Prevent incorrect reschedule of resumed tasks with jobboard

A previous patch [0] introduced a keepalive mechanism that allows a
controller to notify taskflow that a task is being executed.
But it doesn't handle "resumed" tasks (tasks that have been rescheduled
from one controller to another controller).

This patch moves this keepalive task in a thread that is attached to a
taskflow conductor, the thread will extend the expiration of the tasks
that are executed by this conductor.

[0] I6619ba117f7051fe44086389789bc6e2810fd23d

Story 2009998
Task 45089

Change-Id: I29fcad9e121a30d6e8f8178f2f078cf10771a32a

octavia/common/base_taskflow.py

@@ -16,6 +16,7 @@
 import concurrent.futures
 import datetime
 import functools
+import time
 
 from oslo_config import cfg
 from oslo_log import log
@@ -153,17 +154,9 @@ class TaskFlowServiceController(object):
 
     def _wait_for_job(self, job_board):
         # Wait for job to its complete state
-        expiration_time = CONF.task_flow.jobboard_expiration_time
-
-        need_wait = True
-        while need_wait:
-            need_wait = False
         for job in job_board.iterjobs():
-                # If job hasn't finished in expiration_time/2 seconds,
+            LOG.debug("Waiting for job %s to finish", job.name)
-                # extend its TTL
+            job.wait()
-                if not job.wait(timeout=expiration_time / 2):
-                    job.extend_expiry(expiration_time)
-                    need_wait = True
 
     def run_conductor(self, name):
         with self.driver.persistence_driver.get_persistence() as persistence:
@@ -187,4 +180,35 @@ class TaskFlowServiceController(object):
                         name, board, persistence=persistence,
                         engine=CONF.task_flow.engine)
 
+                waiter_th = concurrent.futures.ThreadPoolExecutor(
+                    max_workers=1)
+                waiter_th.submit(self._waiter, conductor)
+
                 conductor.run()
+
+    def _extend_jobs(self, conductor, expiration_time):
+        conductor_name = conductor._name
+
+        with self.driver.persistence_driver.get_persistence() as persistence:
+            with self.driver.job_board(persistence) as board:
+                for job in board.iterjobs():
+                    try:
+                        owner = board.find_owner(job)
+                    except TypeError:
+                        # taskflow throws an exception if a job is not owned
+                        # (probably a bug in taskflow)
+                        continue
+                    # Only extend expiry for jobs that are owner by our
+                    # conductor (from the same process)
+                    if owner == conductor_name:
+                        if job.expires_in() < expiration_time / 2:
+                            LOG.debug("Extend expiry for job %s", job.name)
+                            job.extend_expiry(expiration_time)
+
+    def _waiter(self, conductor):
+        expiration_time = CONF.task_flow.jobboard_expiration_time
+
+        while True:
+            self._extend_jobs(conductor, expiration_time)
+
+            time.sleep(expiration_time / 4)

octavia/tests/unit/common/test_base_taskflow.py

@@ -119,25 +119,20 @@ class TestTaskFlowServiceController(base.TestCase):
 
     def test__wait_for_job(self):
         job1 = mock.MagicMock()
-        job1.wait.side_effect = [False, True]
         job2 = mock.MagicMock()
-        job2.wait.side_effect = [False, True]
-        job3 = mock.MagicMock()
-        job3.wait.return_value = True
         job_board = mock.MagicMock()
         job_board.iterjobs.side_effect = [
-            [job1, job2, job3],
             [job1, job2]
         ]
         self.service_controller._wait_for_job(job_board)
 
-        job1.extend_expiry.assert_called_once()
+        job1.wait.assert_called_once()
-        job2.extend_expiry.assert_called_once()
+        job2.wait.assert_called_once()
-        job3.extend_expiry.assert_not_called()
 
     @mock.patch('octavia.common.base_taskflow.RedisDynamicLoggingConductor')
     @mock.patch('octavia.common.base_taskflow.DynamicLoggingConductor')
-    def test_run_conductor(self, dynamiccond, rediscond):
+    @mock.patch('concurrent.futures.ThreadPoolExecutor')
+    def test_run_conductor(self, mock_threadpoolexec, dynamiccond, rediscond):
         self.service_controller.run_conductor("test")
         rediscond.assert_called_once_with(
             "test", self.jobboard_mock.__enter__(),
@@ -154,3 +149,27 @@ class TestTaskFlowServiceController(base.TestCase):
             "test2", self.jobboard_mock.__enter__(),
             persistence=self.persistence_mock.__enter__(),
             engine='parallel')
+
+    def test__extend_jobs(self):
+        conductor = mock.MagicMock()
+        conductor._name = 'mycontroller'
+
+        job1 = mock.MagicMock()
+        job1.expires_in.return_value = 10
+        job2 = mock.MagicMock()
+        job2.expires_in.return_value = 10
+        job3 = mock.MagicMock()
+        job3.expires_in.return_value = 30
+        self.jobboard_mock.__enter__().iterjobs.return_value = [
+            job1, job2, job3]
+
+        self.jobboard_mock.__enter__().find_owner.side_effect = [
+            'mycontroller',
+            TypeError('no owner'),
+            'mycontroller']
+
+        self.service_controller._extend_jobs(conductor, 30)
+
+        job1.extend_expiry.assert_called_once_with(30)
+        job2.extend_expiry.assert_not_called()
+        job3.extend_expiry.assert_not_called()

releasenotes/notes/fix-reschedule-of-jobboard-tasks-929c066dea9267fd.yaml

@@ -0,0 +1,5 @@
+---
+fixes:
+  - |
+    Fix the rescheduling of taskflow tasks that have been resumed after being
+    interrupted.