AMQ Streams simplifies the process of running Apache Kafka in an OpenShift cluster.
Note:
All references of source cluster, target cluster implicitly mean AMQ Streams v1.4, AMQ Streams v1.7 respectively; unless when otherwise explicitly pointed out. I have made an attempt to define certain commonly utilized terms in the Glossary section.
Furthermore, the post assumes the reader has some basic understanding of Apache kafka.
There is a Kafka CLI command cheat sheet added towards the end of the article; all credits go to the team who put it together; Chapeau a vous.
This guide attempts to discuss lessons learned from a migration project for a large freight company. We helped migrate an AMQ Streams cluster from OpenShift 3.11 to 4.7; additionally, we also migrated over 100s Java event-driven micro-services to the new cluster(AMQ Streams V1.7).
In the section titled Technical Implementation, I will show how to migrate an AMQ Stream cluster between OpenShift environments(Example: OCP3.11 to OCP4.8) using MirrorMaker2; a Kafka component used to mirror data between two or more active Kafka clusters, within or across Kubernetes clusters.
We will begin with an overview of the two clusters(kafka-source and kafka-target), the components deployed, then lay down steps to deploy AMQ Stream v1.8, Prometheus(metrics collection), and Grafana(Dashboard, and Alerts). All these components are set up via a fleet of helm charts with the possibility of selectively installing specific resources(amqs, mirrormaker, grafana).
In condensed terms, there was a need to migrate a pre-existing AMQ Streams(v1.4) cluster from OpenShift(v3.11) to AMQ Streams(v1.7) on OpenShift(v4.7).
These are the primary goals:
-
minimal downtime
-
minimal to no data loss
-
zero message duplication
-
new cluster(AMQ Streams V1.7) should be capable to handle more throughput than the former
-
monitor and alert on some Kafka cluster and applications states and behaviors.
To successfully carry out this migration, we adopted the Think-Do-Check mindset; in other words, we followed the concept of Dry-Run to build a repeatable and predictable migration strategy.
This strategy involved the following points:
-
Make sure source and target clusters are monitored, and all key performance indicators(KPI) needed to track mirroring progress are displayed on easy to access dashboards,
-
Some of these metrics are:
-
Consumer Group CURRENT OFFSET,
-
Consumer Group LOG-END OFFSET,
-
Consumer Group Lag,
-
Bytes In & Out Per Seconds,
-
Incoming & Outgoing Messages Per Seconds
-
Total Messages per Topic/Partition
-
Total Bytes per Partition
-
Balancedness Score
-
Max Bytes Size vs. Used Bytes of PersistenceVolume objects backing up the AMQ Streams cluster
-
-
Prometheus and Grafana were used in this case,
-
Strimzi Kafka Exporter dashboard was to gauge Consumer Group Lag, Consumer Group Offsets, Incoming Messages/Seconds, Outgoing Messages/Seconds,
-
Messages Count/Topic to make rough estimates about topic/partition size.
-
-
Validate target cluster data retention(TTL) configurations,
-
Validate data retention periods match for source and target clusters,
-
spec.kafka.config.log.retention.hoursfor time based TTL, -
spec.kafka.config.log.retention.bytesfor byte size based retention, -
having these two configurations match ensure messages have the same TTL across the two clusters.
-
-
-
Validate all
KafkaTopicresources have been created on target cluster with parallel configs as those on source cluster . -
Confirm all required
KafkaUserresources with their respective RBAC privileges are created,-
for this case, we had one
KafkaUserresource per application.
-
-
Deploy all applications on the target cluster and validate they spin up and down with no prohibitive errors,
-
once all required producers and consumers are healthy, shut them down until ready for migration,
-
above action created all required consumer groups;
-
it set created Consumer Group CURRENT_OFFSET to zero which means the next time a producer or consumer spins up, it will start reading or writing at that CURRENT OFFSET of zero rather than the latest offset;
-
hence preventing apps from skipping messages when they are launched during Cut Over.
-
-
-
Deploy MirrorMaker2 on target cluster and begin the replication,
-
depending on data size, data ingestion rate, replication speed, it might take from minutes, days, to weeks before MirrorMaker2 is caught up with the source cluster .
-
use the monitoring dashboards(source and target), kcat, and kafka cli programs for validation.
-
-
The plan describes steps involved in switching from the source cluster over to the target cluster.
MirrorMaker2 must be running ahead of cut over time, this can be hours, days, weeks. Utilize data size, data ingestion rate and replication speed to estimate for how long you need MirrorMaker2 running.
A side note, this procedure is not set in stone; you may need to tweak the execution plan to match your requirements and scenarios.
-
Turn off data ingestion valve to source cluster and route it to target cluster,
-
this may be
KafkaConnect,Debezium for Change Data Capture, or some other system that pumps data into the AMQ Streams cluster.
-
-
Spin down producers on source cluster and wait for consumers to drain their respective topics,
-
a consumer topic is called drain when CURRENT OFFSET(read position) and LOG-END OFFSET(write position) matches for the associated Consumer Group;
-
when Consumer Group Lag is zero, a topic can also be designated drained.
-
in some cases applications behave as both producers and consumers, you need to be very diligent in this case;
-
I would also recommend that you partition your applications into tiers of Upstream, Dam, and Downstream based on data flow and dependency chain.
-
Another approach could be breaking the applications into buckets(you decide the criteria), and migrate one batch at a time.
-
-
Once all topics are drained, spin down consumer apps on source cluster,
-
this ensures all messages on source cluster have been processed.
-
-
On target cluster, spin down Dam applications if any
-
doing this creates a choke points in the data pipeline; it gives you the ability to dial up/down data flow to Downstream applications.
-
it also provide us the ability to sample and verify data coming in before it reaches Downstream.
-
-
Turn on producers on the target cluster,
-
this may be applications on the left side of the Dam.
-
-
On target cluster, spin up Dam applications if any
-
being the choke point of the data pipeline, spinning up these apps will open up the floodgate.
-
-
Spin up Downstream applications on target cluster.
-
Stop MirrorMaker2 instance
-
this action is performed after confirming target cluster has caught up with source cluster .
-
use monitoring, kcat, kafka cli tools to help with this.
-
hint: using kcat program, grab last n(100, 5000,…) messages from each cluster, encode each message to base64, write the output to files, execute the diff command on the files,
-
in our case we could not compare offsets because for source and target cluster, Consumer Group Offsets were always different regardless of mirroring progress,
-
I hope this gets resolved in a near future.
-
-
-
-
-
Validate Cut Over process is successful,
-
look at data ingestion volume; should be nearly or exactly the same as the it was on the source cluster;
-
if your apps are writing to data stores, validate there is similar read/write rate or traffic patterns.
-
validate apps logs and that there are no errors or abnormal messages appearing in the logs.
-
watch for resource consumption spikes and network traffic patterns in the cluster.
-
and last but not least, verify applications are still able to handle your business transactions.
-
This plan might be per use case basis, but the distilled form is as follow:
-
Switch cluster data ingestion system to source cluster.
-
On source cluster:
-
spin down producers;
-
wait for consumer apps to drain their topics and shut them down.
-
-
On target cluster:
-
turn on producer/upstream apps,
-
turn on consumers.
-
-
Validate data ingestion traffic patterns match that of source cluster ,
-
you may use the monitoring dashboards for this; for instance you may look at:
-
incoming and outgoing messages per second
-
data ingestion rate in the data stores if any
-
application health
-
Consumer Group CURRENT OFFSET, LOG-END OFFSET, LAG
-
bytes in/out per second, and much more.
-
-
In simple terms when mirroring an existing cluster:
-
set
mirrors.sourceConnector.config.auto.offset.reset: latest-
If using this approach, you could just point your producers to the source cluster; however in our case the client wanted to have an new installation of AMQ Streams(v1.4 to v1.7) on a brand new OpenShift Container Platform(v3.11 to v4.7)
-
when replicating historical data is not of concerns
-
an example use case might be when MM2 is set up at the same time as the AMQ Streams cluster
-
or for some requirement, you just want to begin mirroring at the end of topic; no concerns for data loss.
-
-
-
set
mirrors[].sourceConnector.config.auto.offset.reset: earliest-
when replicating historical data is a requirement
-
when you want all available historical data in the source cluster copied over to the target cluster,
-
an example use case might be when migrating an existing AMQ Streams cluster to another with minimal or zero data loss.
-
or you want an active/passive setup whereby the passive cluster is installed at a moment when the active cluster already hold data.
-
Pods were being terminated in the target cluster, at first we thought it was the JVM eating up all the pod memory. We implemented JVM boundaries with -Xms and Xmx values as arguments for the ENTRYPOINT command of the container image; java -jar in this case. Despite this change we still got OOMKilled errors, pods were being terminated.
After few days of troubleshooting and head scratches, we found the root cause. The issue was due to some of the apps creating more processes in their containers than the PID limit set at 1024 by default; hence OOMKilled error after the pod runs for few hours despite having an HorizontalPodAutoscaler resource monitoring this particular application for autoscaling scaling needs.
In this guide, we have explained the strategies and plans involved in the migration of an AMQ Streams cluster in an Active/Passive setting between two OpenShift installations. Have a look at the Technical Implementations section for a walk through demo.
Happy Streaming!
-
Apache Kafka: Software product for building large scale messaging networks.
-
AMQ Streams: RedHat product for simplifying Apache Kafka deployment in an OpenShift Cluster.
-
Source Cluster: AMQ Streams v1.4
-
Target Cluster: AMQ Streams v1.7
-
Producer: Any application that publishes to AMQ Streams.
-
Consumer: Any application that consumes messages or subscribes to topics.
-
Consumer Group: a collection of consumers who collaborate to consume data from topics.
-
Consumer Group Current Offset: Read position of the consumer subscribed to a topic via a Consumer Group.
-
Consumer Group Log End Offset: Write position of the producer subscribed to a topic via a Consumer Group.
-
Consumer Group Lag: indicates the lag(diff between LOD-END OFFSET and CURRENT-OFFSET) between a producers and consumers subscribing to the same topic via a consumer group.
-
Drained Topic: a topic is drained when the associated consumer group lag is zero.
-
Topic: it has unique name across the Kafka cluster; it help with message categorization.
-
Partition: breaks topics into multiple logs each of which can live on a different node; messages are written at the partition level; moreover, partitions provide means of concurrency for the Kafka cluster.