Randall Hauch created KAFKA-10816:
-------------------------------------
Summary: Connect REST API should have a resource that can be used
as a readiness probe
Key: KAFKA-10816
URL: https://issues.apache.org/jira/browse/KAFKA-10816
Project: Kafka
Issue Type: Improvement
Components: KafkaConnect
Reporter: Randall Hauch
There are a few ways to accurately detect whether a Connect worker is
*completely* ready to process all REST requests:
# Wait for `Herder started` in the Connect worker logs
# Use the REST API to issue a request that will be completed only after the
herder has started, such as `GET /connectors/{name}/` or `GET
/connectors/{name}/status`.
Other techniques can be used to detect other startup states, though none of
these will guarantee that the worker has indeed completely started up and can
process all REST requests:
* `GET /` can be used to know when the REST server has started, but this may be
before the worker has started completely and successfully.
* `GET /connectors` can be used to know when the REST server has started, but
this may be before the worker has started completely and successfully. And, for
the distributed Connect worker, this may actually return an older list of
connectors if the worker hasn't yet completely read through the internal config
topic. It's also possible that this request returns even if the worker is
having trouble reading from the internal config topic.
* `GET /connector-plugins` can be used to know when the REST server has
started, but this may be before the worker has started completely and
successfully.
The Connect REST API should have an endpoint that more obviously and more
simply can be used as a readiness probe. This could be a new resource (e.g.,
`GET /status`), though this would only work on newer Connect runtimes, and
existing tooling, installations, and examples would have to be modified to take
advantage of this feature (if it exists).
Alternatively, we could make sure that the existing resources (e.g., `GET /` or
`GET /connectors`) wait for the herder to start completely; this wouldn't
require a KIP and it would not require clients use different technique for
newer and older Connect runtimes. (Whether or not we back port this is another
question altogether, since it's debatable whether the behavior of the existing
REST resources is truly a bug.)
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
