Luca Castelli created FLINK-41074:
-------------------------------------
Summary: Blue/green deployment can retire the active deployment
using readiness from an older candidate revision
Key: FLINK-41074
URL: https://issues.apache.org/jira/browse/FLINK-41074
Project: Flink
Issue Type: Bug
Components: Kubernetes Operator
Reporter: Luca Castelli
Updating a FlinkBlueGreenDeployment during an ongoing transition can change
Green’s spec while its deletion delay is running. Two related bugs can then
cause premature retirement of Blue:
# Stale readiness: Green is considered ready based on RUNNING/STABLE status
even when status.observedGeneration is behind metadata.generation.
# Inherited deletion delay: Once the updated Green revision becomes ready, it
reuses the previous revision’s deploymentReadyTimestamp instead of receiving a
fresh deletion delay.
*Reproduction*
# Deploy Blue, then update the parent to initiate Green.
# Wait until Green is ready and the configured deletion delay starts.
# Update the parent again, for example by changing job parallelism.
# Observe that Blue can retire either:
*
** while Green’s new generation remains unobserved, or
** before a full deletion delay has elapsed for the newly ready generation.
*Expected behavior*
Require readiness for Green’s current desired generation and bind the
deletion-delay timestamp to that generation. Relevant candidate updates or
observed readiness loss should invalidate previous readiness evidence.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
