kafka exporter调研与改进

原创

皮皮熊

修改于 2022-02-28 20:16:19

7.4K5

修改于 2022-02-28 20:16:19

文章被收录于专栏：大数据与实时计算大数据与实时计算

2020年7月的总结文章，2021年再回顾的话发现kafka exporter原作者目前花在本项目的时间过少，很多PR没有处理。好在代码比较清晰，自己拉个独立分支动手修改代码问题也不大。

一、kafka exporter概况

1、项目地址

Github: https://github.com/danielqsj/kafka_exporter

2、项目状态

项目watch、star、fork数量均领先竞品，issue、pull request也比较活跃。

截至2020-07-07:

v0.2.0版本已被Prometheus项目官方推荐
目前官方最新版本已经更新到v1.2.0

3、项目优势

kafka exporter 通过 Kafka Protocol Specification 收集 Brokers, Topics 以及 Consumer Groups的相关指标。

使用简单
配置简单
部署方便支持docker、k8s
运行高效

相比于以往通过kafka内置的脚本进行收集，由于没有了每次脚本启动JVM的开销，指标收集时间从分钟级别降到秒级别，便于大规模集群的监控。

生态丰富
无缝对接prometheus、grafana
grafana有大量开源的DashBoard配置

4、kafka官方项目

KIP-575: build a Kafka-Exporter by Java

kafka后面可能会推出官方的Kafka-Exporter

二、实现分析

1、方案架构

kafka exporter代码层借助大量开源库，所以功能强大但代码量极少，仅600+行，大致架构如下:

Kingpin > go的一个命令行库，处理用户输入的参数
sarama(核心) > go实现的kafka客户端，连接broker获取相关的指标与元数据
kazoo > go实现的zk客户端，连接kafka的zk集群，主要用于zk消费组的lag计算
promhttp > 用于生成 Prometheus HTTP服务器，供prometheus pull指标
其他组件 > 协助将 sarama 和kazoo获取的指标转换成Prometheus的数据格式

2、详细指标

github上指标说明有些滞后，这里加上新的一些指标说明

2.1 Brokers

Name	Exposed informations
`kafka_brokers`	Number of Brokers in the Kafka Cluster

# HELP kafka_brokers Number of Brokers in the Kafka Cluster.
# TYPE kafka_brokers gauge
kafka_brokers 3

2.2 Topics

Name	Exposed informations
`kafka_topic_partitions`	Number of partitions for this Topic
`kafka_topic_partition_current_offset`	Current Offset of a Broker at Topic/Partition
`kafka_topic_partition_oldest_offset`	Oldest Offset of a Broker at Topic/Partition
`kafka_topic_partition_in_sync_replica`	Number of In-Sync Replicas for this Topic/Partition
`kafka_topic_partition_leader`	Leader Broker ID of this Topic/Partition
`kafka_topic_partition_leader_is_preferred`	1 if Topic/Partition is using the Preferred Broker
`kafka_topic_partition_replicas`	Number of Replicas for this Topic/Partition
`kafka_topic_partition_under_replicated_partition`	1 if Topic/Partition is under Replicated

# HELP kafka_topic_partitions Number of partitions for this Topic
# TYPE kafka_topic_partitions gauge
kafka_topic_partitions{topic="__consumer_offsets"} 50

# HELP kafka_topic_partition_current_offset Current Offset of a Broker at Topic/Partition
# TYPE kafka_topic_partition_current_offset gauge
kafka_topic_partition_current_offset{partition="0",topic="__consumer_offsets"} 0

# HELP kafka_topic_partition_oldest_offset Oldest Offset of a Broker at Topic/Partition
# TYPE kafka_topic_partition_oldest_offset gauge
kafka_topic_partition_oldest_offset{partition="0",topic="__consumer_offsets"} 0

# HELP kafka_topic_partition_in_sync_replica Number of In-Sync Replicas for this Topic/Partition
# TYPE kafka_topic_partition_in_sync_replica gauge
kafka_topic_partition_in_sync_replica{partition="0",topic="__consumer_offsets"} 3

# HELP kafka_topic_partition_leader Leader Broker ID of this Topic/Partition
# TYPE kafka_topic_partition_leader gauge
kafka_topic_partition_leader{partition="0",topic="__consumer_offsets"} 0

# HELP kafka_topic_partition_leader_is_preferred 1 if Topic/Partition is using the Preferred Broker
# TYPE kafka_topic_partition_leader_is_preferred gauge
kafka_topic_partition_leader_is_preferred{partition="0",topic="__consumer_offsets"} 1

# HELP kafka_topic_partition_replicas Number of Replicas for this Topic/Partition
# TYPE kafka_topic_partition_replicas gauge
kafka_topic_partition_replicas{partition="0",topic="__consumer_offsets"} 3

# HELP kafka_topic_partition_under_replicated_partition 1 if Topic/Partition is under Replicated
# TYPE kafka_topic_partition_under_replicated_partition gauge
kafka_topic_partition_under_replicated_partition{partition="0",topic="__consumer_offsets"} 0

2.3 Consumer Groups

Name	Exposed informations
`kafka_consumergroup_current_offset`	Current Offset of a ConsumerGroup at Topic/Partition
`kafka_consumergroup_lag`	Current Approximate Lag of a ConsumerGroup at Topic/Partition (broker consumer)
`kafka_consumergroupzookeeper_lag_zookeeper`	Current Approximate Lag of a ConsumerGroup at Topic/Partition (zk consumer)

# HELP kafka_consumergroup_current_offset Current Offset of a ConsumerGroup at Topic/Partition
# TYPE kafka_consumergroup_current_offset gauge
kafka_consumergroup_current_offset{consumergroup="KMOffsetCache-kafka-manager-3806276532-ml44w",partition="0",topic="__consumer_offsets"} -1

# HELP kafka_consumergroup_lag Current Approximate Lag of a ConsumerGroup at Topic/Partition
# TYPE kafka_consumergroup_lag gauge
kafka_consumergroup_lag{consumergroup="KMOffsetCache-kafka-manager-3806276532-ml44w",partition="0",topic="__consumer_offsets"} 1

3、grafana报表

目前grafana主要是基于这2个模板的dashboard按需构建。链接:

三、问题与改进

1、ZK连接串支持chroot

我们现网kafka的zk都是带有chroot，如host1:2181,host2:2181/kafka1，试用发现kafka exporter 并不支持这种用法。分析代码发现，kafka exporter使用zk库kazoo的姿势不太对，使用NewKazooFromConnectionString代替NewKazoo方法就能兼容我们的场景，目前这种改进方案已经提交pr给作者。

2、多kafka集群采集支持

2.1 一个程序实例采集多个集群？

一个kafka exporter实例目前只能采集一个集群，自然而然我们就想通过改造源程序支持采集多个kafka集群的指标。然而试用中发现我们的kafka集群topic\partition\group数量极多，几乎达到LinkedIn内部最繁忙集群的水平，所以prometheus pull单集群指标的时间会比较长(30s左右，开启zk lag采集会更长)且指标内容较大(3w行左右)，如果多个集群的指标汇总在一个实例里，prometheus压力会很大，所以放弃了这种方案。

2.2 给`kafka exporter`实例打标签

经过取舍后，我们决定采用一个集群一个kafka exporter方案，利用kafka.labels的特性，加上clusterId信息来区分不同集群的指标。参考方法：

./kafka_exporter --kafka.server=broker-host:9092  --kafka.labels="clusterId=203" --use.consumelag.zookeeper --zookeeper.server="zk-host:2181/kafka1" --kafka.version ="2.2.0"

3、部署改进

打包成一个docker镜像部署在k8s上。

感兴趣的同学可以关注我的微信公众号~

原创声明：本文系作者授权腾讯云开发者社区发表，未经许可，不得转载。

如有侵权，请联系 cloudcommunity@tencent.com 删除。

消息队列 CKafka 版

原创声明：本文系作者授权腾讯云开发者社区发表，未经许可，不得转载。

如有侵权，请联系 cloudcommunity@tencent.com 删除。

消息队列 CKafka 版

登录后参与评论

0 条评论

热度