* Extract Duplicate Code into Helper Method
DRY up the tests and replace the arbitrary `sleep()` calls with a more
explicit check before progressing.
* Improve Testing of batch:TerminateJob
The test now confirms that the job was terminated by sandwiching a `sleep`
command between two `echo` commands. In addition to the original checks
of the terminated job status/reason, the test now asserts that only the
first echo command succeeded, confirming that the job was indeed terminated
while in progress.
* Fix Race Condition in batch:SubmitJob
The `test_submit_job` in `test_batch.py` kicks off a job, calls `describe_jobs`
in a loop until the job status returned is SUCCEEDED, and then asserts against
the logged events.
The backend code that runs the submitted job does so in a separate thread. If
the job was successful, the job status was being set to SUCCEEDED *before* the
event logs had been written to the logging backend.
As a result, it was possible for the primary thread running the test to detect
that the job was successful immediately after the secondary thread had updated
the job status but before the secondary thread had written the logs to the
logging backend. Under the right conditions, this could cause the subsequent
logging assertions in the primary thread to fail.
Additionally, the code that collected the logs from the container was using
a "dodgy hack" of time.sleep() and a modulo-based conditional that was
ultimately non-deterministic and could result in log messages being dropped
or duplicated in certain scenarios.
In order to address these issues, this commit does the following:
* Carefully re-orders any code that sets a job status or timestamp
to avoid any obvious race conditions.
* Removes the "dodgy hack" in favor of a much more straightforward
(and less error-prone) method of collecting logs from the container.
* Removes arbitrary and unnecessary calls to time.sleep()
Before applying any changes, the flaky test was failing about 12% of the
time. Putting a sleep() call between setting the `job_status` to SUCCEEDED
and collecting the logs, resulted in a 100% failure rate. Simply moving
the code that sets the job status to SUCCEEDED to the end of the code block,
dropped the failure rate to ~2%. Finally, removing the log collection
hack allowed the test suite to run ~1000 times without a single failure.
Taken in aggregate, these changes make the batch backend more deterministic
and should put the nail in the coffin of this flaky test.
Closes#3475
f7467164e4