Fix race condition on task retry

arq/worker.py

@@ -447,8 +447,10 @@ async def start_jobs(self, job_ids: List[bytes]) -> None:
                 await pipe.watch(in_progress_key)
                 ongoing_exists = await pipe.exists(in_progress_key)
                 score = await pipe.zscore(self.queue_name, job_id)
-                if ongoing_exists or not score:
+                if ongoing_exists or not score or score > timestamp_ms():
                     # job already started elsewhere, or already finished and removed from queue
+                    # if score > ts_now,
+                    # it means probably the job was re-enqueued with a delay in another worker
                     self.job_counter = self.job_counter - 1
                     self.sem.release()
                     logger.debug('job %s already running elsewhere', job_id)

tests/test_worker.py

@@ -160,6 +160,40 @@ async def test_job_successful(arq_redis: ArqRedis, worker, caplog):
     assert 'X.XXs → testing:foobar()\n  X.XXs ← testing:foobar ● 42' in log
 
 
+async def test_job_retry_race_condition(arq_redis: ArqRedis, worker):
+    async def retry_job(ctx):
+        if ctx['job_try'] == 1:
+            raise Retry(defer=10)
+
+    job_id = 'testing'
+    await arq_redis.enqueue_job('retry_job', _job_id=job_id)
+
+    worker_one: Worker = worker(functions=[func(retry_job, name='retry_job')])
+    worker_two: Worker = worker(functions=[func(retry_job, name='retry_job')])
+
+    assert worker_one.jobs_complete == 0
+    assert worker_one.jobs_failed == 0
+    assert worker_one.jobs_retried == 0
+
+    assert worker_two.jobs_complete == 0
+    assert worker_two.jobs_failed == 0
+    assert worker_two.jobs_retried == 0
+
+    await worker_one.start_jobs([job_id.encode()])
+    await asyncio.gather(*worker_one.tasks.values())
+
+    await worker_two.start_jobs([job_id.encode()])
+    await asyncio.gather(*worker_two.tasks.values())
+
+    assert worker_one.jobs_complete == 0
+    assert worker_one.jobs_failed == 0
+    assert worker_one.jobs_retried == 1
+
+    assert worker_two.jobs_complete == 0
+    assert worker_two.jobs_failed == 0
+    assert worker_two.jobs_retried == 0
+
+
 async def test_job_successful_no_result_logging(arq_redis: ArqRedis, worker, caplog):
     caplog.set_level(logging.INFO)
     await arq_redis.enqueue_job('foobar', _job_id='testing')