guidellm 0.4.0a21__py3-none-any.whl → 0.4.0a155__py3-none-any.whl

guidellm/scheduler/worker_group.py

@@ -0,0 +1,707 @@
+"""
+Multi-process worker group orchestration for distributed request scheduling.
+
+Provides infrastructure for coordinating worker processes with shared state
+management, inter-process communication, and lifecycle coordination. Handles
+dynamic scaling, load balancing, constraint evaluation, and graceful shutdown
+across distributed workers processing concurrent requests.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import math
+import threading
+import time
+import uuid
+from collections.abc import AsyncIterator, Generator, Iterable
+from multiprocessing import get_context
+from multiprocessing.context import BaseContext
+from multiprocessing.managers import BaseManager
+from multiprocessing.process import BaseProcess
+from multiprocessing.synchronize import Barrier, Event
+from typing import Generic, NamedTuple
+
+from guidellm.logger import logger
+from guidellm.scheduler.constraints import Constraint, RequestsExhaustedConstraint
+from guidellm.scheduler.schemas import (
+    BackendInterface,
+    MultiTurnRequestT,
+    RequestT,
+    ResponseT,
+    SchedulerState,
+    SchedulerUpdateAction,
+)
+from guidellm.scheduler.strategies import SchedulingStrategy
+from guidellm.scheduler.worker import WorkerProcess
+from guidellm.schemas import RequestInfo
+from guidellm.settings import settings
+from guidellm.utils import (
+    InterProcessMessaging,
+    InterProcessMessagingManagerQueue,
+    InterProcessMessagingPipe,
+    InterProcessMessagingQueue,
+    wait_for_sync_objects,
+)
+
+__all__ = ["WorkerGroupState", "WorkerProcessGroup"]
+
+
+class WorkerProcessGroup(Generic[RequestT, ResponseT]):
+    """
+    Orchestrates multiple worker processes for distributed request processing.
+
+    Manages process lifecycle, request distribution, response collection, and state
+    synchronization across workers. Handles dynamic scaling, load balancing, and
+    constraint evaluation with graceful shutdown coordination for high-throughput
+    request processing workloads.
+
+    Example:
+    ::
+        from guidellm.scheduler.worker_group import WorkerProcessGroup
+
+        group = WorkerProcessGroup(
+            requests=request_iterable,
+            backend=backend_instance,
+            strategy=scheduling_strategy,
+            constraints={"max_time": time_constraint}
+        )
+
+        await group.create_processes()
+        await group.start(time.time())
+
+        async for response, request, info, state in group.request_updates():
+            if response is not None:
+                # Process completed request
+                handle_response(response)
+
+        await group.shutdown()
+    """
+
+    def __init__(
+        self,
+        requests: Iterable[RequestT | MultiTurnRequestT[RequestT]],
+        backend: BackendInterface[RequestT, ResponseT],
+        strategy: SchedulingStrategy,
+        startup_duration: float,
+        **constraints: dict[str, Constraint],
+    ):
+        """
+        Initialize a worker process group for distributed request processing.
+
+        :param requests: Finite iterable of requests to process sequentially
+        :param backend: Backend interface for processing requests
+        :param strategy: Scheduling strategy for request timing and distribution
+        :param startup_duration: Duration in seconds for request startup ramping
+        :param constraints: Named constraints for controlling execution behavior
+        """
+        self.requests = requests
+        self.backend = backend
+        self.strategy = strategy
+        self.startup_duration = startup_duration
+        self.constraints = constraints
+
+        # Multiprocessing contexts and primitives, created in create_processes
+        self.mp_context: BaseContext | None = None
+        self.mp_manager: BaseManager | None = None
+        self.processes: list[BaseProcess] | None = None
+        self.startup_barrier: Barrier | None = None
+        self.requests_generated_event: Event | None = None
+        self.constraint_reached_event: Event | None = None
+        self.shutdown_event: Event | None = None
+        self.error_event: Event | None = None
+
+        # Scheduler and messaging state, created in start
+        self.state: WorkerGroupState[RequestT, ResponseT] | None = None
+        self.messaging: (
+            InterProcessMessaging[
+                tuple[
+                    RequestT | MultiTurnRequestT[RequestT],
+                    RequestInfo,
+                ],
+                tuple[
+                    ResponseT | None,
+                    RequestT | MultiTurnRequestT[RequestT],
+                    RequestInfo,
+                    SchedulerState,
+                ],
+            ]
+            | None
+        ) = None
+
+    async def create_processes(self):
+        """
+        Create and initialize worker processes for distributed request processing.
+
+        Sets up multiprocessing infrastructure and worker processes based on
+        strategy constraints, backend capabilities, and system configuration.
+        Determines optimal process count and concurrency limits, then spawns
+        worker processes with distributed request handling capabilities.
+
+        :raises RuntimeError: If process initialization or startup fails
+        """
+        # Processes limits and params
+        max_conc: int
+        if (
+            requests_limit := min(
+                self.strategy.requests_limit or math.inf,
+                self.backend.requests_limit or math.inf,
+            )
+        ) != math.inf:
+            max_conc = int(requests_limit)
+        else:
+            # If concurrency not specified, use settings
+            max_conc = settings.max_concurrency
+        if max_conc <= 0:
+            raise RuntimeError("max_concurrency resolved to 0; increase limits/config")
+
+        # Calculate number of processes, ensure we don't exceed the max concurrency,
+        # or limits from the backend, strategy, or user settings
+        num_processes: int = int(
+            min(
+                max_conc,
+                self.strategy.processes_limit or math.inf,
+                self.backend.processes_limit or math.inf,
+                settings.max_worker_processes,
+            )
+        )
+        if num_processes <= 0:
+            raise RuntimeError("num_processes resolved to 0; increase limits/config")
+
+        per_proc_max_conc = max_conc // num_processes
+        max_pending_size = max(
+            1, math.floor(max_conc * settings.mp_max_pending_buffer_percent)
+        )
+        per_proc_max_buffer_size = 1
+
+        # Initialize multiprocessing components
+        self.mp_context = get_context(settings.mp_context_type)
+        self.mp_manager = self.mp_context.Manager()
+        self.startup_barrier = self.mp_context.Barrier(num_processes + 1)
+        self.requests_generated_event = self.mp_context.Event()
+        self.constraint_reached_event = self.mp_context.Event()
+        self.shutdown_event = self.mp_context.Event()
+        self.error_event = self.mp_context.Event()
+
+        if settings.mp_messaging_object == "queue":
+            self.messaging = InterProcessMessagingQueue(
+                mp_context=self.mp_context,
+                serialization=settings.mp_serialization,
+                encoding=settings.mp_encoding,
+                max_pending_size=max_pending_size,
+                max_buffer_send_size=settings.mp_requests_send_buffer_size,
+                poll_interval=settings.mp_poll_interval,
+            )
+        elif settings.mp_messaging_object == "manager_queue":
+            self.messaging = InterProcessMessagingManagerQueue(
+                manager=self.mp_manager,
+                mp_context=self.mp_context,
+                serialization=settings.mp_serialization,
+                encoding=settings.mp_encoding,
+                max_pending_size=max_pending_size,
+                max_buffer_send_size=settings.mp_requests_send_buffer_size,
+                poll_interval=settings.mp_poll_interval,
+            )
+        elif settings.mp_messaging_object == "pipe":
+            self.messaging = InterProcessMessagingPipe(
+                num_workers=num_processes,
+                mp_context=self.mp_context,
+                serialization=settings.mp_serialization,
+                encoding=settings.mp_encoding,
+                max_pending_size=max_pending_size,
+                max_buffer_send_size=settings.mp_requests_send_buffer_size,
+                poll_interval=settings.mp_poll_interval,
+            )
+
+        # Initialize worker processes
+        self.processes = []
+        self.strategy.init_processes_timings(
+            worker_count=num_processes,
+            max_concurrency=max_conc,
+            startup_duration=self.startup_duration,
+        )
+        for rank in range(num_processes):
+            # Distribute any remainder across the first N ranks
+            async_limit = per_proc_max_conc + (
+                1 if rank < (max_conc % num_processes) else 0
+            )
+
+            worker = WorkerProcess[RequestT, ResponseT](
+                worker_index=rank,
+                messaging=self.messaging.create_worker_copy(
+                    worker_index=rank,
+                    max_buffer_send_size=None,
+                    max_buffer_receive_size=per_proc_max_buffer_size,
+                ),
+                backend=self.backend,
+                strategy=self.strategy,
+                async_limit=async_limit,
+                fut_scheduling_time_limit=0.0,
+                startup_barrier=self.startup_barrier,
+                requests_generated_event=self.requests_generated_event,
+                constraint_reached_event=self.constraint_reached_event,
+                shutdown_event=self.shutdown_event,
+                error_event=self.error_event,
+            )
+            proc = self.mp_context.Process(target=worker.run, daemon=False)
+            proc.start()
+            self.processes.append(proc)
+
+        wait_key = await wait_for_sync_objects(
+            {
+                "startup_barrier": self.startup_barrier,
+                "shutdown_event": self.shutdown_event,
+                "error_event": self.error_event,
+            },
+            poll_interval=settings.mp_poll_interval,
+        )
+
+        if wait_key == "error_event":
+            raise RuntimeError(
+                "Worker process group startup failed: error_event is set"
+            )
+
+    async def start(self, start_time: float):
+        """
+        Begin request processing at the specified start time.
+
+        Initializes scheduler state and background tasks, then waits until the
+        specified start time before beginning operations. Sets up inter-process
+        communication and coordinates synchronized startup across all workers.
+
+        :param start_time: Unix timestamp when processing should begin
+        :raises RuntimeError: If workers encounter errors during startup or
+            if create_processes() was not called first
+        """
+        if (
+            not self.processes
+            or not self.requests_generated_event
+            or not self.constraint_reached_event
+            or not self.shutdown_event
+            or not self.error_event
+            or not self.messaging
+        ):
+            raise RuntimeError("create_processes() must be called before start()")
+
+        self.strategy.init_processes_start(start_time=start_time)
+        stop_send_requests_event = threading.Event()
+        send_requests_stopped_event = threading.Event()
+        self.state = WorkerGroupState[RequestT, ResponseT](
+            start_time=start_time,
+            processes=self.processes,
+            constraints=self.constraints,
+            stop_send_requests_event=stop_send_requests_event,
+            send_requests_stopped_event=send_requests_stopped_event,
+            requests_generated_event=self.requests_generated_event,
+            constraint_reached_event=self.constraint_reached_event,
+            shutdown_event=self.shutdown_event,
+            error_event=self.error_event,
+            messaging=self.messaging,
+        )
+        await self.messaging.start(
+            send_items=self.state.requests_generator(self.requests),
+            receive_callback=self.state.received_callback,
+            send_stopped_event=send_requests_stopped_event,
+            send_stop_criteria=[stop_send_requests_event],
+            receive_stop_criteria=[self.shutdown_event],
+        )
+
+        if (wait_time := start_time - time.time()) > 0:
+            await asyncio.sleep(wait_time)
+        if self.error_event.is_set():
+            raise RuntimeError(
+                "error_event is set in WorkerProcessGroup, "
+                "indicating an error occurred in one of the worker processes."
+            )
+
+    async def request_updates(
+        self,
+    ) -> AsyncIterator[
+        tuple[
+            ResponseT | None,
+            RequestT | MultiTurnRequestT[RequestT],
+            RequestInfo,
+            SchedulerState,
+        ]
+    ]:
+        """
+        Yield request processing updates as they become available.
+
+        Returns an async iterator of request updates including response, request,
+        request scheduling info, and scheduler state. Updates occur on request queued,
+        processing start, and completion. Response is None until processing completes.
+
+        :return: Async iterator yielding (response, request, request_info, state)
+            tuples where response is None until processing is complete
+        :raises RuntimeError: If workers encounter unrecoverable errors
+        """
+        while True:
+            if self.error_event.is_set():  # type: ignore[union-attr]
+                logger.error("Error event set in WorkerProcessGroup")
+                raise RuntimeError(
+                    "error_event is set in WorkerProcessGroup, "
+                    "indicating an error occurred in one of the worker processes."
+                )
+
+            try:
+                (
+                    response,
+                    request,
+                    request_info,
+                    scheduler_state,
+                ) = await self.messaging.get(timeout=settings.mp_poll_interval)  # type: ignore[union-attr]
+
+                yield response, request, request_info, scheduler_state
+            except asyncio.TimeoutError:
+                if self.shutdown_event.is_set():  # type: ignore[union-attr]
+                    # Everything yielded, exit
+                    break
+
+    async def shutdown(self) -> list[Exception]:  # noqa: C901
+        """
+        Gracefully shut down the worker process group and clean up resources.
+
+        Performs safe shutdown of worker processes, background tasks, and
+        multiprocessing resources. Coordinates orderly termination across
+        all workers and collects any exceptions encountered during shutdown.
+
+        :return: List of exceptions encountered during shutdown; empty if no errors
+        """
+        exceptions: list[Exception] = []
+        if self.shutdown_event is not None:
+            self.shutdown_event.set()
+
+        # Clear out start values
+        if self.messaging is not None:
+            try:
+                await asyncio.wait_for(self.messaging.stop(), timeout=5.0)
+            except Exception as err:  # noqa: BLE001
+                exceptions.append(err)
+        self.messaging = None
+        self.state = None
+
+        # Clear out create processes values
+        if self.processes is not None:
+            for proc in self.processes:
+                try:
+                    await asyncio.to_thread(proc.join, timeout=5.0)
+                    if proc.exitcode is not None and proc.exitcode > 0:
+                        exceptions.append(
+                            RuntimeError(
+                                f"Worker {proc.pid} exited with code {proc.exitcode}"
+                            )
+                        )
+                except Exception as err:  # noqa: BLE001
+                    exceptions.append(err)
+        self.processes = None
+        self.startup_barrier = None
+        self.requests_generated_event = None
+        self.constraint_reached_event = None
+        self.shutdown_event = None
+        self.error_event = None
+        if self.mp_manager is not None:
+            try:
+                self.mp_manager.shutdown()
+            except Exception as err:  # noqa: BLE001
+                exceptions.append(err)
+        self.mp_manager = None
+        self.mp_context = None
+
+        return exceptions
+
+
+class _StateUpdate(NamedTuple):
+    """Internal state update result with control flags."""
+
+    state: SchedulerState
+    stop_queueing: bool
+    stop_processing: bool
+
+
+class WorkerGroupState(Generic[RequestT, ResponseT]):
+    """
+    Manages scheduler state and synchronization for worker process groups.
+
+    Handles request generation, state updates, constraint evaluation, and
+    coordination between worker processes. Provides thread-safe state management
+    with request lifecycle tracking and constraint-based termination logic.
+    """
+
+    def __init__(
+        self,
+        start_time: float,
+        processes: list[BaseProcess],
+        constraints: dict[str, Constraint],
+        stop_send_requests_event: threading.Event,
+        send_requests_stopped_event: threading.Event,
+        requests_generated_event: Event,
+        constraint_reached_event: Event,
+        shutdown_event: Event,
+        error_event: Event,
+        messaging: InterProcessMessaging[
+            tuple[RequestT | MultiTurnRequestT[RequestT], RequestInfo],
+            tuple[
+                ResponseT | None,
+                RequestT | MultiTurnRequestT[RequestT],
+                RequestInfo,
+                SchedulerState,
+            ],
+        ],
+    ):
+        """
+        Initialize worker group state management.
+
+        :param start_time: Unix timestamp when processing should begin
+        :param processes: List of worker process instances
+        :param constraints: Named constraints for controlling execution behavior
+        :param stop_send_requests_event: Threading event for stopping request generation
+        :param send_requests_stopped_event: Threading event for request coordination
+        :param requests_generated_event: Multiprocessing event for generation completion
+        :param constraint_reached_event: Multiprocessing event for constraint stopping
+        :param shutdown_event: Multiprocessing event for coordinated shutdown
+        :param error_event: Multiprocessing event for error condition signaling
+        """
+        self.start_time = start_time
+        self.processes = processes
+        self.constraints = constraints
+        self.stop_send_requests_event = stop_send_requests_event
+        self.send_requests_stopped_event = send_requests_stopped_event
+        self.requests_generated_event = requests_generated_event
+        self.constraint_reached_event = constraint_reached_event
+        self.shutdown_event = shutdown_event
+        self.error_event = error_event
+        self.messaging = messaging
+
+        self._update_lock: threading.Lock = threading.Lock()
+        self._state: SchedulerState = SchedulerState(
+            node_id=0,
+            num_processes=len(processes),
+            start_time=start_time,
+        )
+        self._queued_requests: set[RequestT | MultiTurnRequestT[RequestT]] = set()
+        self._pending_requests: set[RequestT | MultiTurnRequestT[RequestT]] = set()
+        self._processing_requests: set[RequestT | MultiTurnRequestT[RequestT]] = set()
+
+    def requests_generator(
+        self, requests: Iterable[RequestT | MultiTurnRequestT[RequestT]]
+    ) -> Generator[
+        tuple[RequestT | MultiTurnRequestT[RequestT], RequestInfo], None, None
+    ]:
+        """
+        Generate request-info pairs for worker processing with constraint evaluation.
+
+        Processes finite requests sequentially then cycles through repeating requests
+        indefinitely. Creates scheduling metadata for each request and evaluates
+        constraints to determine when to stop request generation.
+
+        :param requests: Finite iterable of requests to process sequentially
+        :return: Generator yielding (request, request_info) tuples
+        """
+
+        try:
+            count = 0
+            for request in iter(requests):
+                count += 1
+
+                if hasattr(request, "request_id"):
+                    request_id = request.request_id
+                elif hasattr(request, "id"):
+                    request_id = request.id
+                else:
+                    request_id = str(uuid.uuid4())
+                request_info: RequestInfo = RequestInfo(
+                    request_id=request_id,
+                    status="queued",
+                    scheduler_process_id=0,
+                    scheduler_start_time=self.start_time,
+                )
+                state_update = self._locked_update(request_info)
+                request_info.timings.queued = time.time()
+                self.messaging.buffer_receive_queue.sync_put(
+                    (None, request, request_info, state_update.state)
+                )
+
+                yield (request, request_info)
+
+                if state_update.stop_queueing:
+                    self.stop_send_requests_event.set()
+                    return
+
+            # Reached the end, inject a RequestsExhaustedConstraint to record
+            self._locked_update(
+                info=None,
+                requests_exhausted={
+                    "requests_exhausted": RequestsExhaustedConstraint(
+                        num_requests=count
+                    )
+                },
+            )
+            self.stop_send_requests_event.set()
+        except Exception as err:
+            logger.error(f"Error generating requests: {err}")
+            self.error_event.set()
+            raise err
+
+    def received_callback(
+        self,
+        update: tuple[
+            ResponseT | None,
+            RequestT | MultiTurnRequestT,
+            RequestInfo,
+        ],
+    ) -> tuple[
+        ResponseT | None,
+        RequestT | MultiTurnRequestT,
+        RequestInfo,
+        SchedulerState,
+    ]:
+        """
+        Process received request updates and inject current scheduler state.
+
+        Updates internal state tracking based on request status changes and
+        evaluates constraints to determine if processing should be terminated.
+        Triggers shutdown when stop conditions are met.
+
+        :param update: Tuple containing response, request, and request info
+        :return: Updated tuple with injected scheduler state
+        """
+        try:
+            response, request, request_info = update
+            state_update = self._locked_update(info=request_info)
+
+            # Check if we need to tell workers to stop pulling new requests
+            # based on no more requests sent and all requests removed from queue
+            if (
+                state_update.state.queued_requests == 0
+                and self.stop_send_requests_event.is_set()
+                and not self.requests_generated_event.is_set()
+            ):
+                self.requests_generated_event.set()
+
+            # Check if we need to tell workers to stop processing requests (constraints)
+            if (
+                state_update.stop_processing
+                and not self.constraint_reached_event.is_set()
+            ):
+                self.constraint_reached_event.set()
+
+            # Check if all requests have been processed and can shutdown
+            if (
+                state_update.state.processed_requests
+                == state_update.state.created_requests
+                and self.stop_send_requests_event.is_set()
+                and self.requests_generated_event.is_set()
+                and self.constraint_reached_event.is_set()
+                and not self.shutdown_event.is_set()
+            ):
+                self.shutdown_event.set()
+        except Exception as err:
+            logger.error(f"Error processing received update: {err}")
+            self.error_event.set()
+            raise err
+
+        return (
+            response,
+            request,
+            request_info,
+            state_update.state,  # inject state for updates to be yielded back
+        )
+
+    def _locked_update(
+        self,
+        info: RequestInfo | None = None,
+        **add_constraints: dict[str, Constraint],
+    ) -> _StateUpdate:
+        with self._update_lock:
+            if add_constraints:
+                self.constraints.update(add_constraints)
+
+            if info is not None:
+                self._state.end_time = time.time()  # Always update in case last update
+                self._update_state_request_counts(info)
+                self._update_with_constraints(info)
+
+            state_copy: SchedulerState = self._state.model_copy()
+
+        return _StateUpdate(
+            state_copy,
+            state_copy.end_queuing_time is not None,
+            state_copy.end_processing_time is not None,
+        )
+
+    def _update_state_request_counts(self, info: RequestInfo):
+        if info.status == "queued":
+            self._queued_requests.add(info.request_id)
+            self._state.queued_requests = len(self._queued_requests)
+            self._state.created_requests += 1
+        elif info.status == "pending":
+            self._queued_requests.remove(info.request_id)
+            self._state.queued_requests = len(self._queued_requests)
+            self._pending_requests.add(info.request_id)
+            self._state.pending_requests = len(self._pending_requests)
+        elif info.status == "in_progress":
+            self._pending_requests.remove(info.request_id)
+            self._state.pending_requests = len(self._pending_requests)
+            self._processing_requests.add(info.request_id)
+            self._state.processing_requests = len(self._processing_requests)
+        elif info.status == "completed":
+            self._processing_requests.remove(info.request_id)
+            self._state.processing_requests = len(self._processing_requests)
+            self._state.processed_requests += 1
+            self._state.successful_requests += 1
+        elif info.status in ("errored", "cancelled"):
+            if info.request_id in self._queued_requests:
+                self._queued_requests.remove(info.request_id)
+                self._state.queued_requests = len(self._queued_requests)
+            elif info.request_id in self._pending_requests:
+                self._pending_requests.remove(info.request_id)
+                self._state.pending_requests = len(self._pending_requests)
+            elif info.request_id in self._processing_requests:
+                self._processing_requests.remove(info.request_id)
+                self._state.processing_requests = len(self._processing_requests)
+
+            self._state.processed_requests += 1
+            self._state.errored_requests += 1 if info.status == "errored" else 0
+            self._state.cancelled_requests += 1 if info.status == "cancelled" else 0
+        else:
+            raise ValueError(f"Unknown request_info status {info.status} for {info}")
+
+    def _update_with_constraints(self, info: RequestInfo):
+        actions: dict[str, SchedulerUpdateAction] = {
+            name: const(self._state, info) for name, const in self.constraints.items()
+        }
+        self._state.scheduler_constraints = actions
+        stop_queuing_actions = {}
+        stop_processing_actions = {}
+
+        for key, action in actions.items():
+            # Action updates
+            if (
+                self._state.end_queuing_time is None
+                and action.request_queuing == "stop"
+            ):
+                stop_queuing_actions[key] = action
+            if (
+                self._state.end_processing_time is None
+                and action.request_processing in ("stop_local", "stop_all")
+            ):
+                stop_processing_actions[key] = action
+
+            for progress_key in (
+                "remaining_fraction",
+                "remaining_requests",
+                "remaining_duration",
+            ):
+                if (new_val := action.progress.get(progress_key)) is not None and (
+                    getattr(self._state, progress_key) is None
+                    or new_val < getattr(self._state, progress_key)
+                ):
+                    setattr(self._state, progress_key, new_val)
+
+        if stop_queuing_actions:
+            self._state.end_queuing_constraints = stop_queuing_actions
+            self._state.end_queuing_time = time.time()
+
+        if stop_processing_actions:
+            self._state.end_processing_constraints = stop_processing_actions
+            self._state.end_processing_time = time.time()