A collection of frequently asked questions on the topic of Kafka aimed for advanced users.
Like most Open Source projects, Kafka provides a lot of configuration options to maximize performance. In some cases, it is not obvious how best to map your specific use case to those configuration options. We attempt to address some of those situations.
What can I do to ensure that I never lose a Kafka event?
This is a simple question which has lots of far-reaching implications for your entire Kafka setup. A complete answer includes the next few related FAQs and their answers.
What is the recommended node hardware for best reliability?
Operationally, you need to make sure your Kafka cluster meets the following hardware setup:
- Have a 3 or 5 node cluster only running Zookeeper (higher only necessary at largest scales).
- Have at least a 3 node cluster only running Kafka.
- Have the disks on the Kafka cluster running in RAID 10. (Required for resiliency against disk failure.)
- Have sufficient memory for both the Kafka and Zookeeper roles in the cluster. (Recommended: 4GB for the broker, the rest of memory automatically used by the kernel as file cache.)
- Have sufficient disk space on the Kafka cluster.
- Have a sufficient number of disks to handle the bandwidth requirements for Kafka and Zookeeper.
- You need a number of nodes greater than or equal to the highest replication factor you expect to use.
What are the network requirements for best reliability?
Kafka expects a reliable, low-latency connection between the brokers and the Zookeeper nodes:
- The number of network hops between the Kafka cluster and the Zookeeper cluster is relatively low.
- Have highly reliable network services (such as DNS).
What are the system software requirements for best reliability?
Assuming you’re following the recommendations of the previous two questions, the actual system outside of Kafka must be configured properly.
- The kernel must be configured for maximum I/O usage that Kafka requires.
- Large page cache
- Maximum file descriptions
- Maximum file memory map limits
- Kafka JVM configuration settings:
- Brokers generally don’t need more than 4GB-8GB of heap space.
- Run with the +G1GC garbage collection using Java 8 or later.
How can I configure Kafka to ensure that events are stored reliably?
The following recommendations for Kafka configuration settings make it extremely difficult for data loss to occur.
- Remember to close the producer when it is finished or when there is a long pause.
Topic replication.factor >= 3
Min.insync.replicas = 2
- Disable unclean leader election
- Commit offsets after messages are processed by your consumer client(s).
If you have more than 3 hosts, you can increase the broker settings appropriately on topics that need more protection against data loss.
Once I’ve followed all the previous recommendations, my cluster should never lose data, right?
Kafka does not ensure that data loss never occurs. There are the following tradeoffs:
- Throughput vs. reliability. For example, the higher the replication factor, the more resilient your setup will be against data loss. However, to make those extra copies takes time and can affect throughput.
- Reliability vs. free disk space. Extra copies due to replication use up disk space that would otherwise be used for storing events.
Beyond the above design tradeoffs, there are also the following issues:
- To ensure events are consumed you need to monitor your Kafka brokers and topics to verify sufficient consumption rates are sustained to meet your ingestion requirements.
- Ensure that replication is enabled on any topic that requires consumption guarantees. This protects against Kafka broker failure and host failure.
- Kafka is designed to store events for a defined duration after which the events are deleted. You can increase the duration that events are retained up to the amount of supporting storage space.
- You will always run out of disk space unless you add more nodes to the cluster.
My Kafka events must be processed in order. How can I accomplish this?
After your topic is configured with partitions, Kafka sends each record (based on key/value pair) to a particular partition based on key. So, any given key, the corresponding records are “in order” within a partition.
For global ordering, you have two options:
- Your topic must consist of one partition (but a higher replication factor could be useful for redundancy and failover). However, this will result in very limited message throughput.
- You configure your topic with a small number of partitions and perform the ordering after the consumer has pulled data. This does not result in guaranteed ordering, but, given a large enough time window, will likely be equivalent.
Conversely, it is best to take Kafka’s partitioning design into consideration when designing your Kafka setup rather than rely on global ordering of events.
How do I size my topic? Alternatively: What is the “right” number of partitions for a topic?
Choosing the proper number of partitions for a topic is the key to achieve a high degree of parallelism with respect to writes and reads and to distribute load. Evenly distributed load over partitions is a key factor to have good throughput (avoid hot spots). Making a good decision requires estimation based on the desired throughput of producers and consumers per partition.
For example, if you want to be able to read 1 GB/sec, but your consumer is only able to process 50 MB/sec, then you need at least 20 partitions and 20 consumers in the consumer group. Similarly, if you want to achieve the same for producers, and 1 producer can only write at 100 MB/sec, you need 10 partitions. In this case, if you have 20 partitions, you can maintain 1 GB/sec for producing and consuming messages. You should adjust the exact number of partitions to number of consumers or producers, so that each consumer and producer achieve their target throughput.
So a simple formula could be:
#Partitions = max(NP, NC)
NPis the number of required producers determined by calculating:
NCis the number of required consumers determined by calculating:
TTis the total expected throughput for our system
TPis the max throughput of a single producer to a single partition
TCis the max throughput of a single consumer from a single partition
This calculation gives you a rough indication of the number of partitions. It's a good place to start. Keep in mind the following considerations for improving the number of partitions after you have your system in place:
- The number of partitions can be specified at topic creation time or later.
- Increasing the number of partitions also affects the number of open file descriptors. So make sure you set file descriptor limit properly.
- Reassigning partitions can be very expensive, and therefore it's better to over- than under-provision.
- Changing the number of partitions that are based on keys is challenging and involves manual copying.
- Reducing the number of partitions is not currently supported. Instead, create a new a topic with a lower number of partitions and copy over existing data.
- Metadata about partitions are stored in ZooKeeper in the form of
znodes. Having a large number of partitions has effects on ZooKeeper and on client resources:
- Unneeded partitions put extra pressure on ZooKeeper (more network requests), and might introduce delay in controller and/or partition leader election if a broker goes down.
- Producer and consumer clients need more memory, because they need to keep track of more partitions and also buffer data for all partitions.
- As guideline for optimal performance, you should not have more than 4000 partitions per broker and not more than 200,000 partitions in a cluster.
Make sure consumers don’t lag behind producers by monitoring consumer lag. To check consumers' position in a consumer group (that is, how far behind the end of the log they are), use the following command:
$ kafka-consumer-groups --bootstrap-server BROKER_ADDRESS --describe --group CONSUMER_GROUP --new-consumer
How can I scale a topic that's already deployed in production?
Recall the following facts about Kafka:
- When you create a topic, you set the number of partitions. The higher the partition count, the better the parallelism and the better the events are spread somewhat evenly through the cluster.
- In most cases, as events go to the Kafka cluster, events with the same key go to the same partition. This is a consequence of using a hash function to determine which key goes to which partition.
Now, you might assume that scaling means increasing the number of partitions in a topic. However, due to the way hashing works, simply increasing the number of partitions means that you will lose the "events with the same key go to the same partition" fact.
Given that, there are two options:
- Your cluster may not be scaling well because the partition loads are not balanced
properly (for example, one broker has four very active partitions, while another has
none). In those cases, you can use the
kafka-reassign-partitionsscript to manually balance partitions.
- Create a new topic with more partitions, pause the producers, copy data over from the old topic, and then move the producers and consumers over to the new topic. This can be a bit tricky operationally.
How do I rebalance my Kafka cluster?
This one comes up when new nodes or disks are added to existing nodes. Partitions are not automatically balanced. If a topic already has a number of nodes equal to the replication factor (typically 3), then adding disks does not help with rebalancing.
kafka-reassign-partitions command after adding new
hosts is the recommended method.
There are several caveats to using this command:
- It is highly recommended that you minimize the volume of replica changes to make sure the cluster remains healthy. Say, instead of moving ten replicas with a single command, move two at a time.
- It is not possible to use this command to make an out-of-sync replica into the leader partition.
- If too many replicas are moved, then there could be serious performance impact on the
cluster. When using the
kafka-reassign-partitionscommand, look at the partition counts and sizes. From there, you can test various partition sizes along with the
--throttleflag to determine what volume of data can be copied without affecting broker performance significantly.
- Given the earlier restrictions, it is best to use this command only when all brokers and topics are healthy.
How do I monitor my Kafka cluster?
Cloudera Manager has monitoring for a Kafka cluster.
Currently, there are three GitHub projects as well that provide additional monitoring functionality:
- Doctor Kafka (Pinterest, Apache 2.0 License)
- Kafka Manager (Yahoo, Apache 2.0 License)
- Cruise Control (LinkedIn, BSD 2-clause License)
These projects are Apache-compatible licensed, but are not Open Source (no community, bug filing, or transparency).
What are the best practices concerning consumer group.id?
group.id is just a string that helps Kafka track which consumers are
related (by having the same group id).
- In general, timestamps as part of
group.idare not useful. Because each
group.idcorresponds to multiple consumers, you cannot have a unique timestamp for each consumer.
- Add any helpful identifiers. This could be related to a group (for example, transactions, marketing), purpose (fraud, alerts), or technology (Flume, Spark).
How do I monitor consumer group lag?
kafka-consumer-groupscommand line tool. Copying directly from the upstream documentation, we have this example output (reformatted for readability):
$ bin/kafka-consumer-groups.sh --bootstrap-server localhost:9092 --describe --group my-group TOPIC PARTITION CURRENT-OFFSET LOG-END-OFFSET LAG CONSUMER-ID HOST CLIENT-ID my-topic 0 2 4 2 consumer-1-69d6 /127.0.0.1 consumer-1 my-topic 1 2 3 1 consumer-1-69d6 /127.0.0.1 consumer-1 my-topic 2 2 3 1 consumer-2-9bb2 /127.0.0.1 consumer-2
In general, if everything is going well with a particular topic, each consumer’s
CURRENT-OFFSET should be up-to-date or nearly up-to-date with the
LOG-END-OFFSET. From this command, you can determine whether a particular
host or a particular partition is having issues keeping up with the data rate.
How do I reset the consumer offset to an arbitrary value?
This is also done using the
kafka-consumer-groups command line
tool. This is generally an administration feature used to get around corrupted records, data
loss, or recovering from failure of the broker or host. Aside from those special cases,
using the command line tool for this purpose is not recommended.
By using the
--execute --reset-offsets flags, you can change the consumer
offsets for a consumer group (or even all groups) to a specific setting based on each
partitions log’s beginning/end or a fixed timestamp. Typing the
kafka-consumer-groups command with no arguments will give you the
complete help output.
How do I configure MirrorMaker for bi-directional replication across DCs?
Mirror Maker is a one way copy of one or more topics from a Source Kafka Cluster to a Destination Kafka Cluster. Given this restriction on Mirror Maker, you need to run two instances, one to copy from A to B and another to copy from B to A.
In addition, consider the following:
- Cloudera recommends using the "pull" model for Mirror Maker, meaning that the Mirror Maker instance that is writing to the destination is running on a host "near" the destination cluster.
- The topics must be unique across the two clusters being copied.
- On secure clusters, the source cluster and destination cluster must be in the same Kerberos realm.
How does the consumer max retries vs timeout work?
- Retries: This is generally related to reading data. When a consumer reads from a
brokers, it’s possible for that attempt to fail due to problems such as intermittent
network outages or I/O issues on the broker. To improve reliability, the consumer
retries (up to the configured
max.retriesvalue) before actually failing to read a log offset.
- Timeout. This term is a bit vague because there are two timeouts related to consumers:
- Poll Timeout: This is the timeout between calls to
KafkaConsumer.poll(). This timeout is set based on whatever read latency requirements your particular use case needs.
- Heartbeat Timeout: The newer consumer has a “heartbeat thread” which give a heartbeat to the broker (actually the Group Coordinator within a broker) to let the broker know that the consumer is still alive. This happens on a regular basis and if the broker doesn’t receive at least one heartbeat within the timeout period, it assumes the consumer is dead and disconnects it.
- Poll Timeout: This is the timeout between calls to
How do I size my Kafka cluster?
There are several considerations for sizing your Kafka cluster.
- Disk space
Disk space will primarily consist of your Kafka data and broker logs. When in debug mode, the broker logs can get quite large (10s to 100s of GB), so reserving a significant amount of space could save you some future headaches.
For Kafka data, you need to perform estimates on message size, number of topics, and redundancy. Also remember that you will be using RAID10 for Kafka’s data, so half your hard drives will go towards redundancy. From there, you can calculate how many drives will be needed.
In general, you will want to have more hosts than the minimum suggested by the number of drives. This leaves room for growth and some scalability headroom.
- Zookeeper nodes
One node is fine for a test cluster. Three is standard for most Kafka clusters. At large scale, five nodes is fairly common for reliability.
- Looking at leader partition count/bandwidth usage
This is likely the metric with the highest variability. Any Kafka broker will be overloaded if it has too many leader partitions. In the worst cases, each leader partition requires high bandwidth, high message rates, or both. For other topics, leader partitions will be a tiny fraction of what a broker can handle (limited by software and hardware). To estimate an average that works on a per-host basis, try grouping topics by partition data throughput requirements, such as 2 high bandwidth data partitions, 4 medium bandwidth data partitions, 20 small bandwidth data partitions. From there, you can determine how many hosts are needed.
How can I combine Kafka with Flume to ingest into HDFS?
We have two blog posts on using Kafka with Flume:
- The original post: Flafka: Apache Flume Meets Apache Kafka for Event Processing
- This updated version for CDH 5.8/Apache Kafka 0.9/Apache Flume 1.7: New in Cloudera Enterprise 5.8: Flafka Improvements for Real-Time Data Ingest
How can I build a Spark streaming application that consumes data from Kafka?
You will need to set up your development environment to use both Spark libraries and Kafka libraries:
- Building Spark Applications
- The kafka-examples directory on Cloudera’s public GitHub has an
From there, you should be able to read data using the KafkaConsumer class and using Spark libraries for real-time data processing. The blog post Reading data securely from Apache Kafka to Apache Spark has a pointer to a GitHub repository that contains a word count example.
For further background, read the blog post Architectural Patterns for Near Real-Time Data Processing with Apache Hadoop.