Cluster migration using MirrorMaker2.0
This How-To guide covers executing a cluster migration to a Charmed Apache Kafka K8s deployment using MirrorMaker2.0.
The MirrorMaker runs on the new (destination) cluster as a process on each Juju unit in an active/passive setup. It acts as a consumer from an existing cluster and a producer to the Charmed Apache Kafka cluster. Data and consumer offsets for all existing topics will be synced one-way in parallel (one process on each unit) until both clusters are in-sync, with all data replicated across both in real-time.
MirrorMaker2 overview
Under the hood, MirrorMaker uses Kafka Connect source connectors to replicate data, those being the following:
- MirrorSourceConnector - replicates topics from an original cluster to a new cluster. It also replicates ACLs and is necessary for the MirrorCheckpointConnector to run
- MirrorCheckpointConnector - periodically tracks offsets. If enabled, it also synchronizes consumer group offsets between the original and new clusters
- MirrorHeartbeatConnector - periodically checks connectivity between the original and new clusters
Together, they are used for cluster->cluster replication of topics, consumer groups, topic configuration and ACLs, preserving partitioning and consumer offsets. For more detail on MirrorMaker internals, consult the MirrorMaker README.md and the MirrorMaker 2.0 KIP. In practice, it allows one to sync data one way between two live Apache Kafka clusters with minimal impact on the ongoing production service.
In short, MirrorMaker runs as a distributed service on the new cluster that may not yet be serving traffic to external clients. MirrorMaker consumes all topics, groups and offsets from the still-active original cluster in production to produce them one way on the new one.
The original, in-production cluster is referred to as an ‘active’ cluster, and the new cluster still waiting to serve external clients is ‘passive’. The MirrorMaker service can be configured using a configuration similar to the one available for Kafka Connect.
Pre-requisites
To migrate a cluster we need:
- An “old” existing Apache Apache Kafka cluster to migrate from.
- The cluster needs to be reachable from/to the new Apache Kafka cluster.
- A bootstrapped Juju K8s cloud running Charmed Apache Kafka K8s to migrate to. For guidance on how to deploy a new Charmed Apache Kafka K8s, see:
- The How to deploy guide for Charmed Apache Kafka K8s
- The Charmed Apache Kafka K8s Tutorial
- The CLI tool yq, that can be installed via snap:
snap install yq --channel=v3/stable
Get cluster details and admin credentials
By design, the kafka charm will not expose any available connections until related by a client. In this case, we deploy data-integrator charms and relate them to each kafka application, requesting admin level privileges:
juju deploy data-integrator --channel=edge -n 1 --config extra-user-roles="admin" --config topic-name="default"
juju relate kafka-k8s data-integrator
When the data-integrator charm relates to a kafka-k8s application on the kafka-client relation interface, passing extra-user-roles=admin, a new user with super.user permissions will be created on that cluster, with the charm passing back the credentials and broker addresses in the relation data to the data-integrator.
As we will need full access to both clusters, we must grab these newly generated authorisation credentials from the data-integrator:
# SASL credentials to connect to the Charmed Apache Kafka cluster
export NEW_USERNAME=$(juju show-unit data-integrator/0 | yq -r '.. | .username? // empty')
export NEW_PASSWORD=$(juju show-unit data-integrator/0 | yq -r '.. | .password? // empty')
# list of bootstrap-server IPs
export NEW_SERVERS=$(juju show-unit data-integrator/0 | yq -r '.. | .endpoints? // empty')
# building full sasl.jaas.config for authorisation
export NEW_SASL_JAAS_CONFIG="org.apache.kafka.common.security.scram.ScramLoginModule required username=\""${NEW_USERNAME}"\" password=\""${NEW_PASSWORD}\"\;
Required source cluster credentials
To authenticate MirrorMaker to both clusters, it will need full super.user permissions on BOTH clusters. MirrorMaker supports every possible security.protocol supported by Apache Kafka. In this guide, we will make the assumption that the original cluster is using SASL_PLAINTEXT authentication, as such, the required information is as follows:
# comma-separated list of Kafka server IPs and ports to connect to
OLD_SERVERS
# string of sasl.jaas.config property
OLD_SASL_JAAS_CONFIG
If using SSL or SASL_SSL authentication, review the configuration options supported by Kafka Connect in the Apache Kafka documentation
Generating mm2.properties file on the Apache Kafka cluster
MirrorMaker takes a .properties file for its configuration to fine-tune behaviour. See below an example mm2.properties file that can be placed on each of the Apache Kafka units using the above credentials:
# Aliases for each cluster, can be set to any unique alias
clusters = old,new
# Specifies that data from 'old' should be consumed and produced to 'new', and NOT visa-versa, i.e 'active/passive' setup
old->new.enabled = true
new->old.enabled = false
# comma-separated list of kafka server IPs and ports to connect from both clusters
old.bootstrap.servers=$OLD_SERVERS
new.bootstrap.servers=$NEW_SERVERS
# sasl authentication config for each cluster, in this case using the 'admin' users created by the integrator charm for Charmed Apache Kafka
old.sasl.jaas.config=$OLD_SASL_JAAS_CONFIG
new.sasl.jaas.config=$NEW_SASL_JAAS_CONFIG
# if not deployed with TLS, Charmed Apache Kafka uses SCRAM-SHA-512 for SASL auth, with a SASL_PLAINTEXT listener
sasl.mechanism=SCRAM-SHA-512
security.protocol=SASL_PLAINTEXT
# keeps topic names consistent across clusters - see https://kafka.apache.org/30/javadoc/org/apache/kafka/connect/mirror/IdentityReplicationPolicy.html
replication.policy.class=org.apache.kafka.connect.mirror.IdentityReplicationPolicy
# pattern match for replicating all topics and all consumer groups
topics=.*
groups=.*
# the expected number of concurrent MirrorMaker tasks, usually set to match number of physical cores on the target cluster
tasks.max=3
# the new replication.factor for topics produced to the target cluster
replication.factor=2
# allows new topics and groups created mid-migration, to be copied
refresh.topics.enabled=true
sync.group.offsets.enabled=true
sync.topic.configs.enabled=true
refresh.topics.interval.seconds=5
refresh.groups.interval.seconds=5
sync.group.offsets.interval.seconds=5
emit.checkpoints.interval.seconds=5
# filters out records from aborted transactions
old.consumer.isolation.level=read_committed
new.consumer.isolation.level=read_committed
# Specific Connector configuration for ensuring Exactly-Once-Delivery (EOD)
# NOTE - EOD support guarantees released with Apache Kafka 3.5.0 so some of these options may not work as expected
old.producer.enable.idempotence=true
new.producer.enable.idempotence=true
old.producer.acks=all
new.producer.acks=all
# old.exactly.once.support = enabled
# new.exactly.once.support = enabled
Once these properties have been generated (in this example, saved to /tmp/mm2.properties), it is needed to place them on every Apache Kafka unit:
cat /tmp/mm2.properties | juju ssh kafka-k8s/<id> sudo -i 'sudo tee -a /etc/kafka/mm2.properties'
where <id> is the id of the Charmed Apache Kafka unit.
Starting a dedicated MirrorMaker cluster
It is strongly advised to run MirrorMaker services on the downstream cluster to avoid service impact due to resource use. Now that the properties are set on each unit of the new cluster, the MirrorMaker services can be started with JMX metrics exporters using the following:
# building KAFKA_OPTS env-var for running with an exporter
export KAFKA_OPTS = "-Djava.security.auth.login.config=/etc/kafka/zookeeper-jaas.cfg -javaagent:/opt/kafka/libs/jmx_prometheus_javaagent.jar=9099:/etc/kafka/jmx_kafka_connect.yaml"
# To start MM on kafka-k8s/<id> unit
juju ssh kafka-k8s/<id> sudo -i 'cd /opt/kafka/bin && KAFKA_OPTS=$KAFKA_OPTS ./connect-mirror-maker.sh /etc/kafka/mm2.properties'
Monitoring and validating data replication
The migration process can be monitored using built-in Apache Kafka bin commands on the original cluster. In the Charmed Apache Kafka, these bin commands are also mapped to snap commands on the units (e.g charmed-kafka.get-offsets or charmed-kafka.topics).
To monitor the current consumer offsets, run the following on the original cluster being migrated from:
watch "bin/kafka-consumer-groups.sh --describe --offsets --bootstrap-server $OLD_SERVERS --all-groups
An example output of which may look similar to this:
GROUP TOPIC PARTITION CURRENT-OFFSET LOG-END-OFFSET LAG CONSUMER-ID HOST CLIENT-ID
admin-group-1 NEW-TOPIC 0 95 95 0 kafka-python-2.0.2-a95b3f90-75e9-4a16-b63e-5e021b7344c5 /10.248.204.1 kafka-python-2.0.2
admin-group-1 NEW-TOPIC 3 98 98 0 kafka-python-2.0.2-a95b3f90-75e9-4a16-b63e-5e021b7344c5 /10.248.204.1 kafka-python-2.0.2
admin-group-1 NEW-TOPIC 1 82 82 0 kafka-python-2.0.2-a95b3f90-75e9-4a16-b63e-5e021b7344c5 /10.248.204.1 kafka-python-2.0.2
admin-group-1 NEW-TOPIC 2 89 90 1 kafka-python-2.0.2-a95b3f90-75e9-4a16-b63e-5e021b7344c5 /10.248.204.1 kafka-python-2.0.2
admin-group-1 NEW-TOPIC 4 103 104 1 kafka-python-2.0.2-a95b3f90-75e9-4a16-b63e-5e021b7344c5 /10.248.204.1 kafka-python-2.0.2
There is also a range of different metrics made available by MirrorMaker during the migration. These can be accessed with something similar to:
curl 10.248.204.198:9099/metrics | grep records_count
Switching client traffic
Once happy with data migration, stop all active consumer applications on the original cluster, and redirect them to the Charmed Apache Kafka cluster, making sure to use the Charmed Apache Kafka cluster server addresses and authentication. After doing so, they will re-join their original consumer groups at the last committed offset it had originally, and continue consuming as normal.
Finally, the producer client applications can be stopped, updated with the Charmed Apache Kafka cluster server addresses and authentication, and restarted, with any newly produced messages being received by the migrated consumer client applications, completing the migration of both the data, and the client applications.
Stopping MirrorMaker replication
Once confident in the successful completion of the data an client migration, the running processes on each of the charm units can be killed, stopping the MirrorMaker processes active on the Charmed Apache Kafka cluster.