Fix workflow stop in online mode by sending stop command to RQ worker

.github/workflows/ci.yml

@@ -20,7 +20,7 @@ jobs:
       run: |
         python -m pip install --upgrade pip
         pip install -r requirements.txt # test with requirements file so can easily bump with dependabot
-        pip install pytest
+        pip install pytest fakeredis
     - name: Test
       run: |
         python -m pytest test_gui.py tests/

src/workflow/QueueManager.py

@@ -178,14 +178,16 @@ def get_job_info(self, job_id: str) -> Optional[JobInfo]:
 
             job = Job.fetch(job_id, connection=self._redis)
 
-            # Map RQ status to our enum
+            # 'stopped' is what RQ records after send_stop_job_command runs;
+            # surface it as CANCELED so the UI doesn't show stopped jobs as queued.
             status_map = {
                 "queued": JobStatus.QUEUED,
                 "started": JobStatus.STARTED,
                 "finished": JobStatus.FINISHED,
                 "failed": JobStatus.FAILED,
                 "deferred": JobStatus.DEFERRED,
                 "canceled": JobStatus.CANCELED,
+                "stopped": JobStatus.CANCELED,
             }
 
             status = status_map.get(job.get_status(), JobStatus.QUEUED)
@@ -232,24 +234,61 @@ def cancel_job(self, job_id: str) -> bool:
         """
         Cancel a queued or running job.
 
+        For queued jobs, this removes them from the queue. For jobs that are
+        already executing in a worker, Job.cancel() alone is not enough — it
+        only updates Redis registries while the worker keeps running the
+        workflow. We send a stop-job command to the worker so the work-horse
+        is actually interrupted.
+
         Args:
             job_id: The job ID to cancel
 
         Returns:
-            True if successfully canceled
+            True if the job is canceled (or already was), False otherwise.
         """
         if not self.is_available:
             return False
 
         try:
+            from rq.command import send_stop_job_command
+            from rq.exceptions import InvalidJobOperation, NoSuchJobError
             from rq.job import Job
+        except ImportError:
+            return False
 
+        try:
             job = Job.fetch(job_id, connection=self._redis)
-            job.cancel()
+        except NoSuchJobError:
+            return False
+        except Exception:
+            return False
+
+        # Idempotent: a second Stop click (or rerun) should not surface an error.
+        if job.is_canceled or job.is_stopped:
             return True
+
+        # Tell the worker to interrupt the work-horse before marking canceled.
+        if job.is_started and job.worker_name:
+            try:
+                send_stop_job_command(self._redis, job_id)
+            except InvalidJobOperation:
+                # The worker just finished or the job moved out of 'started';
+                # fall through to cancel() to settle registry state.
+                pass
+            except Exception:
+                pass
+
+        try:
+            job.cancel()
+        except InvalidJobOperation:
+            # Worker already transitioned the job (e.g. to 'stopped'); that
+            # satisfies the user's intent to stop.
+            pass
         except Exception:
             return False
 
+        return True
+
     def get_queue_stats(self) -> dict:
         """
         Get queue statistics.

tests/test_queue_manager_cancel.py

@@ -0,0 +1,161 @@
+"""
+Tests for QueueManager.cancel_job - the "stop workflow" path used in online
+mode where workflows are executed by RQ workers (the vendor's queue).
+
+Bug being fixed: when a workflow is mid-execution in an RQ worker and the
+user clicks "Stop Workflow", QueueManager.cancel_job calls Job.cancel() on
+the RQ Job. For a job in the "started" state this only marks the job as
+canceled in the Redis registries; the worker keeps executing the workflow
+and the user sees inconsistent / "weird" state (worker still appending to
+logs, status flipping around, etc.).
+
+To actually stop a running RQ job, RQ exposes
+rq.command.send_stop_job_command(connection, job_id) which messages the
+worker over Redis pubsub to interrupt the work-horse.
+"""
+
+import os
+import sys
+
+import pytest
+
+fakeredis = pytest.importorskip("fakeredis")
+rq = pytest.importorskip("rq")
+from rq import Queue
+from rq.job import Job, JobStatus
+
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+from src.workflow.QueueManager import QueueManager
+
+
+def _make_queue_manager() -> QueueManager:
+    """Build a QueueManager wired to fake Redis, bypassing __init__."""
+    qm = QueueManager.__new__(QueueManager)
+    qm._redis = fakeredis.FakeStrictRedis()
+    qm._queue = Queue(QueueManager.QUEUE_NAME, connection=qm._redis)
+    qm._is_online = True
+    qm._init_attempted = True
+    qm._default_timeout = 7200
+    qm._default_result_ttl = 86400
+    return qm
+
+
+def _force_started(job: Job, worker_name: str = "rq:worker:test-worker") -> None:
+    """Move a queued job into the 'started' state with a worker assigned."""
+    job.set_status(JobStatus.STARTED)
+    job.worker_name = worker_name
+    job.save()
+
+
+def test_cancel_queued_job_marks_it_canceled():
+    qm = _make_queue_manager()
+    qm._queue.enqueue(os.getcwd, job_id="queued-job")
+
+    assert qm.cancel_job("queued-job") is True
+
+    refreshed = Job.fetch("queued-job", connection=qm._redis)
+    assert refreshed.get_status() == JobStatus.CANCELED
+
+
+def test_cancel_started_job_sends_stop_command_to_worker(monkeypatch):
+    """
+    Reproduces the vendor-queue stop bug.
+
+    A workflow that is actively running in a worker must be stopped by
+    sending a stop-job command to the worker. The previous implementation
+    only called Job.cancel(), which left the worker running.
+    """
+    qm = _make_queue_manager()
+    job = qm._queue.enqueue(os.getcwd, job_id="started-job")
+    _force_started(job)
+
+    stop_calls: list[str] = []
+
+    def fake_send_stop_job_command(connection, job_id, *args, **kwargs):
+        stop_calls.append(job_id)
+
+    import rq.command as rq_command
+    monkeypatch.setattr(
+        rq_command, "send_stop_job_command", fake_send_stop_job_command
+    )
+
+    result = qm.cancel_job("started-job")
+
+    assert result is True, "cancel_job should report success for started jobs"
+    assert stop_calls == ["started-job"], (
+        "cancel_job must call send_stop_job_command for started jobs - "
+        "otherwise the RQ worker keeps running the workflow."
+    )
+
+
+def test_cancel_already_canceled_job_does_not_raise():
+    """
+    User double-clicks 'Stop Workflow' or stop is invoked twice on rerun.
+    The second call must not surface InvalidJobOperation as a 'weird error'.
+    """
+    qm = _make_queue_manager()
+    job = qm._queue.enqueue(os.getcwd, job_id="dup-cancel")
+    job.cancel()
+    assert Job.fetch("dup-cancel", connection=qm._redis).get_status() == JobStatus.CANCELED
+
+    # Must not raise; intent (job is canceled) is already satisfied.
+    assert qm.cancel_job("dup-cancel") is True
+
+
+def test_cancel_missing_job_returns_false():
+    qm = _make_queue_manager()
+    assert qm.cancel_job("does-not-exist") is False
+
+
+def test_started_status_without_worker_is_handled_gracefully(monkeypatch):
+    """
+    Edge case: job is marked started but has no worker_name yet (race between
+    worker pickup and stop click). cancel_job must not raise; it should fall
+    back to canceling the job in the registry.
+    """
+    qm = _make_queue_manager()
+    job = qm._queue.enqueue(os.getcwd, job_id="started-no-worker")
+    job.set_status(JobStatus.STARTED)
+    job.save()
+
+    stop_calls: list[str] = []
+
+    def fake_send_stop_job_command(connection, job_id, *args, **kwargs):
+        stop_calls.append(job_id)
+
+    import rq.command as rq_command
+    monkeypatch.setattr(
+        rq_command, "send_stop_job_command", fake_send_stop_job_command
+    )
+
+    result = qm.cancel_job("started-no-worker")
+
+    assert result is True
+    # Without a worker_name there is nothing to send the stop command to.
+    assert stop_calls == []
+    assert (
+        Job.fetch("started-no-worker", connection=qm._redis).get_status()
+        == JobStatus.CANCELED
+    )
+
+
+def test_stopped_status_is_mapped_in_get_job_info(monkeypatch):
+    """
+    After send_stop_job_command runs, RQ marks the job 'stopped'. The status
+    map in get_job_info must recognise it; otherwise the UI would show the
+    job as still 'queued', which is the user-visible 'weird error'.
+    """
+    qm = _make_queue_manager()
+    job = qm._queue.enqueue(os.getcwd, job_id="stopped-job")
+    job.set_status(JobStatus.STOPPED)
+    job.save()
+
+    info = qm.get_job_info("stopped-job")
+    assert info is not None
+    assert info.status == __import__(
+        "src.workflow.QueueManager", fromlist=["JobStatus"]
+    ).JobStatus.CANCELED, (
+        "RQ 'stopped' status should be reported as CANCELED to the UI; "
+        "otherwise stopped jobs appear stuck in 'queued'."
+    )