sanghyeok An created KAFKA-21515:
------------------------------------
Summary: A Streams static rejoin rejected by topology validation
causes a subsequent valid rejoin to receive an empty task assignment
Key: KAFKA-21515
URL: https://issues.apache.org/jira/browse/KAFKA-21515
Project: Kafka
Issue Type: Bug
Components: group-coordinator, streams
Reporter: sanghyeok An
Assignee: sanghyeok An
When a Streams static member’s rejoin is rejected by topology validation,
changes made to the intermediate assignment cache can remain. A subsequent
member rejoining with the correct topology can therefore receive an empty
assignment in its first response, even though the previous task assignment is
available for inheritance.
*Scenario*
# Static member A is operating normally and owns task T. The coordinator has
an intermediate assignment cached for the current assignment epoch.
# A temporarily leaves the group normally ({{{}memberEpoch=-2{}}}). The
coordinator preserves its member information and target assignment so they can
be inherited upon rejoining.
# Before the session expires, B rejoins with A’s instance ID and a new member
ID. Due to a deployment error, for example, B submits different topology
contents while retaining the existing topology epoch.
# While preparing the static member replacement, the coordinator changes the
intermediate assignment cache key from A to B. It then detects the topology
mismatch and rejects B’s request with {{{}STREAMS_INVALID_TOPOLOGY_EPOCH{}}}.
# B’s member replacement is not applied, so the stored member and target
assignment still refer to A. However, the intermediate assignment cache retains
B’s key.
# Before the session expires, C rejoins with the original topology restored,
the same instance ID and process ID, and a new member ID. The relevant topology
epoch, metadata, and assignment configuration remain unchanged.
# C’s rejoin succeeds, but the coordinator cannot find A’s cached entry to
transfer to C. The first response therefore contains an empty active assignment
instead of task T.
Without B’s rejected join attempt, C receives the existing task T in its first
rejoin response. The cache mutation from the rejected request changes the
outcome of a subsequent valid request.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)