcome"""
+        jobset_name = jobset.get("metadata", {}).get("name", "unknown")
+        namespace = jobset.get("metadata", {}).get("namespace", self._namespace)
+
+        if outcome == JobOutcomes.LEAVE_UNCHANGED:
+            # self.echo(f"JobSet {jobset_name} is in terminal state, leaving unchanged")
+            return None
+        elif outcome == JobOutcomes.DELETE:
+            self.echo(f"Deleting JobSet {jobset_name}")
+            try:
+                self.jobset_api.delete_namespaced_custom_object(
+                    group="jobset.x-k8s.io",
+                    version="v1alpha2",
+                    namespace=namespace,
+                    plural="jobsets",
+                    name=jobset_name,
+                )
+                return True
+            except Exception as e:
+                self.echo(f"Failed to delete JobSet {jobset_name}: {str(e)}")
+                return False
+        else:
+            self.echo(f"Unknown outcome {outcome} for JobSet {jobset_name}")
+            return False
+
+    def process_matching_jobs_and_jobsets(self, flow_name, run_id, user):
+        """Process all matching jobs and jobsets based on their derived outcomes"""
+        results = []
+
+        # Process matching jobs
+        _jobs, _jobsets = [], []
+        jobs = self._find_matching_jobs(flow_name, run_id, user)
+        for job in jobs:
+            outcome = derive_job_outcome(job.status)
+            result = self._handle_job_outcome(job, outcome)
+            # results.append(result)
+            if result is not None:
+                results.append(result)
+                _jobs.append(result)
+
+        # Process matching jobsets
+        jobsets = self._find_matching_jobsets(flow_name, run_id, user)
+        for jobset in jobsets:
+            jobset_status = jobset.get("status", {})
+            outcome = derive_jobset_outcome(jobset_status)
+            result = self._handle_jobset_outcome(jobset, outcome)
+            if result is not None:
+                results.append(result)
+                _jobsets.append(result)
+
+        return results, len(_jobs), len(_jobsets)

{ob_metaflow_extensions-1.1.160rc0.dist-info → ob_metaflow_extensions-1.1.161rc1.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ob-metaflow-extensions
-Version: 1.1.160rc0
+Version: 1.1.161rc1
 Summary: Outerbounds Platform Extensions for Metaflow
 Author: Outerbounds, Inc.
 License: Commercial

{ob_metaflow_extensions-1.1.160rc0.dist-info → ob_metaflow_extensions-1.1.161rc1.dist-info}/RECORD

@@ -23,7 +23,8 @@ metaflow_extensions/outerbounds/plugins/fast_bakery/fast_bakery.py,sha256=PE81ZB
 metaflow_extensions/outerbounds/plugins/fast_bakery/fast_bakery_cli.py,sha256=kqFyu2bJSnc9_9aYfBpz5xK6L6luWFZK_NMuh8f1eVk,1494
 metaflow_extensions/outerbounds/plugins/fast_bakery/fast_bakery_decorator.py,sha256=MXSIp05-jvt8Q2uGaLKjtuM_ToLeRLxhtMbfHc9Kcko,1515
 metaflow_extensions/outerbounds/plugins/kubernetes/__init__.py,sha256=5zG8gShSj8m7rgF4xgWBZFuY3GDP5n1T0ktjRpGJLHA,69
-metaflow_extensions/outerbounds/plugins/kubernetes/kubernetes_client.py,sha256=fx_XUkgR4r6hF2ilDfT5LubRyVrYMVIv5f6clHkCaEk,5988
+metaflow_extensions/outerbounds/plugins/kubernetes/kubernetes_client.py,sha256=sjBhQ4aa-i1UkKsJyTswdDLYOBAFIvHRco4r7wfs9Tc,5003
+metaflow_extensions/outerbounds/plugins/kubernetes/pod_killer.py,sha256=R483_Qg8eRifx5Z-3iHSQsFL6hUDmNTVLBlPWudCsQA,11404
 metaflow_extensions/outerbounds/plugins/nim/card.py,sha256=dXOJvsZed5NyYyxYLPDvtwg9z_X4azL9HTJGYaiNriY,4690
 metaflow_extensions/outerbounds/plugins/nim/nim_decorator.py,sha256=50YVvC7mcZYlPluM0Wq1UtufhzlQb-RxzZkTOJJ3LkM,3439
 metaflow_extensions/outerbounds/plugins/nim/nim_manager.py,sha256=y8U71106KJtrC6nlhsNnzX9Xkv3RnyZ1KEpRFwqZZFk,13686
@@ -74,7 +75,7 @@ metaflow_extensions/outerbounds/toplevel/plugins/kubernetes/__init__.py,sha256=5
 metaflow_extensions/outerbounds/toplevel/plugins/ollama/__init__.py,sha256=GRSz2zwqkvlmFS6bcfYD_CX6CMko9DHQokMaH1iBshA,47
 metaflow_extensions/outerbounds/toplevel/plugins/snowflake/__init__.py,sha256=LptpH-ziXHrednMYUjIaosS1SXD3sOtF_9_eRqd8SJw,50
 metaflow_extensions/outerbounds/toplevel/plugins/torchtune/__init__.py,sha256=uTVkdSk3xZ7hEKYfdlyVteWj5KeDwaM1hU9WT-_YKfI,50
-ob_metaflow_extensions-1.1.160rc0.dist-info/METADATA,sha256=1oYInnVvJpWGKs3SexiSvgx7i30fdQX8Fv25sXlW9Tw,524
-ob_metaflow_extensions-1.1.160rc0.dist-info/WHEEL,sha256=bb2Ot9scclHKMOLDEHY6B2sicWOgugjFKaJsT7vwMQo,110
-ob_metaflow_extensions-1.1.160rc0.dist-info/top_level.txt,sha256=NwG0ukwjygtanDETyp_BUdtYtqIA_lOjzFFh1TsnxvI,20
-ob_metaflow_extensions-1.1.160rc0.dist-info/RECORD,,
+ob_metaflow_extensions-1.1.161rc1.dist-info/METADATA,sha256=QZ_Y0zbT95Qr57WQ4XnvhZUSXoOQy3x2Pd-KT73mnxo,524
+ob_metaflow_extensions-1.1.161rc1.dist-info/WHEEL,sha256=bb2Ot9scclHKMOLDEHY6B2sicWOgugjFKaJsT7vwMQo,110
+ob_metaflow_extensions-1.1.161rc1.dist-info/top_level.txt,sha256=NwG0ukwjygtanDETyp_BUdtYtqIA_lOjzFFh1TsnxvI,20
+ob_metaflow_extensions-1.1.161rc1.dist-info/RECORD,,