Darren Weber created AIRFLOW-5889:
-------------------------------------
Summary: AWS Batch Operator - API request limits should not fail a
task
Key: AIRFLOW-5889
URL: https://issues.apache.org/jira/browse/AIRFLOW-5889
Project: Apache Airflow
Issue Type: Improvement
Components: aws, contrib
Affects Versions: 1.10.4
Reporter: Darren Weber
Assignee: Darren Weber
Fix For: 2.0.0, 1.10.6
The AWS Batch Operator attempts to use a boto3 feature that is not available
and has not been merged in years, see
- [https://github.com/boto/botocore/pull/1307]
- see also [https://github.com/broadinstitute/cromwell/issues/4303]
This is a curious case of premature optimization. So, in the meantime, this
means that the fallback is the exponential backoff routine for the status
checks on the batch job. Unfortunately, when the concurrency of Airflow jobs is
very high (100's of tasks), this fallback polling hits the AWS Batch API too
hard and the AWS API throttle throws an error, which fails the Airflow task,
simply because the status is polled too frequently.
Check the output from the retry algorithm, e.g. within the first 10 retries,
the status of an AWS batch job is checked about 10 times at a rate that is
approx 1 retry/sec. When an Airflow instance is running 10's or 100's of
concurrent batch jobs, this hits the API too frequently and crashes the Airflow
task (plus it occupies a worker in too much busy work).
{code:java}
In [4]: [1 + pow(retries * 0.1, 2) for retries in range(20)]
Out[4]:
[1.0,
1.01,
1.04,
1.09,
1.1600000000000001,
1.25,
1.36,
1.4900000000000002,
1.6400000000000001,
1.81,
2.0,
2.21,
2.4400000000000004,
2.6900000000000004,
2.9600000000000004,
3.25,
3.5600000000000005,
3.8900000000000006,
4.24,
4.61]{code}
Possible solutions are to introduce an initial sleep (say 60 sec?) right after
issuing the request, so that the batch job has some time to spin up. The job
progresses through a through phases before it gets to RUNNING state and polling
for each phase of that sequence might help. Since batch jobs tend to be
long-running jobs (rather than near-real time jobs), it might help to issue
less frequent polls when it's in the RUNNING state. Something on the order of
10's seconds might be reasonable for batch jobs? Maybe the class could expose a
parameter for the rate of polling (or a callable)?
Another option is to use something like the sensor-poke approach, with
rescheduling, e.g.
-
[https://github.com/apache/airflow/blob/master/airflow/sensors/base_sensor_operator.py#L117]
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
