parsl 2024.3.18__py3-none-any.whl → 2025.1.13__py3-none-any.whl

parsl/monitoring/monitoring.py

@@ -1,30 +1,25 @@
 from __future__ import annotations
 
+import logging
+import multiprocessing.synchronize as ms
 import os
-import socket
-import time
 import pickle
-import logging
-import typeguard
-import zmq
-
 import queue
+import time
+from multiprocessing import Event
+from multiprocessing.queues import Queue
+from typing import TYPE_CHECKING, Literal, Optional, Tuple, Union, cast
 
-import parsl.monitoring.remote
+import typeguard
 
-from parsl.multiprocessing import ForkProcess, SizedQueue
-from multiprocessing import Process
-from multiprocessing.queues import Queue
 from parsl.log_utils import set_file_logger
-from parsl.utils import RepresentationMixin
+from parsl.monitoring.errors import MonitoringHubStartError
+from parsl.monitoring.radios.multiprocessing import MultiprocessingQueueRadioSender
+from parsl.monitoring.router import router_starter
+from parsl.monitoring.types import TaggedMonitoringMessage
+from parsl.multiprocessing import ForkProcess, SizedQueue
 from parsl.process_loggers import wrap_with_logs
-from parsl.utils import setproctitle
-
-from parsl.serialize import deserialize
-
-from parsl.monitoring.message_type import MessageType
-from parsl.monitoring.types import AddressedMonitoringMessage, TaggedMonitoringMessage
-from typing import cast, Any, Callable, Dict, Optional, Sequence, Tuple, Union, TYPE_CHECKING
+from parsl.utils import RepresentationMixin, setproctitle
 
 _db_manager_excepts: Optional[Exception]
 
@@ -49,7 +44,6 @@ class MonitoringHub(RepresentationMixin):
                  workflow_name: Optional[str] = None,
                  workflow_version: Optional[str] = None,
                  logging_endpoint: Optional[str] = None,
-                 logdir: Optional[str] = None,
                  monitoring_debug: bool = False,
                  resource_monitoring_enabled: bool = True,
                  resource_monitoring_interval: float = 30):  # in seconds
@@ -78,8 +72,6 @@ class MonitoringHub(RepresentationMixin):
              The database connection url for monitoring to log the information.
              These URLs follow RFC-1738, and can include username, password, hostname, database name.
              Default: sqlite, in the configured run_dir.
-        logdir : str
-             Parsl log directory paths. Logs and temp files go here. Default: '.'
         monitoring_debug : Bool
              Enable monitoring debug logging. Default: False
         resource_monitoring_enabled : boolean
@@ -93,14 +85,6 @@ class MonitoringHub(RepresentationMixin):
              Default: 30 seconds
         """
 
-        self.logger = logger
-
-        # Any is used to disable typechecking on uses of _dfk_channel,
-        # because it is used in the code as if it points to a channel, but
-        # the static type is that it can also be None. The code relies on
-        # .start() being called and initialising this to a real channel.
-        self._dfk_channel = None  # type: Any
-
         if _db_manager_excepts:
             raise _db_manager_excepts
 
@@ -109,7 +93,6 @@ class MonitoringHub(RepresentationMixin):
         self.hub_port_range = hub_port_range
 
         self.logging_endpoint = logging_endpoint
-        self.logdir = logdir
         self.monitoring_debug = monitoring_debug
 
         self.workflow_name = workflow_name
@@ -118,19 +101,15 @@ class MonitoringHub(RepresentationMixin):
         self.resource_monitoring_enabled = resource_monitoring_enabled
         self.resource_monitoring_interval = resource_monitoring_interval
 
-    def start(self, run_id: str, dfk_run_dir: str, config_run_dir: Union[str, os.PathLike]) -> int:
+    def start(self, dfk_run_dir: str, config_run_dir: Union[str, os.PathLike]) -> None:
 
-        if self.logdir is None:
-            self.logdir = "."
+        logger.debug("Starting MonitoringHub")
 
         if self.logging_endpoint is None:
             self.logging_endpoint = f"sqlite:///{os.fspath(config_run_dir)}/monitoring.db"
 
-        os.makedirs(self.logdir, exist_ok=True)
+        os.makedirs(dfk_run_dir, exist_ok=True)
 
-        # Initialize the ZMQ pipe to the Parsl Client
-
-        self.logger.debug("Initializing ZMQ Pipes to client")
         self.monitoring_hub_active = True
 
         # This annotation is incompatible with typeguard 4.x instrumentation
@@ -151,26 +130,22 @@ class MonitoringHub(RepresentationMixin):
         self.exception_q: Queue[Tuple[str, str]]
         self.exception_q = SizedQueue(maxsize=10)
 
-        self.priority_msgs: Queue[Tuple[Any, int]]
-        self.priority_msgs = SizedQueue()
-
-        self.resource_msgs: Queue[AddressedMonitoringMessage]
+        self.resource_msgs: Queue[Union[TaggedMonitoringMessage, Literal["STOP"]]]
         self.resource_msgs = SizedQueue()
 
-        self.node_msgs: Queue[AddressedMonitoringMessage]
-        self.node_msgs = SizedQueue()
-
-        self.block_msgs: Queue[AddressedMonitoringMessage]
-        self.block_msgs = SizedQueue()
+        self.router_exit_event: ms.Event
+        self.router_exit_event = Event()
 
         self.router_proc = ForkProcess(target=router_starter,
-                                       args=(comm_q, self.exception_q, self.priority_msgs, self.node_msgs, self.block_msgs, self.resource_msgs),
-                                       kwargs={"hub_address": self.hub_address,
-                                               "hub_port": self.hub_port,
-                                               "hub_port_range": self.hub_port_range,
-                                               "logdir": self.logdir,
+                                       kwargs={"comm_q": comm_q,
+                                               "exception_q": self.exception_q,
+                                               "resource_msgs": self.resource_msgs,
+                                               "exit_event": self.router_exit_event,
+                                               "hub_address": self.hub_address,
+                                               "udp_port": self.hub_port,
+                                               "zmq_port_range": self.hub_port_range,
+                                               "run_dir": dfk_run_dir,
                                                "logging_level": logging.DEBUG if self.monitoring_debug else logging.INFO,
-                                               "run_id": run_id
                                                },
                                        name="Monitoring-Router-Process",
                                        daemon=True,
@@ -178,8 +153,8 @@ class MonitoringHub(RepresentationMixin):
         self.router_proc.start()
 
         self.dbm_proc = ForkProcess(target=dbm_starter,
-                                    args=(self.exception_q, self.priority_msgs, self.node_msgs, self.block_msgs, self.resource_msgs,),
-                                    kwargs={"logdir": self.logdir,
+                                    args=(self.exception_q, self.resource_msgs,),
+                                    kwargs={"run_dir": dfk_run_dir,
                                             "logging_level": logging.DEBUG if self.monitoring_debug else logging.INFO,
                                             "db_url": self.logging_endpoint,
                                             },
@@ -187,113 +162,97 @@ class MonitoringHub(RepresentationMixin):
                                     daemon=True,
                                     )
         self.dbm_proc.start()
-        self.logger.info("Started the router process {} and DBM process {}".format(self.router_proc.pid, self.dbm_proc.pid))
+        logger.info("Started the router process %s and DBM process %s", self.router_proc.pid, self.dbm_proc.pid)
 
-        self.filesystem_proc = Process(target=filesystem_receiver,
-                                       args=(self.logdir, self.resource_msgs, dfk_run_dir),
-                                       name="Monitoring-Filesystem-Process",
-                                       daemon=True
-                                       )
+        self.filesystem_proc = ForkProcess(target=filesystem_receiver,
+                                           args=(self.resource_msgs, dfk_run_dir),
+                                           name="Monitoring-Filesystem-Process",
+                                           daemon=True
+                                           )
         self.filesystem_proc.start()
-        self.logger.info(f"Started filesystem radio receiver process {self.filesystem_proc.pid}")
+        logger.info("Started filesystem radio receiver process %s", self.filesystem_proc.pid)
+
+        self.radio = MultiprocessingQueueRadioSender(self.resource_msgs)
 
         try:
             comm_q_result = comm_q.get(block=True, timeout=120)
+            comm_q.close()
+            comm_q.join_thread()
         except queue.Empty:
-            self.logger.error("Hub has not completed initialization in 120s. Aborting")
-            raise Exception("Hub failed to start")
+            logger.error("Hub has not completed initialization in 120s. Aborting")
+            raise MonitoringHubStartError()
 
         if isinstance(comm_q_result, str):
-            self.logger.error(f"MonitoringRouter sent an error message: {comm_q_result}")
+            logger.error("MonitoringRouter sent an error message: %s", comm_q_result)
             raise RuntimeError(f"MonitoringRouter failed to start: {comm_q_result}")
 
-        udp_port, ic_port = comm_q_result
+        udp_port, zmq_port = comm_q_result
 
         self.monitoring_hub_url = "udp://{}:{}".format(self.hub_address, udp_port)
 
-        context = zmq.Context()
-        self.dfk_channel_timeout = 10000  # in milliseconds
-        self._dfk_channel = context.socket(zmq.DEALER)
-        self._dfk_channel.setsockopt(zmq.LINGER, 0)
-        self._dfk_channel.set_hwm(0)
-        self._dfk_channel.setsockopt(zmq.SNDTIMEO, self.dfk_channel_timeout)
-        self._dfk_channel.connect("tcp://{}:{}".format(self.hub_address, ic_port))
+        logger.info("Monitoring Hub initialized")
 
-        self.logger.info("Monitoring Hub initialized")
+        self.hub_zmq_port = zmq_port
 
-        return ic_port
-
-    # TODO: tighten the Any message format
-    def send(self, mtype: MessageType, message: Any) -> None:
-        self.logger.debug("Sending message type {}".format(mtype))
-        try:
-            self._dfk_channel.send_pyobj((mtype, message))
-        except zmq.Again:
-            self.logger.exception(
-                "The monitoring message sent from DFK to router timed-out after {}ms".format(self.dfk_channel_timeout))
+    def send(self, message: TaggedMonitoringMessage) -> None:
+        logger.debug("Sending message type %s", message[0])
+        self.radio.send(message)
 
     def close(self) -> None:
-        self.logger.info("Terminating Monitoring Hub")
+        logger.info("Terminating Monitoring Hub")
         exception_msgs = []
         while True:
             try:
                 exception_msgs.append(self.exception_q.get(block=False))
-                self.logger.error("There was a queued exception (Either router or DBM process got exception much earlier?)")
+                logger.error("There was a queued exception (Either router or DBM process got exception much earlier?)")
             except queue.Empty:
                 break
-        if self._dfk_channel and self.monitoring_hub_active:
+        if self.monitoring_hub_active:
             self.monitoring_hub_active = False
-            self._dfk_channel.close()
             if exception_msgs:
                 for exception_msg in exception_msgs:
-                    self.logger.error(
-                        "{} process delivered an exception: {}. Terminating all monitoring processes immediately.".format(
-                            exception_msg[0],
-                            exception_msg[1]
-                        )
+                    logger.error(
+                        "%s process delivered an exception: %s. Terminating all monitoring processes immediately.",
+                        exception_msg[0],
+                        exception_msg[1]
                     )
                 self.router_proc.terminate()
                 self.dbm_proc.terminate()
                 self.filesystem_proc.terminate()
-            self.logger.info("Waiting for router to terminate")
+            logger.info("Setting router termination event")
+            self.router_exit_event.set()
+            logger.info("Waiting for router to terminate")
             self.router_proc.join()
-            self.logger.debug("Finished waiting for router termination")
+            self.router_proc.close()
+            logger.debug("Finished waiting for router termination")
             if len(exception_msgs) == 0:
-                self.logger.debug("Sending STOP to DBM")
-                self.priority_msgs.put(("STOP", 0))
+                logger.debug("Sending STOP to DBM")
+                self.resource_msgs.put("STOP")
             else:
-                self.logger.debug("Not sending STOP to DBM, because there were DBM exceptions")
-            self.logger.debug("Waiting for DB termination")
+                logger.debug("Not sending STOP to DBM, because there were DBM exceptions")
+            logger.debug("Waiting for DB termination")
             self.dbm_proc.join()
-            self.logger.debug("Finished waiting for DBM termination")
+            self.dbm_proc.close()
+            logger.debug("Finished waiting for DBM termination")
 
             # should this be message based? it probably doesn't need to be if
             # we believe we've received all messages
-            self.logger.info("Terminating filesystem radio receiver process")
+            logger.info("Terminating filesystem radio receiver process")
             self.filesystem_proc.terminate()
             self.filesystem_proc.join()
+            self.filesystem_proc.close()
 
-    @staticmethod
-    def monitor_wrapper(f: Any,
-                        args: Sequence,
-                        kwargs: Dict,
-                        try_id: int,
-                        task_id: int,
-                        monitoring_hub_url: str,
-                        run_id: str,
-                        logging_level: int,
-                        sleep_dur: float,
-                        radio_mode: str,
-                        monitor_resources: bool,
-                        run_dir: str) -> Tuple[Callable, Sequence, Dict]:
-        return parsl.monitoring.remote.monitor_wrapper(f, args, kwargs, try_id, task_id, monitoring_hub_url,
-                                                       run_id, logging_level, sleep_dur, radio_mode,
-                                                       monitor_resources, run_dir)
+            logger.info("Closing monitoring multiprocessing queues")
+            self.exception_q.close()
+            self.exception_q.join_thread()
+            self.resource_msgs.close()
+            self.resource_msgs.join_thread()
+            logger.info("Closed monitoring multiprocessing queues")
 
 
 @wrap_with_logs
-def filesystem_receiver(logdir: str, q: "queue.Queue[AddressedMonitoringMessage]", run_dir: str) -> None:
-    logger = set_file_logger("{}/monitoring_filesystem_radio.log".format(logdir),
+def filesystem_receiver(q: Queue[TaggedMonitoringMessage], run_dir: str) -> None:
+    logger = set_file_logger(f"{run_dir}/monitoring_filesystem_radio.log",
                              name="monitoring_filesystem_radio",
                              level=logging.INFO)
 
@@ -302,7 +261,9 @@ def filesystem_receiver(logdir: str, q: "queue.Queue[AddressedMonitoringMessage]
     base_path = f"{run_dir}/monitor-fs-radio/"
     tmp_dir = f"{base_path}/tmp/"
     new_dir = f"{base_path}/new/"
-    logger.debug(f"Creating new and tmp paths under {base_path}")
+    logger.debug("Creating new and tmp paths under %s", base_path)
+
+    target_radio = MultiprocessingQueueRadioSender(q)
 
     os.makedirs(tmp_dir, exist_ok=True)
     os.makedirs(new_dir, exist_ok=True)
@@ -313,202 +274,15 @@ def filesystem_receiver(logdir: str, q: "queue.Queue[AddressedMonitoringMessage]
         # iterate over files in new_dir
         for filename in os.listdir(new_dir):
             try:
-                logger.info(f"Processing filesystem radio file {filename}")
+                logger.info("Processing filesystem radio file %s", filename)
                 full_path_filename = f"{new_dir}/{filename}"
                 with open(full_path_filename, "rb") as f:
-                    message = deserialize(f.read())
-                logger.debug(f"Message received is: {message}")
+                    message = pickle.load(f)
+                logger.debug("Message received is: %s", message)
                 assert isinstance(message, tuple)
-                q.put(cast(AddressedMonitoringMessage, message))
+                target_radio.send(cast(TaggedMonitoringMessage, message))
                 os.remove(full_path_filename)
             except Exception:
-                logger.exception(f"Exception processing {filename} - probably will be retried next iteration")
+                logger.exception("Exception processing %s - probably will be retried next iteration", filename)
 
         time.sleep(1)  # whats a good time for this poll?
-
-
-class MonitoringRouter:
-
-    def __init__(self,
-                 *,
-                 hub_address: str,
-                 hub_port: Optional[int] = None,
-                 hub_port_range: Tuple[int, int] = (55050, 56000),
-
-                 monitoring_hub_address: str = "127.0.0.1",
-                 logdir: str = ".",
-                 run_id: str,
-                 logging_level: int = logging.INFO,
-                 atexit_timeout: int = 3    # in seconds
-                 ):
-        """ Initializes a monitoring configuration class.
-
-        Parameters
-        ----------
-        hub_address : str
-             The ip address at which the workers will be able to reach the Hub.
-        hub_port : int
-             The specific port at which workers will be able to reach the Hub via UDP. Default: None
-        hub_port_range : tuple(int, int)
-             The MonitoringHub picks ports at random from the range which will be used by Hub.
-             This is overridden when the hub_port option is set. Default: (55050, 56000)
-        logdir : str
-             Parsl log directory paths. Logs and temp files go here. Default: '.'
-        logging_level : int
-             Logging level as defined in the logging module. Default: logging.INFO
-        atexit_timeout : float, optional
-            The amount of time in seconds to terminate the hub without receiving any messages, after the last dfk workflow message is received.
-
-        """
-        os.makedirs(logdir, exist_ok=True)
-        self.logger = set_file_logger("{}/monitoring_router.log".format(logdir),
-                                      name="monitoring_router",
-                                      level=logging_level)
-        self.logger.debug("Monitoring router starting")
-
-        self.hub_address = hub_address
-        self.atexit_timeout = atexit_timeout
-        self.run_id = run_id
-
-        self.loop_freq = 10.0  # milliseconds
-
-        # Initialize the UDP socket
-        self.sock = socket.socket(socket.AF_INET,
-                                  socket.SOCK_DGRAM,
-                                  socket.IPPROTO_UDP)
-
-        # We are trying to bind to all interfaces with 0.0.0.0
-        if not hub_port:
-            self.sock.bind(('0.0.0.0', 0))
-            self.hub_port = self.sock.getsockname()[1]
-        else:
-            self.hub_port = hub_port
-            try:
-                self.sock.bind(('0.0.0.0', self.hub_port))
-            except Exception as e:
-                raise RuntimeError(f"Could not bind to hub_port {hub_port} because: {e}")
-        self.sock.settimeout(self.loop_freq / 1000)
-        self.logger.info("Initialized the UDP socket on 0.0.0.0:{}".format(self.hub_port))
-
-        self._context = zmq.Context()
-        self.ic_channel = self._context.socket(zmq.DEALER)
-        self.ic_channel.setsockopt(zmq.LINGER, 0)
-        self.ic_channel.set_hwm(0)
-        self.ic_channel.RCVTIMEO = int(self.loop_freq)  # in milliseconds
-        self.logger.debug("hub_address: {}. hub_port_range {}".format(hub_address, hub_port_range))
-        self.ic_port = self.ic_channel.bind_to_random_port("tcp://*",
-                                                           min_port=hub_port_range[0],
-                                                           max_port=hub_port_range[1])
-
-    def start(self,
-              priority_msgs: "queue.Queue[AddressedMonitoringMessage]",
-              node_msgs: "queue.Queue[AddressedMonitoringMessage]",
-              block_msgs: "queue.Queue[AddressedMonitoringMessage]",
-              resource_msgs: "queue.Queue[AddressedMonitoringMessage]") -> None:
-        try:
-            router_keep_going = True
-            while router_keep_going:
-                try:
-                    data, addr = self.sock.recvfrom(2048)
-                    resource_msg = pickle.loads(data)
-                    self.logger.debug("Got UDP Message from {}: {}".format(addr, resource_msg))
-                    resource_msgs.put((resource_msg, addr))
-                except socket.timeout:
-                    pass
-
-                try:
-                    dfk_loop_start = time.time()
-                    while time.time() - dfk_loop_start < 1.0:  # TODO make configurable
-                        # note that nothing checks that msg really is of the annotated type
-                        msg: TaggedMonitoringMessage
-                        msg = self.ic_channel.recv_pyobj()
-
-                        assert isinstance(msg, tuple), "IC Channel expects only tuples, got {}".format(msg)
-                        assert len(msg) >= 1, "IC Channel expects tuples of length at least 1, got {}".format(msg)
-                        assert len(msg) == 2, "IC Channel expects message tuples of exactly length 2, got {}".format(msg)
-
-                        msg_0: AddressedMonitoringMessage
-                        msg_0 = (msg, 0)
-
-                        if msg[0] == MessageType.NODE_INFO:
-                            msg[1]['run_id'] = self.run_id
-                            node_msgs.put(msg_0)
-                        elif msg[0] == MessageType.RESOURCE_INFO:
-                            resource_msgs.put(msg_0)
-                        elif msg[0] == MessageType.BLOCK_INFO:
-                            block_msgs.put(msg_0)
-                        elif msg[0] == MessageType.TASK_INFO:
-                            priority_msgs.put(msg_0)
-                        elif msg[0] == MessageType.WORKFLOW_INFO:
-                            priority_msgs.put(msg_0)
-                            if 'exit_now' in msg[1] and msg[1]['exit_now']:
-                                router_keep_going = False
-                        else:
-                            # There is a type: ignore here because if msg[0]
-                            # is of the correct type, this code is unreachable,
-                            # but there is no verification that the message
-                            # received from ic_channel.recv_pyobj() is actually
-                            # of that type.
-                            self.logger.error("Discarding message "  # type: ignore[unreachable]
-                                              f"from interchange with unknown type {msg[0].value}")
-                except zmq.Again:
-                    pass
-                except Exception:
-                    # This will catch malformed messages. What happens if the
-                    # channel is broken in such a way that it always raises
-                    # an exception? Looping on this would maybe be the wrong
-                    # thing to do.
-                    self.logger.warning("Failure processing a ZMQ message", exc_info=True)
-
-            self.logger.info("Monitoring router draining")
-            last_msg_received_time = time.time()
-            while time.time() - last_msg_received_time < self.atexit_timeout:
-                try:
-                    data, addr = self.sock.recvfrom(2048)
-                    msg = pickle.loads(data)
-                    self.logger.debug("Got UDP Message from {}: {}".format(addr, msg))
-                    resource_msgs.put((msg, addr))
-                    last_msg_received_time = time.time()
-                except socket.timeout:
-                    pass
-
-            self.logger.info("Monitoring router finishing normally")
-        finally:
-            self.logger.info("Monitoring router finished")
-
-
-@wrap_with_logs
-def router_starter(comm_q: "queue.Queue[Union[Tuple[int, int], str]]",
-                   exception_q: "queue.Queue[Tuple[str, str]]",
-                   priority_msgs: "queue.Queue[AddressedMonitoringMessage]",
-                   node_msgs: "queue.Queue[AddressedMonitoringMessage]",
-                   block_msgs: "queue.Queue[AddressedMonitoringMessage]",
-                   resource_msgs: "queue.Queue[AddressedMonitoringMessage]",
-
-                   hub_address: str,
-                   hub_port: Optional[int],
-                   hub_port_range: Tuple[int, int],
-
-                   logdir: str,
-                   logging_level: int,
-                   run_id: str) -> None:
-    setproctitle("parsl: monitoring router")
-    try:
-        router = MonitoringRouter(hub_address=hub_address,
-                                  hub_port=hub_port,
-                                  hub_port_range=hub_port_range,
-                                  logdir=logdir,
-                                  logging_level=logging_level,
-                                  run_id=run_id)
-    except Exception as e:
-        logger.error("MonitoringRouter construction failed.", exc_info=True)
-        comm_q.put(f"Monitoring router construction failed: {e}")
-    else:
-        comm_q.put((router.hub_port, router.ic_port))
-
-        router.logger.info("Starting MonitoringRouter in router_starter")
-        try:
-            router.start(priority_msgs, node_msgs, block_msgs, resource_msgs)
-        except Exception as e:
-            router.logger.exception("router.start exception")
-            exception_q.put(('Hub', str(e)))

parsl/monitoring/queries/pandas.py

@@ -1,7 +1,6 @@
-import pandas as pd
-
 from typing import Any
 
+import pandas as pd
 
 # pandas can take several different types of database connection,
 # and itself exposes its connection parameters as "Any".

parsl/monitoring/radios/base.py

@@ -0,0 +1,13 @@
+import logging
+from abc import ABCMeta, abstractmethod
+from typing import Optional
+
+_db_manager_excepts: Optional[Exception]
+
+logger = logging.getLogger(__name__)
+
+
+class MonitoringRadioSender(metaclass=ABCMeta):
+    @abstractmethod
+    def send(self, message: object) -> None:
+        pass

parsl/monitoring/radios/filesystem.py

@@ -0,0 +1,52 @@
+import logging
+import os
+import pickle
+import uuid
+
+from parsl.monitoring.radios.base import MonitoringRadioSender
+
+logger = logging.getLogger(__name__)
+
+
+class FilesystemRadioSender(MonitoringRadioSender):
+    """A MonitoringRadioSender that sends messages over a shared filesystem.
+
+    The messsage directory structure is based on maildir,
+    https://en.wikipedia.org/wiki/Maildir
+
+    The writer creates a message in tmp/ and then when it is fully
+    written, moves it atomically into new/
+
+    The reader ignores tmp/ and only reads and deletes messages from
+    new/
+
+    This avoids a race condition of reading partially written messages.
+
+    This radio is likely to give higher shared filesystem load compared to
+    the UDP radio, but should be much more reliable.
+    """
+
+    def __init__(self, *, monitoring_url: str, timeout: int = 10, run_dir: str):
+        logger.info("filesystem based monitoring channel initializing")
+        self.base_path = f"{run_dir}/monitor-fs-radio/"
+        self.tmp_path = f"{self.base_path}/tmp"
+        self.new_path = f"{self.base_path}/new"
+
+        os.makedirs(self.tmp_path, exist_ok=True)
+        os.makedirs(self.new_path, exist_ok=True)
+
+    def send(self, message: object) -> None:
+        logger.info("Sending a monitoring message via filesystem")
+
+        unique_id = str(uuid.uuid4())
+
+        tmp_filename = f"{self.tmp_path}/{unique_id}"
+        new_filename = f"{self.new_path}/{unique_id}"
+        buffer = message
+
+        # this will write the message out then atomically
+        # move it into new/, so that a partially written
+        # file will never be observed in new/
+        with open(tmp_filename, "wb") as f:
+            pickle.dump(buffer, f)
+        os.rename(tmp_filename, new_filename)

parsl/monitoring/radios/htex.py

@@ -0,0 +1,57 @@
+import logging
+import pickle
+
+from parsl.monitoring.radios.base import MonitoringRadioSender
+
+logger = logging.getLogger(__name__)
+
+
+class HTEXRadioSender(MonitoringRadioSender):
+
+    def __init__(self, monitoring_url: str, timeout: int = 10):
+        """
+        Parameters
+        ----------
+
+        monitoring_url : str
+            URL of the form <scheme>://<IP>:<PORT>
+        timeout : int
+            timeout, default=10s
+        """
+        logger.info("htex-based monitoring channel initialising")
+
+    def send(self, message: object) -> None:
+        """ Sends a message to the UDP receiver
+
+        Parameter
+        ---------
+
+        message: object
+            Arbitrary pickle-able object that is to be sent
+
+        Returns:
+            None
+        """
+
+        import parsl.executors.high_throughput.monitoring_info
+
+        result_queue = parsl.executors.high_throughput.monitoring_info.result_queue
+
+        # this message needs to go in the result queue tagged so that it is treated
+        # i) as a monitoring message by the interchange, and then further more treated
+        # as a RESOURCE_INFO message when received by monitoring (rather than a NODE_INFO
+        # which is the implicit default for messages from the interchange)
+
+        # for the interchange, the outer wrapper, this needs to be a dict:
+
+        interchange_msg = {
+            'type': 'monitoring',
+            'payload': message
+        }
+
+        if result_queue:
+            result_queue.put(pickle.dumps(interchange_msg))
+        else:
+            logger.error("result_queue is uninitialized - cannot put monitoring message")
+
+        return