experimaestro 1.6.2__py3-none-any.whl → 1.7.0rc1__py3-none-any.whl

experimaestro/scheduler/base.py

@@ -7,7 +7,16 @@ from pathlib import Path
 from shutil import rmtree
 import threading
 import time
-from typing import Any, Iterator, List, Optional, Set, TypeVar, Union, TYPE_CHECKING
+from typing import (
+    Any,
+    Iterator,
+    List,
+    Optional,
+    Set,
+    TypeVar,
+    Union,
+    TYPE_CHECKING,
+)
 import enum
 import signal
 import asyncio
@@ -18,9 +27,10 @@ from experimaestro.scheduler.services import Service
 from experimaestro.settings import WorkspaceSettings, get_settings
 
 
-from experimaestro.core.objects import Config, ConfigWalkContext
+from experimaestro.core.objects import Config, ConfigWalkContext, WatchedOutput
 from experimaestro.utils import logger
 from experimaestro.locking import Locks, LockError, Lock
+from experimaestro.utils.asyncio import asyncThreadcheck
 from .workspace import RunMode, Workspace
 from .dependencies import Dependency, DependencyStatus, Resource
 import concurrent.futures
@@ -111,7 +121,7 @@ class JobDependency(Dependency):
 
 
 class Job(Resource):
-    """A job is a resouce that is produced by the execution of some code"""
+    """A job is a resource that is produced by the execution of some code"""
 
     # Set by the scheduler
     _readyEvent: Optional[asyncio.Event]
@@ -149,6 +159,11 @@ class Job(Resource):
         # Dependencies
         self.dependencies: Set[Dependency] = set()  # as target
 
+        # Watched outputs
+        self.watched_outputs = {}
+        for watched in config.__xpm__.watched_outputs:
+            self.watch_output(watched)
+
         # Process
         self._process = None
         self.unsatisfied = 0
@@ -160,6 +175,23 @@ class Job(Resource):
         self._progress: List[LevelInformation] = []
         self.tags = config.tags()
 
+    def watch_output(self, watched: "WatchedOutput"):
+        """Monitor task outputs
+
+        :param watched: A description of the watched output
+        """
+        self.scheduler.xp.watch_output(watched)
+
+    def task_output_update(self, subpath: Path):
+        """Notification of an updated task output"""
+        if watcher := self.watched_outputs.get(subpath, None):
+            watcher.update()
+
+    def done_handler(self):
+        """The task has been completed"""
+        for watcher in self.watched_outputs.values():
+            watcher.update()
+
     def __str__(self):
         return "Job[{}]".format(self.identifier)
 
@@ -170,10 +202,8 @@ class Job(Resource):
     @cached_property
     def python_path(self) -> Iterator[str]:
         """Returns an iterator over python path"""
-        return itertools.chain(
-            self.workspace.python_path
-        )
-        
+        return itertools.chain(self.workspace.python_path)
+
     @cached_property
     def environ(self):
         """Returns the job environment
@@ -227,7 +257,7 @@ class Job(Resource):
         return self.state == JobState.READY
 
     @property
-    def jobpath(self):
+    def jobpath(self) -> Path:
         """Deprecated, use `path`"""
         return self.workspace.jobspath / self.relpath
 
@@ -235,6 +265,14 @@ class Job(Resource):
     def path(self) -> Path:
         return self.workspace.jobspath / self.relpath
 
+    @property
+    def experimaestro_path(self) -> Path:
+        return (self.path / ".experimaestro").resolve()
+
+    @cached_property
+    def task_outputs_path(self) -> Path:
+        return self.experimaestro_path / "task-outputs.jsonl"
+
     @property
     def relpath(self):
         identifier = self.config.__xpm__.identifier
@@ -444,7 +482,7 @@ class Scheduler:
         self.jobs: Dict[str, "Job"] = {}
 
         # List of jobs
-        self.waitingjobs = set()
+        self.waitingjobs: Set[Job] = set()
 
         # Listeners
         self.listeners: Set[Listener] = set()
@@ -467,10 +505,12 @@ class Scheduler:
 
     def submit(self, job: Job) -> Optional[Job]:
         # Wait for the future containing the submitted job
+        logger.debug("Registering the job %s within the scheduler", job)
         otherFuture = asyncio.run_coroutine_threadsafe(
             self.aio_registerJob(job), self.loop
         )
         other = otherFuture.result()
+        logger.debug("Job already submitted" if other else "First submission")
         if other:
             return other
 
@@ -606,9 +646,13 @@ class Scheduler:
         if job.state != JobState.DONE:
             self.xp.failedJobs[job.identifier] = job
 
+        # Process all remaining tasks outputs
+        await asyncThreadcheck("End of job processing", job.done_handler)
+
         # Decrement the number of unfinished jobs and notify
         self.xp.unfinishedJobs -= 1
         async with self.xp.central.exitCondition:
+            logging.debug("Updated number of unfinished jobs")
             self.xp.central.exitCondition.notify_all()
 
         job.endtime = time.time()
@@ -861,6 +905,7 @@ class experiment:
             assert self.central is not None
             async with self.central.exitCondition:
                 self.exitMode = True
+                logging.debug("Setting exit mode to true")
                 self.central.exitCondition.notify_all()
 
         assert self.central is not None and self.central.loop is not None
@@ -873,8 +918,19 @@ class experiment:
             assert self.central is not None
             async with self.central.exitCondition:
                 while True:
-                    if self.unfinishedJobs == 0 or self.exitMode:
+                    if self.exitMode:
                         break
+
+                    # If we have still unfinished jobs or possible new tasks, wait
+                    logger.debug(
+                        "Checking exit condition: unfinished jobs=%d, task output queue size=%d",
+                        self.unfinishedJobs,
+                        self.taskOutputQueueSize,
+                    )
+                    if self.unfinishedJobs == 0 and self.taskOutputQueueSize == 0:
+                        break
+
+                    # Wait for more news...
                     await self.central.exitCondition.wait()
 
                 if self.failedJobs:
@@ -906,6 +962,8 @@ class experiment:
         return self.workspace.connector.createtoken(name, count)
 
     def __enter__(self):
+        from .dynamic_outputs import TaskOutputsWorker
+
         if self.workspace.run_mode != RunMode.DRY_RUN:
             logger.info("Locking experiment %s", self.xplockpath)
             self.xplock = self.workspace.connector.lock(self.xplockpath, 0).__enter__()
@@ -934,6 +992,7 @@ class experiment:
         global SIGNAL_HANDLER
         # Number of unfinished jobs
         self.unfinishedJobs = 0
+        self.taskOutputQueueSize = 0
 
         # List of failed jobs
         self.failedJobs: Dict[str, Job] = {}
@@ -942,6 +1001,8 @@ class experiment:
         self.exitMode = False
 
         self.central = SchedulerCentral.create(self.scheduler.name)
+        self.taskOutputsWorker = TaskOutputsWorker(self)
+        self.taskOutputsWorker.start()
 
         SIGNAL_HANDLER.add(self)
 
@@ -975,8 +1036,13 @@ class experiment:
                 service.stop()
 
             if self.central is not None:
+                logger.info("Stopping scheduler event loop")
                 self.central.loop.stop()
 
+            if self.taskOutputsWorker is not None:
+                logger.info("Stopping tasks outputs worker")
+                self.taskOutputsWorker.queue.put(None)
+
             self.central = None
             self.workspace.__exit__(exc_type, exc_value, traceback)
             if self.xplock:
@@ -987,6 +1053,24 @@ class experiment:
             if self.server:
                 self.server.stop()
 
+    async def update_task_output_count(self, delta: int):
+        """Change in the number of task outputs to process"""
+        async with self.central.exitCondition:
+            self.taskOutputQueueSize += delta
+            logging.debug(
+                "Updating queue size with %d => %d", delta, self.taskOutputQueueSize
+            )
+            if self.taskOutputQueueSize == 0:
+                self.central.exitCondition.notify_all()
+
+    def watch_output(self, watched: "WatchedOutput"):
+        """Watch an output
+
+        :param watched: The watched output specification
+        """
+
+        self.taskOutputsWorker.watch_output(watched)
+
     def add_service(self, service: ServiceClass) -> ServiceClass:
         """Adds a service (e.g. tensorboard viewer) to the experiment
 

experimaestro/scheduler/dynamic_outputs.py

@@ -0,0 +1,184 @@
+"""Handles dynamic task outputs"""
+
+import asyncio
+import json
+import logging
+import queue
+import threading
+from collections import defaultdict
+from functools import cached_property
+from pathlib import Path
+from typing import Callable, TYPE_CHECKING
+
+from watchdog.events import FileSystemEventHandler
+
+from experimaestro.ipc import ipcom
+from experimaestro.utils import logger
+
+from .base import Job, experiment
+
+if TYPE_CHECKING:
+    from experimaestro.core.objects import WatchedOutput
+
+
+class TaskOutputCallbackHandler:
+    def __init__(self, converter: Callable):
+        pass
+
+
+class TaskOutputs(FileSystemEventHandler):
+    """Represent and monitors dynamic outputs generated by one task"""
+
+    #: Global dictionary for handles
+    HANDLERS: dict[Path, "TaskOutputs"] = {}
+
+    #: Global lock to access current HANDLERS
+    LOCK = threading.Lock()
+
+    def create(job: Job):
+        with TaskOutputs.LOCK:
+            if instance := TaskOutputs.get(job.task_outputs_path, None):
+                return instance
+
+            instance = TaskOutputs(job.task_outputs_path)
+            TaskOutputs[job.task_outputs_path] = instance
+            return instance
+
+    def __init__(self, path: Path):
+        """Monitors an event path"""
+        logger.debug("Watching dynamic task outputs in %s", path)
+        self.path = path
+        self.handle = None
+        self.count = 0
+        self.lock = threading.Lock()
+        self.listeners: dict[str, dict[Callable, set[Callable]]] = defaultdict(
+            lambda: defaultdict(set)
+        )
+
+        #: The events registered so far
+        self.events = []
+
+    def __enter__(self):
+        """Starts monitoring task outputs"""
+        self.job.task_outputs_path.parent.mkdir(parents=True, exist_ok=True)
+        with self.lock:
+            if self.handle is None:
+                assert self.count == 0
+                self.handle = ipcom().fswatch(self, self.path.parent, False)
+            self.count += 1
+        return self
+
+    def __exit__(self, *args):
+        """Stops monitoring task outputs"""
+        with self.lock:
+            self.count -= 1
+            if self.count == 0:
+                ipcom().fsunwatch(self.handle)
+                self.fh.close()
+
+                self.handle = None
+                self._fh = None
+
+    def watch_output(self, watched: "WatchedOutput"):
+        """Add a new listener"""
+        key = f"{watched.config.__identifier__}/{watched.method_name}"
+        with self.lock:
+            # Process events so far
+            listener = self.listeners[key].get(watched.method, None)
+            if listener is None:
+                listener = TaskOutputCallbackHandler(watched.method)
+
+            # Register
+            self.listeners[key][watched.method].add(watched.callback)
+
+    #
+    # --- Events
+    #
+
+    @cached_property
+    def fh(self):
+        if self._fh is None:
+            self._fh = self.path.open("rt")
+        return self._fh
+
+    def on_modified(self, event):
+        self.handle(Path(event.src_path))
+
+    def on_created(self, event):
+        self.handle(Path(event.src_path))
+
+    def handle(self, path: Path):
+        if path != self.path:
+            return
+
+        with self.lock:
+            logger.debug("[TASK OUTPUT] Handling task output for %s", self.path)
+
+            while json_line := self.fh.readline():
+                # Read the event
+                event = json.loads(json_line)
+                logger.debug("Event: %s", event)
+
+                # FIXME: move elsewhere
+                # # Process the event
+                # event = self.config_method(
+                #     self.job.config.__xpm__.mark_output,
+                #     *event["args"],
+                #     **event["kwargs"],
+                # )
+
+                self.events.append(event)
+                # self.job.scheduler.xp.taskOutputsWorker.add(self, event)
+
+
+class TaskOutputsWorker(threading.Thread):
+    """This worker process dynamic output queue for one experiment"""
+
+    def __init__(self, xp: experiment):
+        super().__init__(name="task outputs worker")
+        self.queue = queue.Queue()
+        self.xp = xp
+
+    def watch_output(self, watched: "WatchedOutput"):
+        """Watch an output
+
+        :param watched: The watched output specification
+        """
+        logger.debug("Registering task output listener %s", watched)
+
+        # path = watched.job.tasks_output_path
+        TaskOutputs.create(watched.job).watch_output(watched)
+
+    def add(self, watcher, event):
+        asyncio.run_coroutine_threadsafe(
+            self.xp.update_task_output_count(1),
+            self.xp.scheduler.loop,
+        ).result()
+        self.queue.put((watcher, event))
+
+    def run(self):
+        logging.debug("Starting output listener queue")
+        while True:
+            # Get the next element in the queue
+            element = self.queue.get()
+            if element is None:
+                # end of processing
+                break
+
+            # Call all the listeners
+            logging.debug("Got one event: %s", element)
+            watcher, event = element
+            for listener in watcher.listeners:
+                try:
+                    logger.debug("Calling listener [%s] with %s", listener, event)
+                    listener(event)
+                    logger.debug(
+                        "[done] Calling listener [%s] with %s", listener, event
+                    )
+                except Exception:
+                    logging.exception("Exception while calling the listener")
+                self.queue.task_done()
+
+                asyncio.run_coroutine_threadsafe(
+                    self.xp.update_task_output_count(-1), self.xp.scheduler.loop
+                ).result()