Gerk Elznik created FLINK-40509:
-----------------------------------
Summary: Move operator defaults out of the Helm chart so
config.yaml carries only user-supplied content
Key: FLINK-40509
URL: https://issues.apache.org/jira/browse/FLINK-40509
Project: Flink
Issue Type: Improvement
Components: Kubernetes Operator
Affects Versions: kubernetes-operator-1.16.0
Reporter: Gerk Elznik
h2. Background
On PR #1126 (FLINK-39791) [~dciupitu] proposed a cleanup so that the operator's
{{config.yaml}} on disk contains only user-supplied content. With no shipped
seed there is no duplicate-key collision surface between the seed and a user's
flat override, which was the residual risk noted in that PR. The approach was
agreed as a follow-up rather than folded into #1126, but it was never filed.
This ticket tracks it.
h2. Proposal
# Drop the operator overrides from {{values.yaml}} (the SLF4J reporter class
and interval, {{kubernetes.operator.reconcile.interval}} and
{{{}kubernetes.operator.observer.progress-check.interval{}}}) and let the
compiled {{ConfigOption}} defaults apply.
# Move the Java 17 opts
({{{}kubernetes.operator.default-configuration.flink-version.*{}}}) out of the
chart's {{conf/config.yaml}} into compiled defaults, leaving that file empty or
a commented example.
# Stop text-writing {{watchNamespaces}} and {{operatorHealth}} into the
operator ConfigMap and set them as env vars instead, mirroring how
{{WATCH_NAMESPACES}} already works.
h2. What was measured while prototyping
Verified in a kind cluster:
* Removing the {{values.yaml}} overrides changes the reconcile cadence from
*15s to 60s* on a default install, because the compiled defaults are 60s and
10s while the chart ships 15s and 5s. Progress check would go from 5s to 10s.
Dropping the SLF4J reporter is also a default-install behavior change, since
operator metrics stop being logged every 5 minutes.
* {{WATCH_NAMESPACES}} scopes the watch with no code change, but it has to be
set on *both* containers. The webhook resolves watched namespaces through the
same config path and has its own {{env:}} block. With the var on the operator
container only and the ConfigMap line removed, the webhook informers stay
cluster-scoped. They are created lazily, so nothing fails at startup and the
first {{FlinkSessionJob}} apply then fails admission with {{{}cannot list
flinkdeployments at the cluster scope{}}}.
* Shipping namespaces through the env var silently disables
{{{}kubernetes.operator.dynamic.namespaces.enabled{}}}. The reload still logs
the new namespace list, but no {{Changing namespaces}} event fires, the webhook
does not rescope, nothing warns, and a FlinkDeployment created in the newly
listed namespace is never reconciled.
* Nothing reads {{KUBERNETES_OPERATOR_HEALTH_PROBE_ENABLED}} or
{{KUBERNETES_OPERATOR_HEALTH_PROBE_PORT}} today, so that half needs new
operator code mirroring {{{}ENV_WATCH_NAMESPACES{}}}. It is only needed to keep
non-default ports and disablement working: the compiled defaults ({{{}true{}}}
and {{{}8085{}}}) already match what the chart writes, so removing those two
ConfigMap lines is behavior-neutral for a default install. The Helm value stays
regardless, since it also wires the containerPort and the liveness and startup
probes.
* The Java 17 opts are dynamic string prefixes resolved at runtime by
{{{}FlinkConfigManager#getRelevantVersionPrefixes{}}}, so they cannot be
expressed as plain {{ConfigOption}} defaults.
h2. Open decisions
# Intervals: change the compiled defaults to 15s and 5s, which also affects
non-Helm deployments and the generated config docs, or let default installs
slow down and carry a release note.
# Where the compiled Java 17 seed gets injected, since user configuration has
to keep winning over it.
# {{defaultConfiguration.append: false}} currently produces a config with no
Java 17 opts at all. Compiled defaults cannot be shed by omitting text, so an
explicit off-switch is probably needed to preserve that escape hatch.
# Whether losing no-restart namespace updates is an acceptable trade for env
vars, given that changing an env var through helm upgrade forces a rolling
restart anyway.
h2. Suggested sequencing
Three separate PRs, smallest blast radius first: the env vars, then the Java 17
opts, then the intervals and SLF4J removal last, since that is the one carrying
a user-visible behavior change and a release note.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
