[
https://issues.apache.org/jira/browse/FLINK-38548?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18032620#comment-18032620
]
Daniel Rossos commented on FLINK-38548:
---------------------------------------
Hi [~premsantosh],
Was just writing up and issue for this as I saw this notification pop-up so
thanks for raising this!
Another issue we were facing is if you have your FlinkBlueGreenDeployment name
set to any keyword key name / subset of a key name it will do a find and
replace on all parts of the spec, not just limited to yaml values but also the
keys. This can create invalid key/fields that will not be properly set in the
FlinkDeployment spec leading to them being seen as missing / invalid.
An easy to see example of this is with a FlinkBlueGreenDeployment that is named
"flink". Your prepared spec will contain the fields like
`"flink-blueConfiguration"` and `"flink-blueVersion"` instead of
`"flinkConfiguration"` and `"flinkVersion"`. This breaks the spec and makes
entire deployment invalid.
Luckily though, I believe the direction your PR is going will cover this
alternative case as well
> Blue/Green Deployments rename too many spec fields
> --------------------------------------------------
>
> Key: FLINK-38548
> URL: https://issues.apache.org/jira/browse/FLINK-38548
> Project: Flink
> Issue Type: Bug
> Components: Kubernetes Operator
> Reporter: Prem Santosh
> Priority: Minor
> Labels: pull-request-available
>
> Hi [~gyfora] and [~schongloo], thanks so much for contributing FLINK-38420.
> In our initial testing of the feature, we found that the adjustNameReferences
> method blanket replaces all occurrences of the deployment name in the spec
> with a suffixed version of the name with -blue or -green:
> [https://github.com/apache/flink-kubernetes-operator/blob/main/flink-kubernetes-operator/src/main/java/org/apache/flink/kubernetes/operator/utils/bluegreen/BlueGreenUtils.java#L62-L81]
> {code:java}
> public static <T> T adjustNameReferences(
> T spec,
> String deploymentName,
> String childDeploymentName,
> String wrapperKey,
> Class<T> valueType) {
> String serializedSpec = SpecUtils.writeSpecAsJSON(spec, wrapperKey);
> String replacedSerializedSpec =
> serializedSpec.replace(deploymentName, childDeploymentName);
> return SpecUtils.readSpecFromJSON(replacedSerializedSpec, wrapperKey,
> valueType);
> }
> {code}
> We use the deployment name in various fields such as metrics scope,
> checkpoint and savepoint directories, and even configmap names in the
> taskmanager and jobmanager pod templates. The blanket replacement breaks
> operations such as restoring the deployment from savepoint since the
> initialSavepointPath field also gets rewriten.
> We would like to submit a patch to remove the call to adjustNameReferences,
> and change only the child deployment names to include the blue/green suffixes
> if possible. Creating a Jira ticket first per the contributor guide.
> Sharing the code change we'll submit as a PR upstream if the contributors
> agree: [https://github.com/lyft/flink-kubernetes-operator/pull/1/files]
> Thanks!
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
