Source code for graphscope.client.session

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
#
# Copyright 2020 Alibaba Group Holding Limited. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#

""" Manage sessions to the GraphScope coordinator.
"""

import atexit
import base64
import contextlib
import copy
import gc
import json
import logging
import os
import signal
import threading
import time
import uuid
import warnings
from typing import Any
from typing import Union

import graphscope
from graphscope.analytical.udf.utils import InMemoryZip
from graphscope.client.archive import OutArchive
from graphscope.client.rpc import GRPCClient
from graphscope.client.utils import CaptureKeyboardInterrupt
from graphscope.client.utils import GSLogger
from graphscope.client.utils import SignalIgnore
from graphscope.client.utils import set_defaults
from graphscope.config import Config
from graphscope.config import gs_config
from graphscope.deploy.hosts.cluster import HostsClusterLauncher
from graphscope.deploy.kubernetes.cluster import KubernetesClusterLauncher
from graphscope.deploy.kubernetes.utils import resolve_api_client
from graphscope.framework.app import App
from graphscope.framework.context import Context
from graphscope.framework.dag import Dag
from graphscope.framework.dag import DAGNode
from graphscope.framework.errors import FatalError
from graphscope.framework.errors import InvalidArgumentError
from graphscope.framework.graph import Graph
from graphscope.framework.graph import GraphDAGNode
from graphscope.framework.operation import Operation
from graphscope.framework.utils import decode_dataframe
from graphscope.framework.utils import decode_numpy
from graphscope.framework.utils import deprecated
from graphscope.framework.utils import random_string
from graphscope.interactive.query import InteractiveQuery
from graphscope.proto import graph_def_pb2
from graphscope.proto import message_pb2
from graphscope.proto import op_def_pb2
from graphscope.proto import types_pb2

try:
    import vineyard
    from kubernetes import client as kube_client
    from kubernetes import config as kube_config
except ImportError:
    kube_client = None
    kube_config = None

_session_dict = {}

logger = logging.getLogger("graphscope")


class _FetchHandler(object):
    """Handler for structured fetches.
    This class takes care of extracting a sub-DAG as targets for a user-provided structure for fetches,
    which can be used for a low level `run` call of grpc_client.

    Given the results of the low level run call, this class can also rebuild a result structure matching
    the user-provided structure for fetches, but containing the corresponding results.
    """

    def __init__(self, dag, fetches):
        self._fetches = fetches
        self._ops = list()
        self._unpack = False
        if not isinstance(self._fetches, (list, tuple)):
            self._fetches = [self._fetches]
            self._unpack = True
        for fetch in self._fetches:
            if hasattr(fetch, "op"):
                fetch = fetch.op
            if not isinstance(fetch, Operation):
                raise ValueError("Expect an `Operation` in sess run method.")
            self._ops.append(fetch)
        # extract sub dag
        self._sub_dag = dag.extract_subdag_for(self._ops)
        if "GRAPHSCOPE_DEBUG" in os.environ:
            logger.info("sub_dag: %s", self._sub_dag)

    @property
    def targets(self):
        return self._sub_dag

    def _rebuild_graph(self, seq, op_result: op_def_pb2.OpResult):
        if isinstance(self._fetches[seq], Operation):
            # for nx Graph
            return op_result.graph_def
        # get graph dag node as base
        graph_dag_node = self._fetches[seq]
        # construct graph
        g = Graph(graph_dag_node)
        # update graph flied from graph_def
        g.update_from_graph_def(op_result.graph_def)
        return g

    def _rebuild_app(self, seq, op_result: op_def_pb2.OpResult):
        from graphscope.framework.app import App

        # get app dag node as base
        app_dag_node = self._fetches[seq]
        # construct app
        app = App(app_dag_node, op_result.result.decode("utf-8", errors="ignore"))
        return app

    def _rebuild_context(self, seq, op_result: op_def_pb2.OpResult):
        from graphscope.framework.context import Context

        # get context dag node as base
        context_dag_node = self._fetches[seq]
        ret = json.loads(op_result.result.decode("utf-8", errors="ignore"))
        context_type = ret["context_type"]
        if context_type == "dynamic_vertex_data":
            # for nx
            from graphscope.framework.context import DynamicVertexDataContext

            return DynamicVertexDataContext(context_dag_node, ret["context_key"])
        return Context(context_dag_node, ret["context_key"], ret["context_schema"])

    def wrap_results(self, response: message_pb2.RunStepResponse):  # noqa: C901
        rets = list()
        for seq, op in enumerate(self._ops):
            for op_result in response.results:
                if op.key == op_result.key:
                    if op.output_types == types_pb2.RESULTS:
                        if op.type == types_pb2.RUN_APP:
                            rets.append(self._rebuild_context(seq, op_result))
                        elif op.type == types_pb2.REPORT_GRAPH:
                            rets.append(OutArchive(op_result.result))
                        else:
                            # for nx Graph
                            rets.append(
                                op_result.result.decode("utf-8", errors="ignore")
                            )
                    if op.output_types == types_pb2.GRAPH:
                        rets.append(self._rebuild_graph(seq, op_result))
                    if op.output_types == types_pb2.APP:
                        rets.append(None)
                    if op.output_types == types_pb2.BOUND_APP:
                        rets.append(self._rebuild_app(seq, op_result))
                    if op.output_types in (
                        types_pb2.VINEYARD_TENSOR,
                        types_pb2.VINEYARD_DATAFRAME,
                    ):
                        rets.append(
                            json.loads(
                                op_result.result.decode("utf-8", errors="ignore")
                            )["object_id"]
                        )
                    if op.output_types in (types_pb2.TENSOR, types_pb2.DATAFRAME):
                        if (
                            op.type == types_pb2.CONTEXT_TO_DATAFRAME
                            or op.type == types_pb2.GRAPH_TO_DATAFRAME
                        ):
                            rets.append(decode_dataframe(op_result.result))
                        if (
                            op.type == types_pb2.CONTEXT_TO_NUMPY
                            or op.type == types_pb2.GRAPH_TO_NUMPY
                        ):
                            rets.append(decode_numpy(op_result.result))
                    if op.output_types == types_pb2.NULL_OUTPUT:
                        rets.append(None)
                    break
        return rets[0] if rets and self._unpack else rets

    def get_dag_for_unload(self):
        """Unload operations (graph, app, context) in dag which are not
        existed in fetches.
        """
        unload_dag = op_def_pb2.DagDef()
        keys_of_fetches = set([op.key for op in self._ops])
        mapping = {
            types_pb2.CREATE_GRAPH: types_pb2.UNLOAD_GRAPH,
            types_pb2.CREATE_APP: types_pb2.UNLOAD_APP,
            types_pb2.RUN_APP: types_pb2.UNLOAD_CONTEXT,
        }
        for op_def in self._sub_dag.op:
            if op_def.op in mapping and op_def.key not in keys_of_fetches:
                unload_op_def = op_def_pb2.OpDef(
                    op=mapping[op_def.op], key=uuid.uuid4().hex
                )
                unload_op_def.parents.extend([op_def.key])
                unload_dag.op.extend([unload_op_def])
        return unload_dag


[docs]class Session(object): """A class for interacting with GraphScope graph computation service cluster. A :class:`Session` object encapsulates the environment in which :class:`Operation` objects are executed/evaluated. A session may own resources. It is important to release these resources when they are no longer required. To do this, invoke the :meth:`close` method on the session. A Session can register itself as default session with :meth:`as_default`, and all operations after that will use the default session. Session deregister itself as a default session when closed. The following example demonstrates its usage: .. code:: python >>> import graphscope as gs >>> # use session object explicitly >>> sess = gs.session() >>> g = sess.g() >>> pg = g.project(vertices={'v': []}, edges={'e': ['dist']}) >>> r = gs.sssp(g, 4) >>> sess.close() >>> # or use a session as default >>> sess = gs.session().as_default() >>> g = gs.g() >>> pg = g.project(vertices={'v': []}, edges={'e': ['dist']}) >>> r = gs.sssp(pg, 4) >>> sess.close() We support setup a service cluster and create a RPC session in following ways: - GraphScope graph computation service run in cluster managed by kubernetes. >>> s = graphscope.session() Also, :class:`Session` provides several keyword params for users to define the cluster. You may use the param :code:`k8s_gs_image` to specify the image for all engine pod, and param :code:`k8s_engine_cpu` or :code:`k8s_engine_mem` to specify the resources. More, you can find all params detail in :meth:`__init__` method. >>> s = graphscope.session( ... k8s_vineyard_cpu=0.1, ... k8s_vineyard_mem="256Mi", ... vineyard_shared_mem="4Gi", ... k8s_engine_cpu=0.1, ... k8s_engine_mem="256Mi") - or all params can be provided by a json configuration file or configuration dict. >>> s = graphscope.session(config='/tmp/config.yaml') >>> # Or >>> s = graphscope.session(config={'k8s_engine_cpu': 5, 'k8s_engine_mem': '5Gi'}) """
[docs] def __init__( self, config: Union[Config, str] = None, api_client: kube_client.ApiClient = None, **kw, ): """Construct a new GraphScope session. Args: config (dict or str, optional): The configuration dict or file about how to launch the GraphScope instance. For str, it will identify it as a path and read the configuration file to build a session if file exist. If not specified, the global default configuration will be used Note that it will overwrite explicit parameters. Defaults to None. api_client: The kube api client used in kubernetes cluster kw: Configurable keys. For backward compatibility. For more details, see `Config` class in `config.py` addr (str, optional): The endpoint of a pre-launched GraphScope instance with '<ip>:<port>' format. A new session id will be generated for each session connection. mode (str, optional): optional values are eager and lazy. Defaults to eager. Eager execution is a flexible platform for research and experimentation, it provides: An intuitive interface: Quickly test on small data. Easier debugging: Call ops directly to inspect running models and test changes. Lazy execution means GraphScope does not process the data till it has to. It just gathers all the information to a DAG that we feed into it, and processes only when we execute :code:`sess.run(fetches)` cluster_type (str, optional): Deploy GraphScope instance on hosts or k8s cluster. Defaults to k8s. Available options: "k8s" and "hosts". Note that only support deployed on localhost with hosts mode. num_workers (int, optional): The number of workers to launch GraphScope engine. Defaults to 2. preemptive (bool, optional): If True, GraphScope instance will treat resource params (e.g. k8s_coordinator_cpu) as limits and provide the minimum available value as requests, but this will make pod has a `Burstable` QOS, which can be preempted by other pods with high QOS. Otherwise, it will set both requests and limits with the same value. k8s_namespace (str, optional): Contains the namespace to create all resource inside. If param missing, it will try to read namespace from kubernetes context, or a random namespace will be created and deleted if namespace not exist. Defaults to None. k8s_service_type (str, optional): Type determines how the GraphScope service is exposed. Valid options are NodePort, and LoadBalancer. Defaults to NodePort. k8s_image_registry (str, optional): The GraphScope image registry. k8s_image_repository (str, optional): The GraphScope image repository. k8s_image_tag (str, optional): The GraphScope image tag. k8s_image_pull_policy (str, optional): Kubernetes image pull policy. Defaults to "IfNotPresent". k8s_image_pull_secrets (List[str], optional): A list of secret name used to authorize pull image. k8s_vineyard_image (str, optional): The image of vineyard. k8s_vineyard_deployment (str, optional): The name of vineyard deployment to use. GraphScope will try to discovery the deployment from kubernetes cluster, then use it if exists, and fallback to launching a bundled vineyard container otherwise. k8s_vineyard_cpu (float, optional): Number of CPU cores request for vineyard container. Defaults to 0.2. k8s_vineyard_mem (str, optional): Number of memory request for vineyard container. Defaults to '256Mi' k8s_engine_cpu (float, optional): Number of CPU cores request for engine container. Defaults to 1. k8s_engine_mem (str, optional): Number of memory request for engine container. Defaults to '4Gi'. k8s_coordinator_cpu (float, optional): Number of CPU cores request for coordinator. Defaults to 0.5. k8s_coordinator_mem (str, optional): Number of memory request for coordinator. Defaults to '512Mi'. etcd_addrs (str, optional): The addr of external etcd cluster, with formats like 'etcd01:port,etcd02:port,etcd03:port' k8s_mars_worker_cpu (float, optional): Minimum number of CPU cores request for Mars worker container. Defaults to 0.2. k8s_mars_worker_mem (str, optional): Minimum number of memory request for Mars worker container. Defaults to '4Mi'. k8s_mars_scheduler_cpu (float, optional): Minimum number of CPU cores request for Mars scheduler container. Defaults to 0.2. k8s_mars_scheduler_mem (str, optional): Minimum number of memory request for Mars scheduler container. Defaults to '4Mi'. k8s_coordinator_pod_node_selector (dict, optional): Node selector to the coordinator pod on k8s. Default is None. See also: https://tinyurl.com/3nx6k7ph k8s_engine_pod_node_selector = None Node selector to the engine pod on k8s. Default is None. See also: https://tinyurl.com/3nx6k7ph with_mars (bool, optional): Launch graphscope with Mars. Defaults to False. enabled_engines (str, optional): Select a subset of engines to enable. Only make sense in k8s mode. with_dataset (bool, optional): Create a container and mount aliyun demo dataset bucket to the path `/dataset`. k8s_volumes (dict, optional): A dict of k8s volume which represents a directory containing data, accessible to the containers in a pod. Defaults to {}. For example, you can mount host path with: k8s_volumes = { "my-data": { "type": "hostPath", "field": { "path": "<path>", "type": "Directory" }, "mounts": [ { "mountPath": "<path1>" }, { "mountPath": "<path2>" } ] } } Or you can mount PVC with: k8s_volumes = { "my-data": { "type": "persistentVolumeClaim", "field": { "claimName": "your-pvc-name" }, "mounts": [ { "mountPath": "<path1>" } ] } } Also, you can mount a single volume with: k8s_volumes = { "my-data": { "type": "hostPath", "field": {xxx}, "mounts": { "mountPath": "<path1>" } } } timeout_seconds (int, optional): For waiting service ready (or waiting for delete if k8s_waiting_for_delete is True). dangling_timeout_seconds (int, optional): After seconds of client disconnect, coordinator will kill this graphscope instance. Defaults to 600. Expect this value to be greater than 5 (heartbeat interval). Disable dangling check by setting -1. k8s_deploy_mode (str, optional): the deploy mode of engines on the kubernetes cluster. Default to eager. eager: create all engine pods at once lazy: create engine pods when called k8s_waiting_for_delete (bool, optional): Waiting for service delete or not. Defaults to False. k8s_client_config (dict, optional): config_file: Name of the kube-config file. Provide configurable parameters for connecting to remote k8s e.g. "~/.kube/config" reconnect (bool, optional): When connecting to a pre-launched GraphScope cluster with :code:`addr`, the connect request would be rejected with there is still an existing session connected. There are cases where the session still exists and user's client has lost connection with the backend, e.g., in a jupyter notebook. We have a :code:`dangling_timeout_seconds` for it, but a more deterministic behavior would be better. If :code:`reconnect` is True, the existing session will be reused. It is the user's responsibility to ensure there's no such an active client. Defaults to :code:`False`. Raises: TypeError: If the given argument combination is invalid and cannot be used to create a GraphScope session. """ # suppress the grpc warnings, see also grpc/grpc#29103 os.environ["GRPC_ENABLE_FORK_SUPPORT"] = "false" self._accessable_params = ( "addr", "mode", "cluster_type", "num_workers", "preemptive", "k8s_namespace", "k8s_service_type", "k8s_image_registry", "k8s_image_repository", "k8s_image_tag", "k8s_image_pull_policy", "k8s_image_pull_secrets", "k8s_coordinator_cpu", "k8s_coordinator_mem", "etcd_addrs", "etcd_listening_client_port", "etcd_listening_peer_port", "k8s_vineyard_image", "k8s_vineyard_deployment", "k8s_vineyard_cpu", "k8s_vineyard_mem", "vineyard_shared_mem", "k8s_engine_cpu", "k8s_engine_mem", "k8s_mars_worker_cpu", "k8s_mars_worker_mem", "k8s_mars_scheduler_cpu", "k8s_mars_scheduler_mem", "k8s_coordinator_pod_node_selector", "k8s_engine_pod_node_selector", "enabled_engines", "reconnect", "k8s_volumes", "k8s_waiting_for_delete", "k8s_deploy_mode", "timeout_seconds", "dangling_timeout_seconds", "with_mars", "with_dataset", "hosts", ) # parse config, which should be a path to config file, or dict # config has the highest priority if config is not None: if isinstance(config, str): self._config = self._load_config_from_file(config, silent=False) else: self._config = copy.deepcopy(config) else: self._config = copy.deepcopy(gs_config) # default config self._api_client = api_client for key, value in kw.items(): self._config.set_option(key, value) self._config.session.instance_id = random_string(6) # initial setting of cluster_type self._cluster_type = self._parse_cluster_type() # initial dag self._dag = Dag() # the mapping table from old vineyard object id to new vineyard object id self._vineyard_object_mapping_table = {} self._log_session_info() self._closed = False # coordinator service endpoint self._coordinator_endpoint = None self._launcher = None self._heartbeat_sending_thread = None self._grpc_client: GRPCClient = None self._session_id: str = None # unique identifier across sessions # engine config: # # { # "experiment": "ON/OFF", # "vineyard_socket": "...", # "vineyard_rpc_endpoint": "..." # } self._engine_config: dict = None # interactive instance related graph map self._interactive_instance_dict = {} # learning engine related graph map self._learning_instance_dict = {} self._default_session = None atexit.register(self.close) # create and connect session with CaptureKeyboardInterrupt(self.close): self._connect() self._disconnected: bool = False # heartbeat self._heartbeat_interval_seconds: int = 5 self._heartbeat_sending_thread = threading.Thread( target=self._send_heartbeat, args=() ) self._heartbeat_sending_thread.daemon = True self._heartbeat_sending_thread.start() self._heartbeat_maximum_failures: int = 3 # networkx module self._nx = None self._lock = threading.RLock()
def __repr__(self): return str(self.info) def __str__(self): return repr(self) @property def session_id(self) -> str: return self._session_id @property def dag(self) -> Dag: return self._dag def _log_session_info(self): if self._config.coordinator.endpoint is not None: logger.info( "Connecting graphscope session with address: %s", self._config.coordinator.endpoint, ) else: logger.info( "Initializing graphscope session with parameters: %s", self._config.dumps_json(), ) def _load_config_from_file(self, path, silent=True): config_path = os.path.expandvars(os.path.expanduser(path)) try: return Config.load(config_path, drop_extra_fields=False) except Exception as exp: # noqa if not silent: raise exp def _parse_cluster_type(self): # get the cluster type after connecting if self._config.launcher_type == "hosts": cluster_type = types_pb2.HOSTS elif self._config.launcher_type == "k8s": cluster_type = types_pb2.K8S else: raise ValueError("Expect 'hosts' or 'k8s' for cluster_type parameter") return cluster_type @property def engine_config(self): """Show the engine configuration associated with session in json format.""" return self._engine_config @property def info(self): """Show all resource info associated with session in json format.""" info = {} if self._closed: info["status"] = "closed" elif self._grpc_client is None or self._disconnected: info["status"] = "disconnected" else: info["status"] = "active" if self._cluster_type == types_pb2.K8S: info["type"] = "k8s" info["engine_hosts"] = ",".join(self._pod_name_list) info["namespace"] = self._config.kubernetes_launcher.namespace else: info["type"] = "hosts" info["engine_hosts"] = self._engine_config["engine_hosts"] info["cluster_type"] = types_pb2.ClusterType.Name(self._cluster_type) info["session_id"] = self.session_id info["num_workers"] = self._config.session.num_workers info["coordinator_endpoint"] = self._coordinator_endpoint info["engine_config"] = self._engine_config return info @property def closed(self): return self._closed @property def disconnected(self): return self._grpc_client is None or self._disconnected def eager(self): return self._config.session.execution_mode == "eager" def _send_heartbeat(self): # >1: failure, 0: reset when success heartbeat_failure_count = 0 while not self._closed: if self._grpc_client: try: self._grpc_client.send_heartbeat() except Exception as exc: if heartbeat_failure_count == 0: logger.warning("Failed to send heartbeat message", exc_info=exc) heartbeat_failure_count = heartbeat_failure_count + 1 if heartbeat_failure_count > self._heartbeat_maximum_failures: logger.error( "The connection between coordinator has lost after %d times " "of heartbeat failure, closing the session ...", heartbeat_failure_count, ) self.close() self._disconnected = True else: heartbeat_failure_count = 0 self._disconnected = False time.sleep(self._heartbeat_interval_seconds)
[docs] def connected(self) -> bool: """Check if the session is still connected and available. Returns: True or False """ return not self._disconnected
[docs] def close(self): """Closes this session. This method frees all resources associated with the session. Note that closing will ignore SIGINT and SIGTERM signal and recover later. """ if threading.current_thread() is threading.main_thread(): with SignalIgnore([signal.SIGINT, signal.SIGTERM]): self._close() else: self._close()
def _close(self): # noqa: C901 if self._closed: return self._closed = True self._coordinator_endpoint = None self._unregister_default() if self._heartbeat_sending_thread: try: self._heartbeat_sending_thread.join( timeout=self._heartbeat_interval_seconds ) except RuntimeError: # ignore the "cannot join current thread" error pass self._heartbeat_sending_thread = None self._disconnected = True # close all interactive instances for instance in self._interactive_instance_dict.values(): try: instance.close() except Exception: pass self._interactive_instance_dict.clear() # close all learning instances for instance in self._learning_instance_dict.values(): try: instance.close() except Exception: pass self._learning_instance_dict.clear() if self._grpc_client: try: self._grpc_client.close() except Exception: pass self._grpc_client = None _session_dict.pop(self._session_id, None) # clean up if self._config.coordinator.endpoint is None: try: if self._launcher: self._launcher.stop() except Exception: pass self._pod_name_list = [] def _close_interactive_instance(self, instance): """Close an interactive instance.""" self._grpc_client.close_interactive_instance(instance.object_id) def _close_learning_instance(self, instance): """Close a learning instance.""" self._grpc_client.close_learning_instance(instance.object_id) def __del__(self): # cleanly ignore all exceptions try: self.close() except Exception: # pylint: disable=broad-except pass def _check_closed(self, msg=None): """Internal: raise a ValueError if session is closed""" if self.closed: raise ValueError(msg or "Operation on closed session.") # Context manager
[docs] def __enter__(self): """Context management protocol. Returns self and register self as default session. """ self._check_closed() self.as_default() return self
[docs] def __exit__(self, exc_type, exc_value, exc_tb): """Unregister self from the default session, close the session and release the resources, ignore all exceptions in close(). """ try: self._unregister_default() self.close() except Exception: pass
[docs] def as_default(self): """Obtain a context manager that make this object as default session. This method is used when a Session is constructed, which will immediately install self as a default session. Raises: ValueError: If default session exist in current context. Returns: A context manager using this session as the default session. """ if not _default_session_stack.is_cleared(): raise ValueError( "A default session is already active. You must explicitly call Session.close()." ) # session context manager self._default_session = default_session(self) self._default_session.__enter__()
def _unregister_default(self): """Remove self from the default session stack.""" if self._default_session: self._default_session.__exit__(None, None, None) self._default_session = None def _wrapper(self, dag_node: DAGNode) -> Union[DAGNode, App, Context, Graph, Any]: if self.eager(): return self.run(dag_node) return dag_node def run(self, fetches): """Run operations of `fetches`. Args: fetches: :class:`Operation` Raises: RuntimeError: Client disconnect to the service. Or run on a closed session. ValueError: If fetch is not an instance of :class:`Operation`. Or the fetch has been evaluated. InvalidArgumentError: Not recognized on output type. Returns: Different values for different output types of :class:`Operation` """ # There might be a deadlock without `gc.collect()`: # # - thread 1 uses `run()` to issue grpc requests # - during the process, e.g., print traceback, it triggers certain `__del__()` # and that issues a `run_fetches()` again, that further requires the lock # - then a deadlock been introduced. # # Thus, we simply choose to call `gc.collect()` to force those `__del__()` been # invoked before actually issuing the grpc request to avoid the deadlock. # gc.collect() with self._lock: return self.run_fetches(fetches)
[docs] def run_fetches(self, fetches): """Run operations of `fetches` without the session lock.""" if self._closed: raise RuntimeError("Attempted to use a closed Session.") if not self._grpc_client: raise RuntimeError("Session disconnected.") fetch_handler = _FetchHandler(self.dag, fetches) try: response = self._grpc_client.run(fetch_handler.targets) except FatalError: self.close() raise if not self.eager(): # Unload operations that cannot be touched anymore dag_to_unload = fetch_handler.get_dag_for_unload() try: self._grpc_client.run(dag_to_unload) except FatalError: self.close() raise return fetch_handler.wrap_results(response)
def _connect(self): if self._config.coordinator.endpoint is not None: # try to connect to exist coordinator self._coordinator_endpoint = self._config.coordinator.endpoint elif self._cluster_type == types_pb2.K8S: self._launcher = KubernetesClusterLauncher( config=self._config, api_client=self._get_api_client() ) elif self._cluster_type == types_pb2.HOSTS: # launch coordinator with hosts self._launcher = HostsClusterLauncher(config=self._config) else: raise RuntimeError( f"Unrecognized cluster type {types_pb2.ClusterType.Name(self._cluster_type)}." ) # launching graphscope service if self._launcher is not None: self._launcher.start() self._coordinator_endpoint = self._launcher.coordinator_endpoint # waiting service ready self._grpc_client = GRPCClient( self._launcher, self._coordinator_endpoint, self._config.session.reconnect ) self._grpc_client.waiting_service_ready( timeout_seconds=self._config.session.timeout_seconds, ) # connect and fetch logs from rpc server try: ( self._session_id, self._cluster_type, self._config.session.num_workers, self._config.kubernetes_launcher.namespace, self._engine_config, pod_name_list, ) = self._grpc_client.connect( cleanup_instance=not bool(self._config.coordinator.endpoint), dangling_timeout_seconds=self._config.session.dangling_timeout_seconds, ) self._pod_name_list = list(pod_name_list) # fetch logs if self._config.coordinator.endpoint or self._cluster_type == types_pb2.K8S: self._grpc_client.fetch_logs() _session_dict[self._session_id] = self # Launch analytical engine right after session connected. # This may be changed to on demand launching in the future if not self._engine_config and not self._pod_name_list: ( self._engine_config, pod_name_list, ) = self._grpc_client.create_analytical_instance() self._pod_name_list = list(pod_name_list) except Exception: self.close() raise
[docs] def get_config(self): """Get configuration of the session.""" return self._config
def _get_api_client(self): if self._api_client is not None: return self._api_client try: config_file = self._config.kubernetes_launcher.config_file self._api_client = resolve_api_client(config_file) except kube_config.ConfigException as e: raise RuntimeError( "Kubernetes environment not found, you may want to" ' launch session locally with param cluster_type="hosts"' ) from e return self._api_client def _ensure_pvc_exists(self, pvc_name, pvc_namespace): _core_api = kube_client.CoreV1Api(self._get_api_client()) try: _core_api.read_namespaced_persistent_volume_claim( name=pvc_name, namespace=pvc_namespace, ) except kube_client.rest.ApiException as e: raise RuntimeError( f"PVC {pvc_name} not found in namespace {pvc_namespace}" ) from e def _ensure_vineyard_deployment_exists( self, vineyard_deployment_name, vineyard_deployment_namespace ): _app_api = kube_client.AppsV1Api(self._get_api_client()) try: _app_api.read_namespaced_deployment( name=vineyard_deployment_name, namespace=vineyard_deployment_namespace, ) except kube_client.rest.ApiException as e: if e.status == 404: raise RuntimeError( f"Vineyard deployment {vineyard_deployment_name} not found in namespace {vineyard_deployment_namespace}" ) from e else: raise e
[docs] def store_to_pvc(self, graphIDs, path: str, pvc_name: str): """ Stores the given graph IDs to the given path with the given PVC. Also, if you want to store graphs of different sessions to the same pv, you'd better to create different pvc for different sessions at first. Notice, before calling this function, the KUBECONFIG environment variable should be set to the path of your kubeconfig file. And you should make sure that the pvc is bound to the pv and the pv's capacity is enough to store the graphs. The method uses the vineyardctl to create a kubernetes job to serialize the selected graphs. For more information, see the vineyardctl documentation. https://github.com/v6d-io/v6d/tree/main/k8s/cmd#vineyardctl-deploy-backup-job Args: graph_ids: The list of graph IDs to store. Supported types: - list: list of vineyard.ObjectID or graphscope.Graph path: The path in the pv to which the pvc is bound. pvc_name: The name of the PVC. Raises: RuntimeError: If the cluster type is not Kubernetes. """ if self._cluster_type != types_pb2.K8S: raise RuntimeError("Only support kubernetes cluster") object_ids = [] for object in graphIDs: if isinstance(object, Graph): object_ids.append(object.vineyard_id) else: object_ids.append(vineyard.ObjectID(object)) object_ids = ",".join(repr(id) for id in object_ids) vineyard_deployment_name = self._config.vineyard.deployment_name namespace = self._config.kubernetes_launcher.namespace self._ensure_vineyard_deployment_exists(vineyard_deployment_name, namespace) self._ensure_pvc_exists(pvc_name, namespace) # The next function will create a kubernetes job for backing up # the specific graphIDs to the specific path of the specific pvc vineyard.deploy.vineyardctl.deploy.backup_job( backup_name="vineyard-backup-" + random_string(6), vineyard_deployment_name=vineyard_deployment_name, vineyard_deployment_namespace=namespace, namespace=namespace, path=path, objectids=object_ids, pvc_name=pvc_name, )
[docs] def restore_from_pvc(self, path: str, pvc_name: str): """ Restores the graphs from the given path in the given PVC. Notice, before calling this function, the KUBECONFIG environment variable should be set to the path of your kubeconfig file. Args: path: The path in the pv to which the pvc is bound. pvc_name: The name of the PVC. Raises: RuntimeError: If the cluster type is not Kubernetes. """ if self._cluster_type != types_pb2.K8S: raise RuntimeError("Only support kubernetes cluster") vineyard_deployment_name = self._config.vineyard.deployment_name namespace = self._config.kubernetes_launcher.namespace self._ensure_vineyard_deployment_exists(vineyard_deployment_name, namespace) self._ensure_pvc_exists(pvc_name, namespace) random_suffix = random_string(6) vineyard.deploy.vineyardctl.deploy.recover_job( recover_name="vineyard-recover-" + random_suffix, vineyard_deployment_name=vineyard_deployment_name, vineyard_deployment_namespace=namespace, namespace=namespace, recover_path=path, pvc_name=pvc_name, ) _core_api = kube_client.CoreV1Api(self._get_api_client()) try: config_map = _core_api.read_namespaced_config_map( name="vineyard-recover-" + random_suffix + "-mapping-table", namespace=namespace, ) except kube_client.rest.ApiException as e: if e.status == 404: raise RuntimeError( f"ConfigMap vineyard-recover-{random_suffix}-mapping-table not found in namespace {namespace}" ) from e else: raise e # parse configmap data to the self._vineyard_object_mapping_table self._vineyard_object_mapping_table = config_map.data
[docs] def get_vineyard_object_mapping_table(self): """ Get the vineyard object mapping table from the old object id to new object id during storing and restoring graph to pvc on the kubernetes cluster. """ return self._vineyard_object_mapping_table
[docs] def g( self, incoming_data=None, oid_type="int64", vid_type="uint64", directed=True, generate_eid=True, retain_oid=True, vertex_map="global", compact_edges=False, use_perfect_hash=False, ) -> Union[Graph, GraphDAGNode]: """Construct a GraphScope graph object on the default session. It will launch and set a session to default when there is no default session found. See params detail in :class:`graphscope.framework.graph.GraphDAGNode` Returns: :class:`graphscope.framework.graph.GraphDAGNode`: Evaluated in eager mode. Examples: .. code:: python >>> import graphscope >>> g = graphscope.g() >>> import graphscope >>> sess = graphscope.session() >>> g = sess.g() # creating graph on the session "sess" """ if ( isinstance(incoming_data, vineyard.ObjectID) and repr(incoming_data) in self._vineyard_object_mapping_table ): graph_vineyard_id = self._vineyard_object_mapping_table[repr(incoming_data)] logger.info("Restore graph from original graph: %s", graph_vineyard_id) incoming_data = vineyard.ObjectID(graph_vineyard_id) return self._wrapper( GraphDAGNode( self, incoming_data, oid_type, vid_type, directed, generate_eid, retain_oid, vertex_map, compact_edges, use_perfect_hash, ) )
[docs] def load_from(self, *args, **kwargs): """Load a graph within the session. See more information in :meth:`graphscope.load_from`. """ with default_session(self): return graphscope.load_from(*args, **kwargs)
@deprecated("Please use `sess.interactive` instead.") def gremlin(self, graph, params=None): """This method is going to be deprecated. Use :meth:`interactive` to get an interactive engine handler supports both gremlin and cypher queries """ return self.interactive(graph, params)
[docs] def interactive(self, graph, params=None, with_cypher=False): """Get an interactive engine handler to execute gremlin and cypher queries. It will return an instance of :class:`graphscope.interactive.query.InteractiveQuery`, .. code:: python >>> # close and recreate InteractiveQuery. >>> interactive_query = sess.interactive(g) >>> interactive_query.close() >>> interactive_query = sess.interactive(g) Args: graph (:class:`graphscope.framework.graph.GraphDAGNode`): The graph to create interactive instance. params: A dict consists of configurations of GIE instance. Raises: InvalidArgumentError: - :code:`graph` is not a property graph. Returns: :class:`graphscope.interactive.query.InteractiveQuery`: InteractiveQuery to execute gremlin and cypher queries. """ if self._session_id != graph.session_id: raise RuntimeError( "Failed to create interactive engine on the graph with different session: {0} vs {1}".format( self._session_id, graph.session_id ) ) if not graph.graph_type == graph_def_pb2.ARROW_PROPERTY: raise InvalidArgumentError("The graph should be a property graph.") if not isinstance(graph, Graph): # Is a GraphDAGNode graph = self.run(graph) object_id = graph.vineyard_id schema_path = graph.schema_path ( gremlin_endpoint, cypher_endpoint, ) = self._grpc_client.create_interactive_instance( object_id, schema_path, params, with_cypher ) interactive_query = InteractiveQuery(graph, gremlin_endpoint, cypher_endpoint) self._interactive_instance_dict[object_id] = interactive_query graph._attach_interactive_instance(interactive_query) return interactive_query
@deprecated("Please use `graphlearn` instead.") def learning(self, graph, nodes=None, edges=None, gen_labels=None): """Start a graph learning engine. Note that this method has been deprecated, using `graphlearn` replace. """ warnings.warn( "The method 'learning' has been deprecated, using graphlearn replace." ) return self.graphlearn(graph, nodes, edges, gen_labels)
[docs] def graphlearn(self, graph, nodes=None, edges=None, gen_labels=None): """Start a graph learning engine. Args: graph (:class:`graphscope.framework.graph.GraphDAGNode`): The graph to create learning instance. nodes (list, optional): list of node types that will be used for GNN training, the element of list can be `"node_label"` or `(node_label, features)`. If the element of the list is a tuple and contains selected feature list, it would use the selected feature list for training. Default is None which use all type of nodes and for the GNN training. edges (list, optional): list of edge types that will be used for GNN training. We use `(src_label, edge_label, dst_label)` to specify one edge type. Default is None which use all type of edges for GNN training. gen_labels (list, optional): Alias node and edge labels and extract train/validation/test dataset from original graph for supervised GNN training. The detail is explained in the examples below. Examples -------- >>> # Assume the input graph contains one label node `paper` and one edge label `link`. >>> features = ["weight", "name"] # use properties "weight" and "name" as features >>> lg = sess.graphlearn( graph, nodes=[("paper", features)]) # use "paper" node and features for training edges=[("paper", "links", "paper")] # use the `paper->links->papers` edge type for training gen_labels=[ # split "paper" nodes into 100 pieces, and uses random 75 pieces (75%) as training dataset ("train", "paper", 100, (0, 75)), # split "paper" nodes into 100 pieces, and uses random 10 pieces (10%) as validation dataset ("val", "paper", 100, (75, 85)), # split "paper" nodes into 100 pieces, and uses random 15 pieces (15%) as test dataset ("test", "paper", 100, (85, 100)), ] ) Note that the training, validation and test datasets are not overlapping. And for unsupervised learning: >>> lg = sess.graphlearn( graph, nodes=[("paper", features)]) # use "paper" node and features for training edges=[("paper", "links", "paper")] # use the `paper->links->papers` edge type for training gen_labels=[ # split "paper" nodes into 100 pieces, and uses all pieces as training dataset ("train", "paper", 100, (0, 100)), ] ) """ if self._session_id != graph.session_id: raise RuntimeError( "Failed to create learning engine on the graph with different session: {0} vs {1}".format( self._session_id, graph.session_id ) ) if not graph.graph_type == graph_def_pb2.ARROW_PROPERTY: raise InvalidArgumentError("The graph should be a property graph.") from graphscope.learning.graph import Graph as LearningGraph from graphscope.learning.graph import get_gl_handle handle = get_gl_handle( graph.schema, graph.vineyard_id, self._pod_name_list, self._engine_config, graph.fragments, ) config = LearningGraph.preprocess_args(handle, nodes, edges, gen_labels) config = base64.b64encode( json.dumps(config).encode("utf-8", errors="ignore") ).decode("utf-8", errors="ignore") handle, config, endpoints = self._grpc_client.create_learning_instance( graph.vineyard_id, handle, config, message_pb2.LearningBackend.GRAPHLEARN ) handle = json.loads(base64.b64decode(handle).decode("utf-8", errors="ignore")) handle["server"] = ",".join(endpoints) handle["client_count"] = 1 # construct learning graph g = LearningGraph(graph, handle, config, graph.vineyard_id) self._learning_instance_dict[graph.vineyard_id] = g graph._attach_learning_instance(g) return g
def graphlearn_torch( self, graph, edges, edge_weights=None, node_features=None, edge_features=None, node_labels=None, edge_dir="out", random_node_split=None, num_clients=1, manifest_path=None, client_folder_path="./", return_pyg_remote_backend=False, ): from graphscope.learning.gl_torch_graph import GLTorchGraph from graphscope.learning.gs_feature_store import GsFeatureStore from graphscope.learning.gs_graph_store import GsGraphStore from graphscope.learning.utils import fill_params_in_yaml from graphscope.learning.utils import read_folder_files_content handle = { "vineyard_socket": self._engine_config["vineyard_socket"], "vineyard_id": graph.vineyard_id, "fragments": graph.fragments, "num_servers": len(graph.fragments), "num_clients": num_clients, } manifest_params = { "NUM_CLIENT_NODES": handle["num_clients"], "NUM_SERVER_NODES": handle["num_servers"], "NUM_WORKER_REPLICAS": handle["num_clients"] - 1, } if manifest_path is not None: handle["manifest"] = fill_params_in_yaml(manifest_path, manifest_params) if client_folder_path is not None: handle["client_content"] = read_folder_files_content(client_folder_path) handle = base64.b64encode( json.dumps(handle).encode("utf-8", errors="ignore") ).decode("utf-8", errors="ignore") config = { "edges": edges, "edge_weights": edge_weights, "node_features": node_features, "edge_features": edge_features, "node_labels": node_labels, "edge_dir": edge_dir, "random_node_split": random_node_split, } GLTorchGraph.check_params(graph.schema, config) config = GLTorchGraph.transform_config(config) config = base64.b64encode( json.dumps(config).encode("utf-8", errors="ignore") ).decode("utf-8", errors="ignore") handle, config, endpoints = self._grpc_client.create_learning_instance( graph.vineyard_id, handle, config, message_pb2.LearningBackend.GRAPHLEARN_TORCH, ) g = GLTorchGraph(endpoints) self._learning_instance_dict[graph.vineyard_id] = g graph._attach_learning_instance(g) if return_pyg_remote_backend: feature_store = GsFeatureStore(config) graph_store = GsGraphStore(config) return g, feature_store, graph_store return g def nx(self): if not self.eager(): raise RuntimeError( "Networkx module need the session to be eager mode. " "Current session is lazy mode." ) if self._nx: return self._nx import importlib.util spec = importlib.util.find_spec("graphscope.nx") mod = importlib.util.module_from_spec(spec) spec.loader.exec_module(mod) graph = type("Graph", (mod.Graph.__base__,), dict(mod.Graph.__dict__)) digraph = type("DiGraph", (mod.DiGraph.__base__,), dict(mod.DiGraph.__dict__)) setattr(graph, "_session", self) setattr(digraph, "_session", self) setattr(mod, "Graph", graph) setattr(mod, "DiGraph", digraph) self._nx = mod return self._nx
[docs] def add_lib(self, resource_name): """ add the specified resource to the k8s cluster from client machine. """ logger.info("client: adding lib %s", resource_name) if not os.path.isfile(resource_name): raise RuntimeError("Resource {} can not be found".format(resource_name)) # pack into a gar file garfile = InMemoryZip() resource_reader = open(resource_name, "rb") bytes_ = resource_reader.read() if len(bytes_) <= 0: raise RuntimeError("Expect a non-empty file.") # the uploaded file may be placed in the same directory garfile.append(resource_name.split("/")[-1], bytes_) self._grpc_client.add_lib(garfile.read_bytes().getvalue())
session = Session
[docs]def set_option(**kwargs): """Set the value of specified options. Find params detail in :class:`graphscope.Session` Available options: - num_workers - log_level - show_log - vineyard_shared_mem - k8s_namespace - k8s_service_type - k8s_gs_image - k8s_etcd_image - k8s_image_pull_policy - k8s_image_pull_secrets - k8s_coordinator_cpu - k8s_coordinator_mem - k8s_vineyard_deployment - k8s_vineyard_cpu - k8s_vineyard_mem - k8s_engine_cpu - k8s_engine_mem - k8s_mars_worker_cpu - k8s_mars_worker_mem - k8s_mars_scheduler_cpu - k8s_mars_scheduler_mem - enabled_engines - with_mars - with_dataset - k8s_volumes - k8s_waiting_for_delete - timeout_seconds - dataset_download_retries - k8s_deploy_mode Args: kwargs: dict kv pair of GraphScope config you want to set. Raises: ValueError: If no such option exists. Returns: None """ for k, v in kwargs.items(): gs_config.set_option(k, v) # use string as log level if k == "log_level" and isinstance(v, int): level = logging.getLevelName(v) if " " not in level: # invalid number will return "Level xxx" gs_config.set_option(k, level.upper()) GSLogger.update()
def default_session(session): """Python's :code:`with` handler for defining a default session. This function provides a means of registering a session for handling and code that need a default session calls. The :code:`with` keyword to specify that code invocations within the scope of a block should be executed by a particular session. Args: session: :class:`Session` The session to be installed as the default session. Returns: A context manager for the default session. """ return _default_session_stack.get_controller(session)
[docs]def has_default_session() -> bool: """True if default session exists in current context.""" return not _default_session_stack.empty()
[docs]def get_default_session() -> Session: """Returns the default session for the current context. Note that a new session will be created if there is no default session in current context. Returns: The default :class:`graphscope.Session`. """ return _default_session_stack.get_default()
def get_session_by_id(handle): """Return the session by handle.""" if handle not in _session_dict: raise ValueError(f"Session {handle} not exists.") return _session_dict.get(handle) class _DefaultSessionStack(object): """A stack of objects for providing implicit defaults.""" def __init__(self): super().__init__() self.stack = [] def get_default(self) -> Session: if not self.stack: logger.info("Creating default session ...") sess = session( cluster_type="hosts", num_workers=gs_config.session.default_local_num_workers, ) sess.as_default() return self.stack[-1] def empty(self) -> bool: return len(self.stack) == 0 def reset(self): self.stack = [] def is_cleared(self): return not self.stack @contextlib.contextmanager def get_controller(self, default): """A context manager for manipulating a default stack.""" self.stack.append(default) try: yield default finally: # stack may be empty if reset() was called if self.stack: self.stack.remove(default) _default_session_stack = _DefaultSessionStack() # pylint: disable=protected-access
[docs]def g( incoming_data=None, oid_type="int64", vid_type="uint64", directed=True, generate_eid=True, retain_oid=True, vertex_map="global", compact_edges=False, use_perfect_hash=False, ): """Construct a GraphScope graph object on the default session. It will launch and set a session to default when there is no default session found. See params detail in :class:`graphscope.framework.graph.GraphDAGNode` Returns: :class:`graphscope.framework.graph.GraphDAGNode`: Evaluated in eager mode. Examples: .. code:: python >>> import graphscope >>> g = graphscope.g() >>> import graphscope >>> sess = graphscope.session() >>> sess.as_default() >>> g = graphscope.g() # creating graph on the session "sess" """ return get_default_session().g( incoming_data, oid_type, vid_type, directed, generate_eid, retain_oid, vertex_map, compact_edges, use_perfect_hash, )
@deprecated("Please use `graphscope.interactive` instead.") def gremlin(graph, params=None): """This method is going to be deprecated in the future. Use :meth:`graphscope.interactive` instead. """ return interactive(graph, params) def interactive(graph, params=None, with_cypher=False): """Create an interactive engine and get the handler to execute gremlin and cypher queries. See params detail in :meth:`graphscope.Session.interactive` Returns: :class:`graphscope.interactive.query.InteractiveQueryDAGNode`: InteractiveQuery to execute gremlin and cypher queries, evaluated in eager mode. Examples: .. code:: python >>> import graphscope >>> g = graphscope.g() >>> interactive_query = graphscope.interactive() """ assert graph is not None, "graph cannot be None" sess = graph._session # pylint: disable=protected-access assert sess is not None, "The graph object is invalid" return sess.interactive(graph, params, with_cypher)
[docs]def graphlearn(graph, nodes=None, edges=None, gen_labels=None): """Create a graph learning engine. See params detail in :meth:`graphscope.Session.graphlearn` Returns: :class:`graphscope.learning.GraphDAGNode`: An instance of learning graph that could be feed to the learning engine, evaluated in eager node. Example: .. code:: python >>> import graphscope >>> g = graphscope.g() >>> lg = graphscope.learning(g) """ assert graph is not None, "graph cannot be None" assert ( graph._session is not None ), "The graph object is invalid" # pylint: disable=protected-access return graph._session.graphlearn( graph, nodes, edges, gen_labels ) # pylint: disable=protected-access
def graphlearn_torch( graph, edges, edge_weights=None, node_features=None, edge_features=None, node_labels=None, edge_dir="out", random_node_split=None, num_clients=1, manifest_path=None, client_folder_path="./", return_pyg_remote_backend=False, ): assert graph is not None, "graph cannot be None" assert ( graph._session is not None ), "The graph object is invalid" # pylint: disable=protected-access return graph._session.graphlearn_torch( graph, edges, edge_weights, node_features, edge_features, node_labels, edge_dir, random_node_split, num_clients, manifest_path, client_folder_path, return_pyg_remote_backend, ) # pylint: disable=protected-access