jaseci 1.4.2.6__py3-none-any.whl → 2.0.0__py3-none-any.whl

jaseci/utils/actions/actions_optimizer.py

@@ -1,517 +0,0 @@
-"""
-Module that manage and optimizes the actions configuration of Jaseci
-"""
-
-from jaseci.jsorc.jsorc import JsOrc
-from jaseci.extens.svc.kube_svc import KubeService
-from jaseci.jsorc.remote_actions import ACTIONS_SPEC_LOC
-from jaseci.utils.utils import logger
-from jaseci.jsorc.live_actions import (
-    load_module_actions,
-    unload_module,
-    unload_remote_actions,
-    load_remote_actions,
-    live_actions,
-    action_configs,
-)
-
-import requests
-import copy
-import time
-
-from .actions_state import ActionsState
-
-POLICIES = ["Default", "Evaluation"]
-THRESHOLD = 0.2
-NODE_MEM_THRESHOLD = 0.8
-
-
-class ActionsOptimizer:
-    def __init__(
-        self,
-        namespace: str = "default",
-        policy: str = "Default",
-        benchmark: dict = {},
-        actions_history: dict = {},
-        actions_calls: dict = {},
-    ) -> None:
-        self.actions_state = ActionsState()
-        self.actions_change = {}
-        self.jsorc_interval = 0
-        self.namespace = namespace
-        self.policy = policy
-        self.benchmark = benchmark
-        self.actions_history = actions_history
-        self.actions_calls = actions_calls
-        self.policy_params = {}
-        self.policy_state = {}
-        self.last_eval_configs = []
-
-    def kube_create(self, config):
-        kube = JsOrc.svc("kube").poke(cast=KubeService)
-        for kind, conf in config.items():
-            name = conf["metadata"]["name"]
-            kube.create(kind, name, conf, kube.namespace, "ActionsOptimzer:")
-
-    def kube_delete(self, config):
-        kube = JsOrc.svc("kube").poke(cast=KubeService)
-        for kind, conf in config.items():
-            name = conf["metadata"]["name"]
-            kube.delete(kind, name, kube.namespace, "ActionsOptimzer:")
-
-    def get_actions_status(self, name=""):
-        """
-        Return the state of action
-        """
-        if name == "":
-            return self.actions_state.get_all_state()
-        else:
-            return self.actions_state.get_state(name)
-
-    def retire_remote(self, name):
-        """
-        Retire a microservice through the kube service
-        """
-        config = action_configs[name]["remote"]
-        self.kube_delete(config)
-        self.actions_state.remove_remote(name)
-
-    def spawn_remote(self, name):
-        """
-        Spawn a microservice through the kube service
-        """
-        config = action_configs[name]["remote"]
-        self.kube_create(config)
-        url = f"http://{config['Service']['metadata']['name']}/"
-        return url
-
-    def call_action(self, action_name, *params):
-        """
-        Call an action via live_actions
-        """
-        func = live_actions[action_name]
-        func(*params)
-
-    def action_prep(self, name):
-        """
-        Any action preparation that needs to be called right after action is loaded
-        """
-        pass
-
-    def load_action_remote(self, name, unload_existing=False):
-        """
-        Load a remote action.
-        JSORC will get the URL of the remote microservice
-        and stand up a microservice if there isn't currently one in the cluster.
-        Return True if the remote action is loaded successfully,
-        False otherwise
-        """
-        cur_state = self.actions_state.get_state(name)
-        if cur_state is None:
-            cur_state = self.actions_state.init_state(name)
-
-        if cur_state["mode"] == "remote" and cur_state["remote"]["status"] == "READY":
-            # Check if there is already a remote action loaded
-            return True
-
-        url = self.actions_state.get_remote_url(name)
-        if url is None:
-            # Spawn a remote microservice
-            url = self.spawn_remote(name)
-            self.actions_state.start_remote_service(name, url)
-            cur_state = self.actions_state.get_state(name)
-
-        if cur_state["remote"]["status"] == "STARTING":
-            if_ready = self.remote_action_ready_check(name, url)
-            if if_ready:
-                self.actions_state.set_remote_action_ready(name)
-                cur_state = self.actions_state.get_state(name)
-
-        if cur_state["remote"]["status"] == "READY":
-            if unload_existing:
-                self.unload_action_module(name)
-            load_remote_actions(url)
-            self.action_prep(name)
-            self.actions_state.remote_action_loaded(name)
-            return True
-
-        return False
-
-    def load_action_module(self, name, unload_existing=False):
-        """
-        Load an action module
-        """
-        cur_state = self.actions_state.get_state(name)
-        logger.info(cur_state)
-        if cur_state is None:
-            cur_state = self.actions_state.init_state(name)
-
-        if cur_state["mode"] == "module":
-            logger.info("ALREADY A MODULE LOADED")
-            # Check if there is already a local action loaded
-            return
-
-        if name not in action_configs:
-            return
-
-        module = action_configs[name]["module"]
-        loaded_module = action_configs[name]["loaded_module"]
-        if unload_existing:
-            self.unload_action_remote(name)
-
-        load_module_actions(module, loaded_module)
-        self.action_prep(name)
-        self.actions_state.module_action_loaded(name, module, loaded_module)
-
-    def unload_action_auto(self, name):
-        """
-        Unload an action based on how it is currently loaded
-        """
-        cur_state = self.actions_state.get_state(name)
-        if cur_state is None:
-            return False, "Action is not loaded."
-        if cur_state["mode"] == "module":
-            return self.unload_action_module(name)
-        elif cur_state["mode"] == "remote":
-            return self.unload_action_remote(name)
-        return False, f"Unrecognized action loaded status {cur_state['mode']}"
-
-    def unload_action_module(self, name):
-        """
-        Unload an action module
-        """
-        cur_state = self.actions_state.get_state(name)
-        if cur_state is None:
-            return False, "Action is not loaded."
-
-        if cur_state["mode"] != "module":
-            return False, "Action is not loaded as module."
-
-        module_name = cur_state["module"]["name"]
-        loaded_module = cur_state["module"]["loaded_module"]
-
-        unload_module(module_name)
-        unload_module(loaded_module)
-        self.actions_state.module_action_unloaded(name)
-
-        return (True, f"Action module {name} unloaded.")
-
-    def unload_action_remote(self, name):
-        """
-        Unload a remote action
-        """
-        cur_state = self.actions_state.get_state(name)
-        if cur_state is None:
-            return False, "Action is not loaded."
-
-        if cur_state["mode"] != "remote":
-            return False, "Action is not loaded as remote."
-
-        if cur_state["remote"]["status"] != "READY":
-            return False, "Remote action is not ready."
-
-        # Get the list of actions from the action spec of the server
-        url = cur_state["remote"]["url"]
-
-        unload_remote_actions(url)
-        self.actions_state.remote_action_unloaded(name)
-
-        return (True, f"Remote actions from {url} unloaded.")
-
-    def remote_action_ready_check(self, name, url):
-        """
-        Check if a remote action is ready by querying the action_spec endpoint
-        """
-        if url is None:
-            return False
-        spec_url = url.rstrip("/") + ACTIONS_SPEC_LOC
-        headers = {"content-type": "application/json"}
-        try:
-            res = requests.get(spec_url, headers=headers, timeout=1)
-        except (requests.exceptions.Timeout, requests.exceptions.ConnectionError):
-            # Remote service not ready yet
-            return False
-
-        return res.status_code == 200
-
-    def set_action_policy(self, policy_name: str, policy_params: dict = {}):
-        """
-        Set the action optimization policy for JSORC
-        """
-        # TODO: manage policy switching if there are unresolved actions state
-        if policy_name in POLICIES:
-            self.policy = policy_name
-            self.policy_state[policy_name] = {}
-            self.policy_params = policy_params
-            return True
-        else:
-            return f"Policy {policy_name} not found."
-
-    def get_action_policy(self):
-        """
-        Return the currently active action policy
-        """
-        return self.policy
-
-    def run(self, jsorc_interval: int):
-        """
-        The main optimization function.
-        This gets invoked by JSROC regularly at a configured interval.
-        """
-        self.jsorc_interval = jsorc_interval
-        if self.policy == "Default":
-            # Default policy does not manage action automatically
-            return
-        elif self.policy == "Evaluation":
-            self._actionpolicy_evaluation()
-        if len(self.actions_change) > 0:
-            self.apply_actions_change()
-
-    def _init_evalution_policy(self, policy_state):
-        # 999 is just really large memory size so everything can fits in local
-        node_mem = self.policy_params.get("node_mem", 999 * 1024)
-        jaseci_runtime_mem = self.policy_params.get("jaseci_runtime_mem", 300)
-        # Initialize configs to eval
-        actions = self.actions_state.get_active_actions()
-        # construct list of possible configurations
-        all_configs = [{"local_mem": jaseci_runtime_mem}]
-        for act in actions:
-            new_configs = []
-            for con in all_configs:
-                for m in ["local", "remote"]:
-                    c = copy.deepcopy(con)
-                    c[act] = m
-                    if m == "local":
-                        local_mem_requirement = action_configs[act][
-                            "local_mem_requirement"
-                        ]
-                        c["local_mem"] = c["local_mem"] + local_mem_requirement
-                        if c["local_mem"] < (node_mem * NODE_MEM_THRESHOLD):
-                            new_configs.append(dict(c))
-                        else:
-                            logger.info(
-                                f"config dropped for memory constraint: {c},\n\tcurrent node memory: {node_mem}\n\tavailable memory: {(node_mem * NODE_MEM_THRESHOLD)-c['local_mem'] }"
-                            )
-                    else:
-                        new_configs.append(dict(c))
-            all_configs = list(new_configs)
-        policy_state["remain_configs"] = all_configs
-
-    def _actionpolicy_evaluation(self):
-        """
-        A evaluation based policy.
-        JSORC cycle through possible action configurations and evaluate request performance and select the one with the best performance.
-        Use the post_request_hook from JSORC to track request performance
-        """
-        logger.info("===Evaluation Policy===")
-        policy_state = self.policy_state["Evaluation"]
-
-        if len(policy_state) == 0:
-            # Initialize policy tracking state
-            policy_state = {
-                "phase": "eval",  # current phase of policy: eval|perf
-                "cur_config": None,  # current active configuration
-                "remain_configs": [],  # remaining configurations that need to be evaluated
-                "past_configs": [],  # configurations already evaluated
-                "eval_phase": self.policy_params.get(
-                    "eval_phase", 10
-                ),  # how long is evaluatin period (in seconds)
-                "perf_phase": self.policy_params.get(
-                    "perf_phase", 100
-                ),  # how long is the performance period (in seconds)
-                "cur_phase": 0,  # how long the current period has been running
-                "prev_best_config": self.actions_state.get_all_state(),
-            }
-        policy_state["cur_phase"] += self.jsorc_interval
-
-        # check if we should go into evaluation phase
-        if (
-            policy_state["phase"] == "perf"
-            and policy_state["cur_phase"] >= policy_state["perf_phase"]
-        ):
-            # if no enough walker were execueted in this period, keep in perf phase
-            if "walker_run" not in self.benchmark["requests"]:
-                policy_state["cur_phase"] = 0
-            else:
-                logger.info("===Evaluation Policy=== Switching to evaluation mode")
-                policy_state["phase"] = "eval"
-                policy_state["cur_phase"] = 0
-                policy_state["cur_config"] = None
-                if len(policy_state["remain_configs"]) == 0:
-                    self._init_evalution_policy(policy_state)
-        if policy_state["phase"] == "eval":
-            # In evaluation phase
-            if policy_state["cur_config"] is None:
-                self._init_evalution_policy(policy_state)
-
-                # This is the start of evaluation period
-                policy_state["cur_config"] = policy_state["remain_configs"][0]
-                del policy_state["remain_configs"][0]
-                policy_state["cur_phase"] = 0
-                self.benchmark["active"] = True
-                self.benchmark["requests"] = {}
-                self.actions_change = self._get_action_change(
-                    policy_state["cur_config"]
-                )
-                if len(self.actions_change) > 0:
-                    logger.info(
-                        f"===Evaluation Policy=== Switching eval config to {policy_state['cur_config']}"
-                    )
-                    policy_state["phase"] = "eval_switching"
-                    self.benchmark["active"] = False
-            else:
-                if policy_state["cur_phase"] >= policy_state["eval_phase"]:
-                    # The eval phase for the current configuration is complete
-                    # Get performance
-                    if "walker_run" not in self.benchmark["requests"]:
-                        # meaning no incoming requests during this period.
-                        # stay in this phase
-                        logger.info(f"===Evaluation Policy=== No walkers were executed")
-                        self.policy_state["Evaluation"] = policy_state
-                        return
-
-                    walker_runs = []
-                    for walker, times in self.benchmark["requests"][
-                        "walker_run"
-                    ].items():
-                        if walker == "_default_":
-                            continue
-                        else:
-                            walker_runs.extend(times)
-
-                    avg_walker_lat = sum(walker_runs) / len(walker_runs)
-                    policy_state["cur_config"]["avg_walker_lat"] = avg_walker_lat
-                    policy_state["past_configs"].append(policy_state["cur_config"])
-                    logger.info(
-                        f"===Evaluation Policy=== Complete evaluation period for {policy_state['cur_config']} latency: {avg_walker_lat}"
-                    )
-
-                    # check if all configs have been evaluated
-                    if len(policy_state["remain_configs"]) == 0:
-                        # best config is the one with the fastest walker latency during the evaluation period
-                        logger.info(f"===Evaluation Policy=== Evaluation phase over. ")
-                        best_config = min(
-                            policy_state["past_configs"],
-                            key=lambda x: x["avg_walker_lat"],
-                        )
-                        # Switch the system to the best config
-                        del best_config["avg_walker_lat"]
-                        self.actions_change = self._get_action_change(best_config)
-
-                        # ADAPTIVE: if the selected best config is the same config as the previous best one, double the performance period
-                        if all(
-                            [
-                                best_config[act]
-                                == policy_state["prev_best_config"][act]["mode"]
-                                for act in best_config.keys()
-                                if act in action_configs.keys()
-                            ]
-                        ):
-                            policy_state["perf_phase"] *= 2
-                            logger.info(
-                                f"===Evaluation Policy=== Best config is the same as previous one. Doubling performance phase to {policy_state['perf_phase']}"
-                            )
-
-                        policy_state["phase"] = "perf"
-                        policy_state["cur_config"] = None
-                        policy_state["past_configs"] = []
-                        policy_state["cur_phase"] = 0
-                        self.benchmark["requests"] = {}
-                        self.benchmark["active"] = True
-                        logger.info(
-                            f"===Evaluation Policy=== Evaluation phase over. Selected best config as {best_config}"
-                        )
-                    else:
-                        next_config = policy_state["remain_configs"][0]
-                        del policy_state["remain_configs"][0]
-                        self.actions_change = self._get_action_change(next_config)
-                        policy_state["cur_config"] = next_config
-                        policy_state["cur_phase"] = 0
-                        self.benchmark["requests"] = {}
-                        if len(self.actions_change) > 0:
-                            logger.info(
-                                f"===Evaluation Policy=== Switching eval config to {policy_state['cur_config']}"
-                            )
-                            policy_state["phase"] = "eval_switching"
-                            self.benchmark["active"] = False
-                        else:
-                            policy_state["phase"] = "eval"
-                            self.benchmark["active"] = True
-                        logger.info(
-                            f"===Evaluation Policy=== Switching to next config to evaluate {next_config}"
-                        )
-        elif policy_state["phase"] == "eval_switching":
-            # in the middle of switching between configs for evaluation
-            if len(self.actions_change) == 0:
-                # this means all actions change have been applied, start evaluation phase
-                logger.info(
-                    f"===Evaluation Policy=== All actions change have been applied. Start evaluation phase."
-                )
-                policy_state["phase"] = "eval"
-                policy_state["cur_phase"] = 0
-                self.benchmark["active"] = True
-                self.benchmark["requests"] = {}
-        self.policy_state["Evaluation"] = policy_state
-
-    def _get_action_change(self, new_action_state):
-        """
-        Given a new desired action state and the current action_state tracking, return the change set
-        """
-        change_state = {}
-        for name, new_state in new_action_state.items():
-            if name not in action_configs.keys():
-                continue
-            cur = self.actions_state.get_state(name)
-            if cur is None:
-                cur = self.actions_state.init_state(name)
-            if new_state == "local":
-                new_state = "module"
-            if new_state != cur["mode"]:
-                change_str = (
-                    f"{cur['mode'] if cur['mode'] is not None else ''}_to_{new_state}"
-                )
-                change_state[name] = change_str
-        return change_state
-
-    def apply_actions_change(self):
-        """
-        Apply any action configuration changes
-        """
-        actions_change = dict(self.actions_change)
-        # For now, to_* and *_to_* are the same logic
-        # But this might change down the line
-        for name, change_type in actions_change.items():
-            logger.info(f"==Actions Optimizer== Changing {name} {change_type}")
-            if change_type in ["to_local", "_to_local", "_to_module", "to_module"]:
-                # Switching from no action loaded to local
-                self.load_action_module(name)
-                del self.actions_change[name]
-            elif change_type == "to_remote":
-                loaded = self.load_action_remote(name)
-                if loaded:
-                    del self.actions_change[name]
-            elif change_type == "local_to_remote" or change_type == "module_to_remote":
-                # loaded = self.load_action_remote(name, unload_existing=True)
-                loaded = self.load_action_remote(name)
-                if loaded:
-                    del self.actions_change[name]
-            elif change_type == "remote_to_local" or change_type == "remote_to_module":
-                # self.load_action_module(name, unload_existing=True)
-                self.load_action_module(name)
-                del self.actions_change[name]
-
-        if len(actions_change) > 0 and self.actions_history["active"]:
-            # Summarize action stats during this period and add to previous state
-            self.summarize_action_calls()
-            self.actions_history["history"].append(
-                {"ts": time.time(), "actions_state": self.actions_state.get_all_state()}
-            )
-
-    def summarize_action_calls(self):
-        actions_summary = {}
-        for action_name, calls in self.actions_calls.items():
-            actions_summary[action_name] = sum(calls) / len(calls)
-        self.actions_calls.clear()
-
-        if len(self.actions_history["history"]) > 0:
-            self.actions_history["history"][-1]["actions_calls"] = actions_summary

jaseci/utils/actions/actions_state.py

@@ -1,95 +0,0 @@
-class ActionsState:
-    """
-    An object tracking the various states of available actions. Utilized by the ActionsOptimizer and ActionsOptimizerPolicy
-    """
-
-    def __init__(self):
-        """
-        Actions state format:
-        actions_state = {
-            "ACTION_NAME": {
-                "mode": "local|remote",
-                "module": {
-                    "name": MDOULE_NAME
-                },
-                "local": {
-                    "path": LOCAL MODULE FILE PATH
-                },
-                "remote": {
-                    "url": "remote_url of the action microservice",
-                    "status": "READY|STARTING|RETIRING"
-                }
-            }
-        }
-        """
-        self.state = {}
-
-    def init_state(self, name):
-        """
-        Initialize the data structure for the action state tracking for a specific action
-        """
-        self.state[name] = {
-            "mode": None,
-            "module": {"name": None, "loaded_module": None},
-            "local": {"path": None},
-            "remote": {"url": None, "status": None},
-        }
-
-        return self.state[name]
-
-    def get_active_actions(self):
-        """
-        Return list of action modules currently used
-        """
-        return list(self.state.keys())
-
-    def local_action_loaded(self, name):
-        self.state[name]["mode"] = "local"
-
-    def module_action_loaded(self, name, module, loaded):
-        self.state[name]["mode"] = "module"
-        self.state[name]["module"]["name"] = module
-        self.state[name]["module"]["loaded_module"] = loaded
-
-    def module_action_unloaded(self, name):
-        self.state[name]["mode"] = None
-        self.state[name]["module"]["name"] = None
-
-    def remote_action_unloaded(self, name):
-        self.state[name]["mode"] = None
-
-    def remove_remote(self, name):
-        self.state[name]["remote"] = {"url": None, "status": None}
-
-    def remote_action_loaded(self, name):
-        self.state[name]["mode"] = "remote"
-
-    def start_remote_service(self, name, url):
-        self.state[name]["remote"] = {"status": "STARTING", "url": url}
-
-    def set_remote_action_ready(self, name):
-        self.state[name]["remote"]["status"] = "READY"
-
-    def stop_remote_action(self, name):
-        self.state[name]["remote"] = {}
-
-    def get_remote_url(self, name):
-        if name not in self.state or "remote" not in self.state[name]:
-            return None
-        return self.state[name]["remote"].get("url", None)
-
-    def get_all_state(self):
-        ret = {}
-        for name, state in self.state.items():
-            ret[name] = {"mode": state["mode"]}
-        return ret
-
-    def get_state(self, name):
-        return self.state.get(name, None)
-
-    def set_url(self, name, url):
-        if name not in self.state:
-            self.state[name] = {}
-        if "remote" not in self.state[name]:
-            self.state[name]["remote"] = {}
-        self.state[name]["remote"]["url"] = url