chomingi created KAFKA-20951:
--------------------------------
Summary: StreamsMetricsImpl retains duplicate sensor names when a
sensor is removed and recreated
Key: KAFKA-20951
URL: https://issues.apache.org/jira/browse/KAFKA-20951
Project: Kafka
Issue Type: Improvement
Components: metrics, streams
Affects Versions: 4.3.0
Reporter: chomingi
Assignee: chomingi
StreamsMetricsImpl keeps a per-level collection of sensor names
(clientLevelSensors, threadLevelSensors, taskLevelSensors, nodeLevelSensors,
topicLevelSensors, cacheLevelSensors, and storeLevelSensors) so that the
corresponding sensors can be removed later.
A name is added to its collection in getSensors() or clientLevelSensor() only
when metrics.getSensor(fullSensorName) returns null. removeSensor(Sensor)
removes the sensor from the Metrics registry but leaves its name in the
collection.
If the same sensor is later recreated, the registry lookup returns null again
and the same name is recorded a second time. Repeating this cycle continues to
append duplicate entries.
RecordQueue#close(), RecordCollectorImpl#removeAllProducedSensors(), and
DefaultStateUpdater.StateUpdaterMetrics#clear() remove sensors through this
path.
If the corresponding sensors are later recreated, their names are recorded
again.
The sensors and their metrics are not retained; only their names remain.
Therefore, the impact is limited to these bookkeeping collections accumulating
duplicate entries over the lifetime of a client.
KAFKA-9152 avoids duplicate bookkeeping while a sensor remains registered.
After a sensor is removed, however, the registry-based guard no longer detects
the existing bookkeeping entry when the sensor is recreated.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
