Introducing Kafka Backup

Anatoly Zelenin
Jun 14 · 9 min read

Backing up Kafka is not trivial. Here, we introduce Kafka Backup, currently the only available viable solution to backup your Kafka Clusters

I assume you are familiar with Apache Kafka. If not, stop reading here and grab a book or two or google for a few introductory peaces. Otherwise let us skip the introduction and get straight to the topic: Why there is a need to back up Kafka and how to do so.

The Need for a Backup Solution for Kafka

Kafka is a highly distributed system and can be configured to provide a high level of resilience on its own: Using a large replication factor we can survive the loss of several brokers and when we stretch our Kafka Cluster across multiple data centers (the latency should stay below 30ms) we can even survive the loss of data centers! So why should we care about backups if we can just increase the distribution of Kafka?

Replication does not Replace Backups

Replication handles many error cases but by far not all. What about the case that there is a bug in Kafka that deletes old data? What about a misconfiguration of the topic (are you sure, that your value of is a millisecond value?)? What about an admin that accidentally deleted the whole Prod Cluster because they thought they were on dev? What about security breaches? If an attacker gets access to your Kafka Management interface, they can do whatever they like.

Of course, this does not matter too much if you are using Kafka to distribute click-streams data for your analytics department and it is tolerable to loose some data. But if you use Kafka as your “central nervous system” for your company and you store your core business data in Kafka you better think about a cold storage backup for your Kafka Cluster.

Do you Really Need Additional Kafka Clusters?

If you do not use Kafka for “Big Data” applications you are probably totally fine with just one smallish Kafka Cluster. Maybe your applications run only in one data center and this is totally ok for you. Then there is probably no need to set up and operate additional Kafka Clusters. If your data center with all your applications shuts down what use is it that your stretched high available Kafka Cluster survived? You would be probably absolutely fine if you have just a backup of all your data on an (offline) storage that you could replay to restore operations.

What is Kafka Backup?

Kafka Backup is a tool to back up and restore your Kafka data including all (configurable) topic data and especially also consumer group offsets. To the best of our knowledge, Kafka Backup is the only viable solution to take a cold backup of your Kafka data and restore it correctly.

Kafka Backup consists of two Connectors for Kafka Connect: One for Backup (Implemented as a Sink Connector) and one for the restore of data (Implemented as a Source Connector). There is nothing special about the Connectors — instead of using a third party system like a database or external service to write data to or read data from, the Connectors write or read data to/from the local file system.

The Sink Connector connects to the Kafka Broker and continuously fetches data from the specified topics (using a regex or a list of topics) and writes it to disk. Similar as Kafka, Kafka Backup does not interpret data in any way but writes it without modification as bytes to files. To simplify the handling of the data, Kafka Backup splits the data into segments (similar to Kafka) and creates indices to access that data faster. You can mount directly your storage solution to a directory and let Kafka Backup write to that directory or you use a cron job to move that data to a storage system of your choice. As Kafka Backup is an append-only system, incremental backups are trivial. Similarly, you can simply delete old data by deleting old segments if they reached their time to live (In that case, simply recreate the indices before restoring to minimize confusion).

The Source Connector connects to the Kafka Broker and imports the data stored in the files to Kafka. To reduce errors, you need to explicitly define the list of the topics to restore. And again, Kafka Connect handles many edge cases for us and it is not a problem if the Connector is restarted (or crashes) during restoration. It continues where it paused. Additionally to the actual data in the topic, the source Connector also restores the offsets of the consumers. It is not enough to just copy the __consumer_offsets topic because the actual offsets of the messages may have changed. Kafka Backup handles this. As there is no way to gracefully stop a Connector when its finished, the restore Connector will just Log every few seconds a message that it finished and will not continue to push data to Kafka.


Setting up Kafka Backup is very similar to setting up any other Kafka Connect Connector. We describe this procedure in more detail in the documentation.

First, you need to add the Kafka Backup jar to the Kafka Connect Connector directory. This enables Kafka Connect to discover the Connector. There is only one jar file for backup and restore.

The Kafka Backup Sink requires no special attention. Just define the topics to backup and where to store it. To be able to sync the consumer group offsets the Connector requires the connection details to the Kafka Cluster.

The restore Connector is a bit tricky: The Kafka Backup Source Connector requires an API which is currently (June 2019) under review and which is also required by Mirror Maker 2. At the time of writing it is planned for Kafka 2.4 (Current version is 2.2). Using the old Kafka Connect results in the consumer offsets not being synced. This may be acceptable in some cases but otherwise you need to compile Kafka Connect 2.4 yourself or use our provided Docker image. You find the details in the documentation.

Apart from that, the setup of the Kafka Backup Source Connector is very similar to the Sink Connector: Define the topics to backup explicitly and configure the source directory and cluster credentials.


As written above, at the time of writing, we believe that Kafka Backup is the only viable solution if you really want to take a cold backup of your Kafka data. Nevertheless there are several different approaches that other people use to backup Kafka data and recover from disasters.

File System Snapshots

This was the easiest and most reliable way to backup data and consumer offsets from Kafka. The procedure basically shuts down one broker after another and performs a file system snapshot which is stored on another (cold) disk.

Backup Procedure:

Repeat for each Kafka broker:

  1. Shut down the broker
  2. Take a snapshot of the Filesystem (optional)
  3. Copy the snapshot (or simply the files) to the backup storage
  4. Turn on the broker and wait until all partitions are in sync

Restore Procedure:

  1. Restore the snapshot for each broker
  2. Boot the brokers


  • Uses native OS tools
  • As this procedure needs to be done very often, the fear of shutting down a broker is minimized (especially for a team and environment with few Kafka expertise)
  • Offsets are backed up and restored correctly
  • Internal topics are backed up and restored correctly
  • Compacted messages are deleted too
  • Messages older than the retention time are deleted too
  • Uses cold storage


  • Each message is backed up at least replication_factor-times. Even if it enough to store it without replication.
  • Reduced availability as every broker needs to be turned of for a backup
  • Incremental Backups are harder to achieve (e.g. due to partition rebalancing)
  • POTENTIAL DATA LOSS: If the backup is performed during a partition rebalance (very likely when the backup takes a loooong time) the backup could miss a whole partition due to bad timing.

Using Mirror Maker 2 to backup data to another Cluster

The traditional Mirror Maker has many issues as discussed in KIP-382. Mirror Maker 2 approaches many of them and can be used to back up data from one cluster to another.

Mirror Maker 2 is also (as Kafka Backup) based on Kafka Connect and copies consumer offsets too.

Backup Procedure A+B (normal setup):

  • Set up the MM2 Connector that copies the data from the topic [topic] on the source cluster to the topic [source-cluster-name].[topic] on the cluster name.
  • Mirror Maker 2 ensures that the messages are copied continuously. Offsets are also copied to a separate topic

Backup Procedure C (for consistent Snapshots):

  • Set up the sink (backup) cluster with one broker
  • Set up the topics on the sink cluster with a replication factor of 1
  • Set up Mirror Maker 2 to copy data from the source cluster to the sink cluster
  • Use a cron job to shut down the sink cluster (with one broker) regularly and take a snapshot of the file system and store them on cold storage.

Restore Procedure A (Use other cluster):

  • Use the offset sync topic to configure the consumer groups to consume from the correct offset.
  • Setup the consumers to use the other cluster. Throw away the old one.
  • Set up the clients to produce and consume from the new topics in the new cluster. Edit: To update the new consumer offsets you can either update them using the data found in the Mirror Maker 2 offset topic or you can use the new RemoteClusterUtils.translateOffsets() method in KIP-382. It gives you the right topics and offsets for a given consumer group — you just need a script to apply the new offsets across the board. (Thanks to Ryanne Dolan for the hint)
  • Set up a new Backup Cluster

Restore Procedure B (Mirror data back):

  • Create a new Kafka Cluster
  • Set up Mirror Maker 2 to copy the data to the new cluster
  • Continue with procedure A

Restore Procedure C (Mirror + Snapshot):

  • Use Procedure B or restore a new cluster from the file system snapshots
  • Add more nodes accordingly
  • Increase the replication factor to match the requirements
  • Rebalance the partitions if needed
  • Continue with procedure A


  • Support for warm cluster fail-over (active-active, active-passive)
  • Support for more advanced cluster topologies


  • Requires a second Kafka Cluster
  • Apart from C this is a warm backup and does not protect from bugs in Kafka or the underlying OS
  • Requires custom implementation of the switch-over handling to the restored cluster
  • Adds a lot of complexity in the setup


kafka-connect-s3 is a popular Kafka Connect connector to mirror the data from topics to Amazon S3 (or compatible other services like Minio). Zalando describes a setup in their article Surviving Data Loss

Backup procedure:

  • Set up the sink connector to use your S3 endpoint
  • Set up another sink connector that backs up the __consumer_offsets topic.

Restore procedure:

  • Set up the source connector to read the data from S3 into Kafka
  • Manually extract the new offset for the consumers and manually identify which offset on the new Kafka cluster matches the old one. (This is not a trivial task — you would need to count the ACK’d messages from the beginning to find out the exact offset — and not forgetting about compacted and deleted messages)


  • Cold backup (to S3)
  • Possible to use in downstream services that work only with S3 (e.g. Data Warehouses)


  • Supports only S3 (and compatible systems) as the storage backend
  • No support for restoring consumer offsets (the method described above could be described as guesstimating and will not work in many edge cases)

Kafka Backup

Kafka Backup is inspired heavily by the Mirror Maker 2 and kafka-connect-s3. It consists of a sink and a source connector both of which support the backup and restore of the topic data and also consumer offsets.

Backup Procedure:

  • Set up the Kafka Backup Sink connector
  • Copy the backed up data to a backup storage of your choice
  • See GitHub for more information of how to back up your Kafka Cluster

Restore Procedure

  • Set up the Kafka Backup Source connector
  • Wait until it finished (see logs for information)
  • Use the restored cluster


  • Only solution which is able to restore topic data and also consumer offsets
  • Only solution designed to take cold backups of Kafka
  • Simple to do incremental backups


  • See GitHub for the current maturity status of the project
  • Currently supports only the file system as the storage backend
  • Requires Kafka Connect binaries of Kafka 2.4

Future Work

I think that a Backup strategy should be an essential part of any data store. I aim for getting more feedback from the Kafka community and to let Kafka Backup become a standardized piece of the Kafka Ecosystem.

You can find Kafka Backup on GitHub.

Need Help With Your Kafka Architecture?

As a shameless plug for the end: Drop me a line at if you want to talk about your Kafka Architecture

Anatoly Zelenin

Written by

Computer Scientist who brings together Software Engineering and Theoretical CS to solve real world problems.

Welcome to a place where words matter. On Medium, smart voices and original ideas take center stage - with no ads in sight. Watch
Follow all the topics you care about, and we’ll deliver the best stories for you to your homepage and inbox. Explore
Get unlimited access to the best stories on Medium — and support writers while you’re at it. Just $5/month. Upgrade