ob-metaflow-extensions 1.1.45rc3__py2.py3-none-any.whl → 1.5.1__py2.py3-none-any.whl

metaflow_extensions/outerbounds/plugins/kubernetes/kubernetes_client.py

@@ -1,8 +1,11 @@
+from concurrent.futures import ThreadPoolExecutor
 import os
 import sys
 import time
 
 from metaflow.exception import MetaflowException
+from metaflow.metaflow_config import KUBERNETES_NAMESPACE
+from .pod_killer import PodKiller
 
 
 CLIENT_REFRESH_INTERVAL_SECONDS = 300
@@ -27,6 +30,7 @@ class KubernetesClient(object):
                 % sys.executable
             )
         self._refresh_client()
+        self._namespace = KUBERNETES_NAMESPACE
 
     def _refresh_client(self):
         from metaflow_extensions.outerbounds.plugins.auth_server import get_token
@@ -50,7 +54,82 @@ class KubernetesClient(object):
 
         return self._client
 
+    def _find_active_pods(self, flow_name, run_id=None, user=None):
+        def _request(_continue=None):
+            # handle paginated responses
+            return self._client.CoreV1Api().list_namespaced_pod(
+                namespace=self._namespace,
+                # limited selector support for K8S api. We want to cover multiple statuses: Running / Pending / Unknown
+                field_selector="status.phase!=Succeeded,status.phase!=Failed",
+                limit=1000,
+                _continue=_continue,
+            )
+
+        results = _request()
+
+        if run_id is not None:
+            # handle argo prefixes in run_id
+            run_id = run_id[run_id.startswith("argo-") and len("argo-") :]
+
+        while results.metadata._continue or results.items:
+            for pod in results.items:
+                match = (
+                    # arbitrary pods might have no annotations at all.
+                    pod.metadata.annotations
+                    and pod.metadata.labels
+                    and (
+                        run_id is None
+                        or (pod.metadata.annotations.get("metaflow/run_id") == run_id)
+                        # we want to also match pods launched by argo-workflows
+                        or (
+                            pod.metadata.labels.get("workflows.argoproj.io/workflow")
+                            == run_id
+                        )
+                    )
+                    and (
+                        user is None
+                        or pod.metadata.annotations.get("metaflow/user") == user
+                    )
+                    and (
+                        pod.metadata.annotations.get("metaflow/flow_name") == flow_name
+                    )
+                )
+                if match:
+                    yield pod
+            if not results.metadata._continue:
+                break
+            results = _request(results.metadata._continue)
+
+    def list(self, flow_name, run_id, user):
+        results = self._find_active_pods(flow_name, run_id, user)
+
+        return list(results)
+
+    def kill_pods(self, flow_name, run_id, user, echo):
+        # 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
 
         return KubernetesJob(self, **kwargs)
+
+    def jobset(self, **kwargs):
+        from metaflow.plugins.kubernetes.kubernetes_job import KubernetesJobSet
+
+        return KubernetesJobSet(self, **kwargs)

metaflow_extensions/outerbounds/plugins/kubernetes/pod_killer.py

@@ -0,0 +1,374 @@
+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.DELETE
+
+    # 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, progress_bar=None):
+        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()
+        self.progress_bar = progress_bar
+
+    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"""
+
+        def paginated_job_finder(namespace):
+            continue_token = None
+            while True:
+                response = self.job_api.list_namespaced_job(
+                    namespace=namespace, limit=100, _continue=continue_token
+                )
+                yield response.items
+                continue_token = response.metadata._continue
+                if not continue_token:
+                    break
+
+        try:
+            matching_jobs = []
+            for _jobs in paginated_job_finder(self._namespace):
+                for job in _jobs:
+                    _match = self._metaflow_matching_spec(
+                        run_id=run_id,
+                        user=user,
+                        flow_name=flow_name,
+                        annotations=job.metadata.annotations,
+                        labels=job.metadata.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 []
+
+        def paginated_jobset_finder(namespace):
+            continue_token = None
+            responses = []
+            while True:
+                response = self.jobset_api.list_namespaced_custom_object(
+                    group="jobset.x-k8s.io",
+                    version="v1alpha2",
+                    namespace=namespace,
+                    plural="jobsets",
+                    limit=100,
+                    **({"_continue": continue_token} if continue_token else {}),
+                )
+                continue_token = response.get("metadata", {}).get("continue", None)
+                responses.append(response)
+                if not continue_token:
+                    break
+            return responses
+
+        try:
+            matching_jobsets = []
+
+            for jobset_response in paginated_jobset_finder(self._namespace):
+                for jobset in jobset_response.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 extract_matching_jobs_and_jobsets(self, flow_name, run_id, user):
+        """Extract matching jobs and jobsets based on the flow_name, run_id, and user criteria"""
+        jobs = self._find_matching_jobs(flow_name, run_id, user)
+        jobsets = self._find_matching_jobsets(flow_name, run_id, user)
+        return [(j, derive_job_outcome(j.status)) for j in jobs], [
+            (j, derive_jobset_outcome(j.get("status", {}))) for j in jobsets
+        ]
+
+    def process_matching_jobs_and_jobsets(self, flow_name, run_id, user):
+        """Process all matching jobs and jobsets based on their derived outcomes"""
+        results = []
+        progress_update = lambda x: x
+        if self.progress_bar:
+            progress_update = lambda x: self.progress_bar.update(1, x)
+
+        # 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:
+                progress_update("💀 Killing Job %s" % job.metadata.name)
+                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:
+                progress_update(
+                    "💀 Deleting JobSet %s"
+                    % jobset.get("metadata", {}).get("name", "unknown")
+                )
+                results.append(result)
+                _jobsets.append(result)
+
+        return results, len(_jobs), len(_jobsets)
+
+    def process_matching_jobs_and_jobsets_force_all(self, flow_name, run_id, user):
+        """Force process ALL matching jobs and jobsets regardless of their status/outcome"""
+        results = []
+        progress_update = lambda x: x
+        if self.progress_bar:
+            progress_update = lambda x: self.progress_bar.update(1, x)
+
+        # Process matching jobs - FORCE DELETE ALL
+        _jobs, _jobsets = [], []
+        jobs = self._find_matching_jobs(flow_name, run_id, user)
+        for job in jobs:
+            # Force DELETE outcome regardless of actual status
+            result = self._handle_job_outcome(job, JobOutcomes.DELETE)
+            progress_update("🔥 FORCE Deleting Job %s" % job.metadata.name)
+            results.append(
+                result if result is not None else True
+            )  # Treat None as success for force mode
+            _jobs.append(result if result is not None else True)
+
+        # Process matching jobsets - FORCE DELETE ALL
+        jobsets = self._find_matching_jobsets(flow_name, run_id, user)
+        for jobset in jobsets:
+            # Force DELETE outcome regardless of actual status
+            result = self._handle_jobset_outcome(jobset, JobOutcomes.DELETE)
+            progress_update(
+                "🔥 FORCE Deleting JobSet %s"
+                % jobset.get("metadata", {}).get("name", "unknown")
+            )
+            results.append(
+                result if result is not None else True
+            )  # Treat None as success for force mode
+            _jobsets.append(result if result is not None else True)
+
+        return results, len(_jobs), len(_jobsets)

metaflow_extensions/outerbounds/plugins/nim/card.py

@@ -0,0 +1,140 @@
+from metaflow.cards import Markdown, Table
+from metaflow.metaflow_current import current
+
+from .utils import get_storage_path
+from ..card_utilities.async_cards import CardRefresher
+from ..card_utilities.extra_components import BarPlot, ViolinPlot
+
+
+class NimMetricsRefresher(CardRefresher):
+    CARD_ID = "nim_metrics"
+
+    def __init__(self) -> None:
+        self._metrics_charts = {}
+        self._last_updated_on = None
+        self._already_rendered = False
+        self._file_name = get_storage_path(current.task_id)
+
+    def sqlite_fetch_func(self, conn):
+        try:
+            cursor = conn.cursor()
+            cursor.execute(
+                "SELECT error, success, status_code, prompt_tokens, completion_tokens, e2e_time, model FROM metrics"
+            )
+            rows = cursor.fetchall()
+            data = {
+                "error": 0,
+                "success": 0,
+                "status_code": [],
+                "prompt_tokens": [],
+                "completion_tokens": [],
+                "e2e_time": [],
+                "model": [],
+            }
+            for row in rows:
+                data["error"] += row[0]
+                data["success"] += row[1]
+                data["status_code"].append(row[2])
+                data["prompt_tokens"].append(row[3])
+                data["completion_tokens"].append(row[4])
+                data["e2e_time"].append(row[5])
+                data["model"].append(row[6])
+            return data
+        finally:
+            conn.close()
+
+    def render_card_fresh(self, current_card, data):
+        self._already_rendered = True
+        current_card.clear()
+        current_card.append(Markdown("## Metrics"))
+
+        self._metrics_charts["request_success"] = BarPlot(
+            title="Request success",
+            category_name="category",
+            value_name="amount",
+            orientation="horizontal",
+        )
+        self._metrics_charts["latency_distribution"] = ViolinPlot(
+            title="Latency distribution (s)",
+            category_col_name="model",
+            value_col_name="e2e_time",
+        )
+
+        current_card.append(
+            Table(
+                data=[
+                    [
+                        self._metrics_charts["request_success"],
+                    ],
+                    [self._metrics_charts["latency_distribution"]],
+                ]
+            )
+        )
+        current_card.refresh()
+
+    def on_startup(self, current_card):
+        current_card.append(Markdown("# Task-level NIM API metrics"))
+        current_card.append(
+            Markdown(
+                "_waiting for data to appear_",
+            )
+        )
+        current_card.refresh()
+
+    def on_error(self, current_card, error_message):
+        if isinstance(error_message, FileNotFoundError):
+            return
+
+        if not self._already_rendered:
+            current_card.clear()
+            current_card.append(
+                Markdown(
+                    f"## Error: {str(error_message)}",
+                )
+            )
+            current_card.refresh()
+
+    def update_only_components(self, current_card, data_object):
+        # update request success data
+        self._metrics_charts["request_success"].spec["data"][0]["values"] = [
+            {
+                "category": "Successful requests",
+                "amount": data_object["metrics"]["success"],
+            },
+            {"category": "Errors", "amount": data_object["metrics"]["error"]},
+        ]
+
+        latency_data = []
+        times = []
+        for m, e in zip(
+            data_object["metrics"]["model"], data_object["metrics"]["e2e_time"]
+        ):
+            latency_data.append({"model": m, "e2e_time": e})
+            times.append(e)
+
+        # update latency data
+        self._metrics_charts["latency_distribution"].spec["data"][0][
+            "values"
+        ] = latency_data
+
+        # update domain for latency plot
+        min_time = min(times)
+        max_time = max(times)
+        for scale in self._metrics_charts["latency_distribution"].spec["scales"]:
+            if scale["name"] == "xscale":
+                scale["domain"] = [min_time - max_time * 0.1, max_time + max_time * 0.1]
+
+        current_card.refresh()
+
+    def on_update(self, current_card, data_object):
+        data_object_keys = set(data_object.keys())
+        if len(data_object_keys) == 0:
+            return
+        if len(self._metrics_charts) == 0:
+            self.render_card_fresh(current_card, data_object)
+            return
+        elif len(data_object["metrics"]["status_code"]) == 0:
+            return
+        else:
+            self.update_only_components(current_card, data_object)
+            return