Giannis Polyzos created FLINK-41459:
---------------------------------------
Summary: Documentation Does Not Match The Behaviour Of Health
Checks, Periodic Savepoints And Disabling The Autoscaler
Key: FLINK-41459
URL: https://issues.apache.org/jira/browse/FLINK-41459
Project: Flink
Issue Type: Improvement
Components: Kubernetes Operator, Documentation
Affects Versions: kubernetes-operator-1.16.1
Reporter: Giannis Polyzos
h3. Summary
Three operator behaviours differ from the documentation or are not documented.
Each affects how a production rollout is planned.
h3. Problem
||Area||Documentation||Actual behaviour||
|Checkpoint progress health check|The job management page states a default
window of 5 minutes|Since FLINK-35126 the window is derived from the checkpoint
interval, timeout and tolerated failures, and a shorter configured window is
ignored. With a 10 minute timeout and 10 tolerated failures the effective
window is 2 hours|
|Periodic savepoint schedule|The snapshot management page states that cron
schedules are not supported|The interval option also accepts a cron expression|
|Disabling the autoscaler|Not described|The job is not redeployed, and the last
scaling decisions remain in effect until the resource is deleted or the
autoscaler state is reset|
h3. Impact
* Users expect a job with stalled checkpoints to be restarted within minutes,
when with common settings it takes hours.
* Users may rely on, or avoid, cron schedules based on documentation that
contradicts the implementation.
* Users who disable the autoscaler expect the job to return to its configured
parallelism, which does not happen.
h3. Proposed Solution
* Update the job management page to describe how the checkpoint progress window
is derived.
* Either document cron support for periodic savepoints, or confirm it is
unintended and remove it.
* Describe in the autoscaler page what disabling the autoscaler does and does
not change, and how to return a job to its configured parallelism.
A pull request can follow once the intended behaviour for cron schedules is
confirmed.
h3. Environment
Flink Kubernetes Operator 1.16.1.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
