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

experimaestro/launcherfinder/registry.py

@@ -1,5 +1,6 @@
 # Configuration registers
 
+from contextlib import contextmanager
 from typing import ClassVar, Dict, Optional, Set, Type, Union
 
 from pathlib import Path
@@ -7,7 +8,6 @@ import typing
 from omegaconf import DictConfig, OmegaConf, SCMode
 import pkg_resources
 from experimaestro.utils import logger
-
 from .base import ConnectorConfiguration, TokenConfiguration
 from .specs import HostRequirement
 
@@ -36,6 +36,16 @@ def load_yaml(schema, path: Path):
         )
 
 
+@contextmanager
+def ensure_enter(fp):
+    """Behaves as a resource, whether it is one or not"""
+    if hasattr(fp, "__enter__"):
+        with fp as _fp:
+            yield _fp
+    else:
+        yield fp
+
+
 class LauncherRegistry:
     INSTANCES: ClassVar[Dict[Path, "LauncherRegistry"]] = {}
     CURRENT_CONFIG_DIR: ClassVar[Optional[Path]] = None
@@ -78,13 +88,16 @@ class LauncherRegistry:
 
             from importlib import util
 
-            spec = util.spec_from_file_location("xpm_launchers_conf", launchers_py)
-            module = util.module_from_spec(spec)
-            spec.loader.exec_module(module)
+            with ensure_enter(launchers_py.__fspath__()) as fp:
+                spec = util.spec_from_file_location("xpm_launchers_conf", fp)
+                module = util.module_from_spec(spec)
+                spec.loader.exec_module(module)
 
             self.find_launcher_fn = getattr(module, "find_launcher", None)
             if self.find_launcher_fn is None:
-                logger.warn("No find_launcher() function was found in %s", launchers_py)
+                logger.warning(
+                    "No find_launcher() function was found in %s", launchers_py
+                )
 
         # Read the configuration file
         self.connectors = load_yaml(
@@ -144,9 +157,14 @@ class LauncherRegistry:
                 specs.append(spec)
 
         # Use launcher function
+        from experimaestro.launchers import Launcher
+
         if self.find_launcher_fn is not None:
             for spec in specs:
                 if launcher := self.find_launcher_fn(spec, tags):
+                    assert isinstance(
+                        launcher, Launcher
+                    ), "f{self.find_launcher_fn} did not return a Launcher but {type(launcher)}"
                     return launcher
 
         return None

experimaestro/launchers/slurm/base.py

@@ -11,6 +11,7 @@ from typing import (
 )
 from experimaestro.connectors.local import LocalConnector
 import re
+from shlex import quote as shquote
 from contextlib import contextmanager
 from dataclasses import dataclass
 from experimaestro.launcherfinder.registry import (
@@ -235,15 +236,15 @@ class SlurmProcessBuilder(ProcessBuilder):
         super().__init__()
         self.launcher = launcher
 
-    def start(self) -> BatchSlurmProcess:
+    def start(self, task_mode: bool = False) -> BatchSlurmProcess:
         """Start the process"""
         builder = self.launcher.connector.processbuilder()
-        builder.workingDirectory = self.workingDirectory
         builder.environ = self.launcher.launcherenv
         builder.detach = False
 
         if not self.detach:
             # Simplest case: we wait for the output
+            builder.workingDirectory = self.workingDirectory
             builder.command = [f"{self.launcher.binpath}/srun"]
             builder.command.extend(self.launcher.options.args())
             builder.command.extend(self.command)
@@ -255,11 +256,17 @@ class SlurmProcessBuilder(ProcessBuilder):
             return builder.start()
 
         builder.command = [f"{self.launcher.binpath}/sbatch", "--parsable"]
-        builder.command.extend(self.launcher.options.args())
 
-        addstream(builder.command, "-e", self.stderr)
-        addstream(builder.command, "-o", self.stdout)
-        addstream(builder.command, "-i", self.stdin)
+        if not task_mode:
+            # Use command line parameters when not running a task
+            builder.command.extend(self.launcher.options.args())
+
+            if self.workingDirectory:
+                workdir = self.launcher.connector.resolve(self.workingDirectory)
+                builder.command.append(f"--chdir={workdir}")
+            addstream(builder.command, "-e", self.stderr)
+            addstream(builder.command, "-o", self.stdout)
+            addstream(builder.command, "-i", self.stdin)
 
         builder.command.extend(self.command)
         logger.info(
@@ -427,12 +434,43 @@ class SlurmLauncher(Launcher):
 
         We assume *nix, but should be changed to PythonScriptBuilder when working
         """
-        builder = PythonScriptBuilder()
-        builder.processtype = "slurm"
-        return builder
+        return SlurmScriptBuilder(self)
 
     def processbuilder(self) -> SlurmProcessBuilder:
         """Returns the process builder for this launcher
 
         By default, returns the associated connector builder"""
         return SlurmProcessBuilder(self)
+
+
+class SlurmScriptBuilder(PythonScriptBuilder):
+    def __init__(self, launcher: SlurmLauncher, pythonpath=None):
+        super().__init__(pythonpath)
+        self.launcher = launcher
+        self.processtype = "slurm"
+
+    def write(self, job):
+        py_path = super().write(job)
+        main_path = py_path.parent
+
+        def relpath(path: Path):
+            return shquote(self.launcher.connector.resolve(path, main_path))
+
+        # Writes the sbatch shell script containing all the options
+        sh_path = job.jobpath / ("%s.sh" % job.name)
+        with sh_path.open("wt") as out:
+            out.write("""#!/bin/sh\n\n""")
+
+            workdir = self.launcher.connector.resolve(main_path)
+            out.write(f"#SBATCH --chdir={shquote(workdir)}\n")
+            out.write(f"""#SBATCH --error={relpath(job.stderr)}\n""")
+            out.write(f"""#SBATCH --output={relpath(job.stdout)}\n""")
+
+            for arg in self.launcher.options.args():
+                out.write(f"""#SBATCH {arg}\n""")
+
+            # We finish by the call to srun
+            out.write(f"""\nsrun ./{relpath(py_path)}\n\n""")
+
+        self.launcher.connector.setExecutable(sh_path, True)
+        return sh_path

experimaestro/notifications.py

@@ -78,7 +78,6 @@ class Reporter(threading.Thread):
 
         self.progress_threshold = 0.01
         self.cv = threading.Condition()
-        self.start()
 
     def stop(self):
         self.stopping = True
@@ -222,6 +221,7 @@ class Reporter(threading.Thread):
             taskpath = TaskEnv.instance().taskpath
             assert taskpath is not None, "Task path is not defined"
             Reporter.INSTANCE = Reporter(taskpath)
+            Reporter.INSTANCE.start()
         return Reporter.INSTANCE
 
 

experimaestro/run.py

@@ -140,10 +140,10 @@ class TaskRunner:
                 run(workdir / "params.json")
 
                 # ... remove the handlers
-                logger.info("Task ended successfully")
                 remove_signal_handlers(remove_cleanup=False)
 
                 # Everything went OK
+                logger.info("Task ended successfully")
                 sys.exit(0)
         except Exception:
             logger.exception("Got exception while running")

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()
@@ -696,6 +740,7 @@ class Scheduler:
                     code = await process.aio_code()
                     logger.debug("Got return code %s for %s", code, job)
 
+                    # Check the file if there is no return code
                     if code is None:
                         # Case where we cannot retrieve the code right away
                         if job.donepath.is_file():
@@ -861,6 +906,7 @@ 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
@@ -871,10 +917,22 @@ class experiment:
 
         async def awaitcompletion():
             assert self.central is not None
+            logger.debug("Waiting to exit scheduler...")
             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 +964,8 @@ 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 +994,7 @@ 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 +1003,8 @@ class experiment:
         self.exitMode = False
 
         self.central = SchedulerCentral.create(self.scheduler.name)
+        self.taskOutputsWorker = TaskOutputsWorker(self)
+        self.taskOutputsWorker.start()
 
         SIGNAL_HANDLER.add(self)
 
@@ -950,6 +1013,7 @@ class experiment:
         return self
 
     def __exit__(self, exc_type, exc_value, traceback):
+        logger.debug("Exiting scheduler context")
         # If no exception and normal run mode, remove old "jobs"
         if self.workspace.run_mode == RunMode.NORMAL:
             if exc_type is None and self.jobsbakpath.is_dir():
@@ -975,8 +1039,13 @@ 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:
@@ -985,8 +1054,27 @@ class experiment:
             # Put back old experiment as current one
             experiment.CURRENT = self.old_experiment
             if self.server:
+                logger.info("Stopping web 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", daemon=True)
+        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()

experimaestro/scriptbuilder.py

@@ -51,6 +51,8 @@ class PythonScriptBuilder:
         self.lockfiles: List[Path] = []
         self.notificationURL: Optional[str] = None
         self.command: Optional[AbstractCommand] = None
+
+        # This is used to serialize the full process identifier on disk
         self.processtype = "local"
 
     def write(self, job: CommandLineJob):
@@ -63,7 +65,7 @@ class PythonScriptBuilder:
             job {CommandLineJob} -- [description]
 
         Returns:
-            [type] -- [description]
+            str -- The script path on disk
         """
         assert isinstance(
             job, CommandLineJob