Add function to determine when the job checker is assumed dead

app/config.py

@@ -91,7 +91,7 @@ STATUS_AGENT_CONFIG = RUN_CONFIG.get('status_agent', {})
 RUN_CONFIG['status_agent'] = {
     'lock_validity_seconds': 1,
     'sleep_time': 1,
-    'dead_assumption_seconds': 10,
+    'death_assumption_seconds': 10,
     **STATUS_AGENT_CONFIG,
 }
 

app/models/delayed_job_models.py

@@ -209,8 +209,18 @@ class DelayedJob(DB.Model):
         """
         :return: True if the job needs to be checked in lsf, false otherwise
         """
-        needs_to_be_checked_by_status = self.status in [JobStatuses.QUEUED, JobStatuses.RUNNING, JobStatuses.UNKNOWN]
-        return needs_to_be_checked_by_status
+        return self.status in [JobStatuses.QUEUED, JobStatuses.RUNNING, JobStatuses.UNKNOWN]
+
+    def job_checker_seems_to_have_died(self):
+        """
+        :return: True if the checker seems to have died, false otherwise
+        """
+        lsf_script_errored = self.last_lsf_check_status != 0
+        death_assumption_seconds = RUN_CONFIG.get('status_agent').get('death_assumption_seconds')
+        checker_assumed_dead_time = app_utils.get_utc_now() - datetime.timedelta(seconds=death_assumption_seconds)
+        checker_is_assumed_dead = self.last_lsf_checked_at < checker_assumed_dead_time
+
+        return lsf_script_errored and checker_is_assumed_dead
 
 
 # ----------------------------------------------------------------------------------------------------------------------

app/models/test/test_job_lsf_status.py

@@ -6,6 +6,8 @@ import datetime
 
 from app import create_app
 from app.models import delayed_job_models
+from app.config import RUN_CONFIG
+from app import utils
 
 
 class TestJobLSFStatus(unittest.TestCase):
@@ -63,15 +65,33 @@ class TestJobLSFStatus(unittest.TestCase):
             }
             docker_image_url_must_be = 'some_url'
             job = delayed_job_models.get_or_create(job_type, params, docker_image_url_must_be)
+            death_assumption_seconds = RUN_CONFIG.get('status_agent').get('death_assumption_seconds')
+            last_lsf_checked_at = utils.get_utc_now() - datetime.timedelta(seconds=death_assumption_seconds)
+            job.last_lsf_checked_at = last_lsf_checked_at
+            job.last_lsf_check_status = 1
 
-            last_lsf_checked_at = datetime.datetime.utcnow() + datetime.timedelta(seconds=-3)
-            print('last_lsf_checked_at: ', last_lsf_checked_at)
+            job_checker_seems_to_have_died_got = job.job_checker_seems_to_have_died()
+            self.assertTrue(job_checker_seems_to_have_died_got,
+                            msg='When the last script execution errors and the last time it reported is greater than '
+                                'the assumed dead time it must assume that the checker died.')
 
-            # job.last_lsf_checked_at = 0
-            # the last checked at time must be older than the seconds at which it is considered dead
+            job.last_lsf_checked_at = utils.get_utc_now()
+            job_checker_seems_to_have_died_got = job.job_checker_seems_to_have_died()
+            self.assertFalse(job_checker_seems_to_have_died_got,
+                             msg='When the last script execution errors but the last time it reported is less than '
+                                 'the assumed dead time it must NOT assume that the checker died.')
 
-            # for status in [delayed_job_models.JobStatuses.QUEUED, delayed_job_models.JobStatuses.RUNNING,
-            #                delayed_job_models.JobStatuses.UNKNOWN]:
-            #     job.status = status
-            #     needs_to_be_checked_got = job.needs_to_be_checked_in_lsf()
-            #     self.assertTrue(needs_to_be_checked_got, msg=f'A job with status {status} need to be checked in LSF!')
+            job.last_lsf_checked_at = utils.get_utc_now()
+            job.last_lsf_check_status = 0
+            job_checker_seems_to_have_died_got = job.job_checker_seems_to_have_died()
+            self.assertFalse(job_checker_seems_to_have_died_got,
+                             msg='When the last script execution was successful and the last time it reported is less '
+                                 'than the assumed dead time it must NOT assume that the checker died.')
+
+            last_lsf_checked_at = utils.get_utc_now() - datetime.timedelta(seconds=death_assumption_seconds)
+            job.last_lsf_checked_at = last_lsf_checked_at
+            job.last_lsf_check_status = 0
+            job_checker_seems_to_have_died_got = job.job_checker_seems_to_have_died()
+            self.assertFalse(job_checker_seems_to_have_died_got,
+                             msg='When the last script execution was successful but the last time it reported is '
+                                 'greater than the assumed dead time it must NOT assume that the checker died.')