Edoardo Comar created KAFKA-20453:
-------------------------------------
Summary: Metrics of group "kafka.network.SocketServer" should have
tags unique to each SocketServer
Key: KAFKA-20453
URL: https://issues.apache.org/jira/browse/KAFKA-20453
Project: Kafka
Issue Type: Bug
Components: metrics, unit tests
Reporter: Edoardo Comar
kafka.network.SocketServer scala creates metrics which have no tags to
distinguish them from those of another SocketServer
e.g. for MemoryPool
{color:#871094}metricsGroup{color}.newGauge({color:#067d17}"MemoryPoolAvailable"{color},
() => {color:#871094}memoryPool{color}.availableMemory)
This is a problem found in ClusterTests - which run in a single JVM and have a
singleton KafkaYammerRegistry. So there is a single gauge even though there are
multiple MemoryPools.
The test created for https://issues.apache.org/jira/browse/KAFKA-20302
could not use the metric to access the memory pool available memory.
Outside tests, when kafka is running in Combined KRaft mode there are 2 socket
servers per node. So the metric value is the one of the socket server who gers
created first (the controller, where arguably the memory pool is less useful).
A partial fix would be to add a nodeId tag - this would help just ClusterTests
in KRAFT mode but this would still not work for CO-KRAFT mode where we need to
find a way to distinguish the multiple socket servers per process.
Given that co-kraft is not for production modes, perhaps a simple counter of
SocketServer instances could suffice.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
