Dennis-Mircea Ciupitu created FLINK-39819:
---------------------------------------------
Summary: Avoid redundant JSON String round-trips in
SpecUtils.clone and DiffResult.toString
Key: FLINK-39819
URL: https://issues.apache.org/jira/browse/FLINK-39819
Project: Flink
Issue Type: Improvement
Components: Kubernetes Operator
Affects Versions: kubernetes-operator-1.15.0
Reporter: Dennis-Mircea Ciupitu
Fix For: kubernetes-operator-1.16.0
h1. Summary
The operator deep-copies CRD objects and renders spec diffs by routing data
through a redundant JSON String. Both code paths serialize an in-memory object
to a String and then immediately parse that String back into an object or a
tree. The intermediate String is pure overhead and can be removed by converting
directly between the object and the in-memory representation.
h1. Motivation
Two utility paths build an intermediate JSON String that nothing else ever
consumes.
h2. SpecUtils.clone (hot path)
{{SpecUtils.clone}} is used by {{ReconciliationUtils.clone}} to deep-copy Flink
resources and their nested specs. It runs on every reconcile cycle, for every
managed resource. Each controller clones the full custom resource at the start
of reconciliation ({{FlinkDeploymentController}},
{{FlinkSessionJobController}}), and additional clones happen for specs and pod
templates. The current implementation serializes the object to a String and
then parses that String back, which allocates and re-tokenizes the entire
object graph on a path that executes constantly.
h2. DiffResult.toString (cold path)
{{DiffResult.toString}} builds the human-readable message for the
{{SpecChanged}} Kubernetes event. It converts the before and after values to
JSON trees by going through an intermediate String. This path is low frequency
because it only runs when a user changes a spec, but it carries the same
redundant round-trip and the same pattern worth cleaning up for consistency.
h1. Expected Impact
Removing the String hop lets Jackson convert straight from the object to its
in-memory form (object to tree to object for the clone, object to tree for the
diff). This avoids one full serialization to text plus the matching parse,
reducing allocations and CPU on the per-reconcile clone path. The clone is the
meaningful win given how often it runs. The diff change is a consistency
cleanup with negligible runtime effect.
The change is behavior preserving for the CRD and Fabric8 model types involved,
which are JSON-native by design and already round-tripped through JSON by
Kubernetes. Existing reconciler and diff tests cover the affected code.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
