Fix race condition for published futures with annotations (dask#8577)

distributed/client.py

@@ -2650,18 +2650,23 @@ def retry(self, futures, asynchronous=None):
     @log_errors
     async def _publish_dataset(self, *args, name=None, override=False, **kwargs):
         coroutines = []
+        uid = uuid.uuid4().hex
+        self._send_to_scheduler({"op": "publish_flush_batched_send", "uid": uid})
 
         def add_coro(name, data):
             keys = [f.key for f in futures_of(data)]
-            coroutines.append(
-                self.scheduler.publish_put(
+
+            async def _():
+                await self.scheduler.publish_wait_flush(uid=uid)
+                await self.scheduler.publish_put(
                     keys=keys,
                     name=name,
                     data=to_serialize(data),
                     override=override,
                     client=self.id,
                 )
-            )
+
+            coroutines.append(_())
 
         if name:
             if len(args) == 0:

distributed/publish.py

@@ -1,5 +1,7 @@
 from __future__ import annotations
 
+import asyncio
+from collections import defaultdict
 from collections.abc import MutableMapping
 
 from dask.utils import stringify
@@ -25,9 +27,21 @@ def __init__(self, scheduler):
             "publish_put": self.put,
             "publish_get": self.get,
             "publish_delete": self.delete,
+            "publish_wait_flush": self.flush_wait,
+        }
+        stream_handlers = {
+            "publish_flush_batched_send": self.flush_receive,
         }
 
         self.scheduler.handlers.update(handlers)
+        self.scheduler.stream_handlers.update(stream_handlers)
+        self._flush_received = defaultdict(asyncio.Event)
+
+    def flush_receive(self, uid, **kwargs):
+        self._flush_received[uid].set()
+
+    async def flush_wait(self, uid):
+        await self._flush_received[uid].wait()
 
     @log_errors
     def put(self, keys=None, data=None, name=None, override=False, client=None):

distributed/tests/test_publish.py

@@ -11,6 +11,7 @@
 from distributed.metrics import time
 from distributed.protocol import Serialized
 from distributed.utils_test import gen_cluster, inc
+from distributed.worker import get_worker
 
 
 @gen_cluster()
@@ -301,3 +302,30 @@ async def test_deserialize_client(c, s, a, b):
     from distributed.client import _current_client
 
     assert _current_client.get() is c
+
+
+@gen_cluster(client=True, worker_kwargs={"resources": {"A": 1}})
+async def test_publish_submit_ordering(c, s, a, b):
+    RESOURCES = {"A": 1}
+
+    def _retrieve_annotations():
+        worker = get_worker()
+        task = worker.state.tasks.get(worker.get_current_task())
+        return task.annotations
+
+    # If publish does not take the same comm channel as the submit, it can
+    # happen that the publish message reaches the scheduler before the submit
+    # such that the state of the published future is not the one that has been
+    # requested from the submit. Particularly, this lets us drop annotations
+    # The current implementation does in fact not use the same channel due to
+    # serialization issue (including Futures in BatchedSend appends them to the
+    # "recent messages" log which screws with the refcounting) but ensure that
+    # all queued up messages are flushed and received by the schduler befure
+    # publishing
+    future = c.submit(_retrieve_annotations, resources=RESOURCES, pure=False)
+
+    await c.publish_dataset(future, name="foo")
+    assert await c.list_datasets() == ("foo",)
+
+    result = await future.result()
+    assert result == {"resources": RESOURCES}