Kafka Health Check

Health checker for Kafka brokers and clusters that operates by checking whether:

• a message inserted in a dedicated health check topic becomes available for consumers,
• the broker can stay in the ISR of a replication check topic,
• the broker is in the in-sync replica set for all partitions it replicates,
• under-replicated partitions exist,
• out-of-sync replicas exist,
• offline partitions exist, and
• the metadata of the cluster and the ZooKeeper metadata are consistent with each other.

Status

Release version is 0.1.0

Compiled binaries are available for Linux, macOS, and FreeBSD.

Use Cases

Submit a pull request to have your use case listed here!

Self-healing cluster

At AutoScout24, in order to reduce operational workload, we use kafka-health-check to automatically restart broker nodes as they become unhealthy.

In-place rolling updates

At AutoScout24, to keep the OS up to date of our clusters running on AWS, we perform regular in-place rolling updates. As we run immutable servers, we terminate each broker and replace them with fresh EC2 instances (keeping the previous broker ids). In order not to jeopardy the cluster stability when terminating brokers, we verify that the cluster is healthy before taking one broker offline. Similarly, we wait for the broker coming back online to fully catch up before proceeding with the next broker. To achieve this, we use the cluster health information provided by kafka-health-check.

Usage

Usage of kafka-health-check:
  -broker-host string
        ip address or hostname of broker host (default "localhost")
  -broker-id uint
        id of the Kafka broker to health check
  -broker-port uint
        Kafka broker port (default 9092)
  -check-interval duration
        how frequently to perform health checks (default 10s)
  -no-topic-creation
        disable automatic topic creation and deletion
  -replication-failures-count uint
        number of replication failures before broker is reported unhealthy (default 5)
  -replication-topic string
        name of the topic to use for replication checks - use one per cluster, defaults to broker-replication-check
  -server-port uint
        port to open for http health status queries (default 8000)
  -topic string
        name of the topic to use - use one per broker, defaults to broker--health-check
  -zookeeper string
        ZooKeeper connect string (e.g. node1:2181,node2:2181,.../chroot)

Broker Health

Broker health can be queried at

/

$ curl -s :8000/
{
    "broker": 1,
    "status": "sync"
}

Return codes and status values are: *

200

sync

200

imok

500

nook

replication-failures-count

The returned json contains details about replicas the broker is lagging behind:

$ curl -s :8000/
{
    "broker": 3,
    "status": "imok",
    "out-of-sync": [
        {
            "topic": "mytopic",
            "partition": 0
        }
    ],
    "replication-failures": 1
}

Cluster Health

Cluster health can be queried at

/cluster

$ curl -s :8000/cluster
{
    "status": "green"
}

Return codes and status values are: *

200

green

200

yellow

500

red

The returned json contains details about metadata status and partition replication:

$ curl -s :8000/cluster
{
    "status": "yellow",
    "topics": [
        {
            "topic": "mytopic",
            "status": "yellow",
            "partitions": {
                "1": {
                    "status": "yellow",
                    "OSR": [
                        3
                    ]
                },
                "2": {
                    "status": "yellow",
                    "OSR": [
                        3
                    ]
                }
            }
        }
    ]
}

The fields for additional info and structures are: *

topics

[{"topic":"mytopic","status":"yellow","partitions":{"2":{"status":"yellow","OSR":[3]}}}]

OSR

metadata

[{"broker":3,"status":"red","problem":"Missing in ZooKeeper"}]

zookeeper

"Fetching brokers failed: ..."

Supported Kafka Versions

Tested with the following Kafka versions:

• 2.0.0
• 1.1.1
• 1.1.0
• 1.0.0
• 0.11.0.2
• 0.11.0.1
• 0.11.0.0
• 0.10.2.1
• 0.10.2.0
• 0.10.1.1
• 0.10.1.0
• 0.10.0.1
• 0.10.0.0
• 0.9.0.1
• 0.9.0.0

Kafka 0.8 is not supported.

see the compatibility spec for the full list of executed compatibility checks. To execute the compatibility checks, run

make compatibility

Building

Run

make

make deps

Prerequisites

• Make to run the Makefile
• Go 1.8 since it's written in Go

Notable Details on Health Check Behavior

• When first started, the check tries to find the Kafka broker to check in the cluster metadata. Then, it tries to find the health check topic, and creates it if missing by communicating directly with ZooKeeper(configuration: 10 seconds message lifetime, one single partition assigned to the broker to check). This behavior can be disabled by using

  -no-topic-creation

  .
• The check also creates one replication check topic for the whole cluster. This topic is expanded to all brokers that are checked.
• When shutting down, the check deletes to health check topic partition by communicating directly with ZooKeeper. It also shrinks the partition assignment of the replication check topic, and deletes it when stopping the last health check process. This behavior can be disabled by using

  -no-topic-creation

  .
• The check will try to create the health check and replication check topics only on its first connection after startup. If the topic disappears later while the check is running, it will not try to re-create its topics.
• If the broker health check fails, the cluster health will be set to

  red

  .
• For each check pass, the Kafka cluster metadata is fetched from ZooKeeper, i.e. the full data on brokers and topic partitions with replicas.