Best practices
This topic outlines some best practices to follow when using Amazon MSK.
Right-size your cluster: Number of partitions per broker
The following table shows the recommended number of partitions (including leader and follower replicas) per broker.
| Broker type | Recommended number of partitions (including leader and follower replicas) per broker |
|---|---|
kafka.t3.small |
300 |
kafka.m5.large or kafka.m5.xlarge |
1000 |
kafka.m5.2xlarge |
2000 |
kafka.m5.4xlarge, kafka.m5.8xlarge, kafka.m5.12xlarge,
kafka.m5.16xlarge, or kafka.m5.24xlarge |
4000 |
kafka.m7g.large or kafka.m7g.xlarge |
1000 |
kafka.m7g.2xlarge |
2000 |
kafka.m7g.4xlarge, kafka.m7g.8xlarge,
kafka.m7g.12xlarge, or kafka.m7g.16xlarge |
4000 |
If the number of partitions per broker exceeds the recommended value and your cluster becomes overloaded, you may be prevented from performing the following operations:
-
Update the cluster configuration
-
Update the Apache Kafka version for the cluster
-
Update the cluster to a smaller broker type
-
Associate an AWS Secrets Manager secret with a cluster that has SASL/SCRAM authentication
A high number of partitions can also result in missing Kafka metrics on CloudWatch and on Prometheus scraping.
For guidance on choosing the number of partitions, see Apache Kafka Supports 200K Partitions Per Cluster
Right-size your cluster: Number of brokers per cluster
To determine the right number of brokers for your MSK cluster and understand
costs, see the MSK Sizing
and Pricing
To understand how the underlying infrastructure affects Apache Kafka performance, see
Best practices for right-sizing your Apache Kafka clusters to optimize performance
and cost
Optimize cluster throughput for m5.4xl, m7g.4xl or larger instances
When using m5.4xl, m7g.4xl, or larger instances, you can optimize the cluster throughput by tuning the num.io.threads and num.network.threads configurations.
Num.io.threads is the number of threads that a broker uses for processing requests. Adding more threads, up to the number of CPU cores supported for the instance type, can help improve cluster throughput.
Num.network.threads is the number of threads the broker uses for receiving all incoming requests and returning responses. Network threads place incoming requests on a request queue for processing by io.threads. Setting num.network.threads to half the number of CPU cores supported for the instance type allows for full usage of the new instance type.
Important
Do not increase num.network.threads without first increasing num.io.threads as this can lead to congestion related to queue saturation.
Recommended settings | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Instance type | Recommended value for num.io.threads | Recommended value for num.network.threads | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
m5.4xl |
16 |
8 |
|||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
m5.8xl |
32 |
16 |
|||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
m5.12xl |
48 |
24 |
|||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
m5.16xl |
64 |
32 |
|||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
m5.24xl |
96 |
48 |
|||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
m7g.4xlarge |
16 |
8 |
|||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
m7g.8xlarge |
32 |
16 |
|||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
m7g.12xlarge |
48 |
24 |
|||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
m7g.16xlarge |
64 |
32 |
|||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Use latest Kafka AdminClient to avoid topic ID mismatch issue
The ID of a topic is lost (Error: does not match the topic Id for partition) when you use a Kafka AdminClient version lower than 2.8.0 with the flag --zookeeper to increase or reassign topic partitions for a cluster using Kafka version 2.8.0 or higher. Note that the --zookeeper flag is deprecated in Kafka 2.5 and is removed starting with Kafka 3.0. See Upgrading to 2.5.0 from any version 0.8.x through 2.4.x
To prevent topic ID mismatch, use a Kafka client version 2.8.0 or higher for Kafka admin operations. Alternatively, clients 2.5 and higher can use the --bootstrap-servers flag instead of the --zookeeper flag.
Build highly available clusters
Use the following recommendations so that your MSK cluster can be highly available during an update (such as when you're updating the broker type or Apache Kafka version, for example) or when Amazon MSK is replacing a broker.
-
Set up a three-AZ cluster.
-
Ensure that the replication factor (RF) is at least 3. Note that a RF of 1 can lead to offline partitions during a rolling update; and a RF of 2 may lead to data loss.
-
Set minimum in-sync replicas (minISR) to at most RF - 1. A minISR that is equal to the RF can prevent producing to the cluster during a rolling update. A minISR of 2 allows three-way replicated topics to be available when one replica is offline.
-
Ensure client connection strings include at least one broker from each availability zone. Having multiple brokers in a client's connection string allows for failover when a specific broker is offline for an update. For information about how to get a connection string with multiple brokers, see Getting the bootstrap brokers for an Amazon MSK cluster.
Monitor CPU usage
Amazon MSK strongly recommends that you maintain the total CPU utilization for your brokers
(defined as CPU User + CPU System) under 60%. When you have at least 40% of
your cluster's total CPU available, Apache Kafka can redistribute CPU load across
brokers in the cluster when necessary. One example of when this is necessary is when
Amazon MSK detects and recovers from a broker fault; in this case, Amazon MSK performs automatic
maintenance, like patching. Another example is when a user requests a broker-type change
or version upgrade; in these two cases, Amazon MSK deploys rolling workflows that take one
broker offline at a time. When brokers with lead partitions go offline, Apache Kafka
reassigns partition leadership to redistribute work to other brokers in the cluster. By
following this best practice you can ensure you have enough CPU headroom in your cluster
to tolerate operational events like these.
You can use Amazon CloudWatch metric math to create a composite metric that is
CPU User + CPU System. Set an alarm that gets triggered when the
composite metric reaches an average CPU utilization of 60%. When this alarm is
triggered, scale the cluster using one of the following options:
-
Option 1 (recommended): Update your broker type to the next larger type. For example, if the current type is
kafka.m5.large, update the cluster to usekafka.m5.xlarge. Keep in mind that when you update the broker type in the cluster, Amazon MSK takes brokers offline in a rolling fashion and temporarily reassigns partition leadership to other brokers. A size update typically takes 10-15 minutes per broker. -
Option 2: If there are topics with all messages ingested from producers that use round-robin writes (in other words, messages aren't keyed and ordering isn't important to consumers), expand your cluster by adding brokers. Also add partitions to existing topics with the highest throughput. Next, use
kafka-topics.sh --describeto ensure that newly added partitions are assigned to the new brokers. The main benefit of this option compared to the previous one is that you can manage resources and costs more granularly. Additionally, you can use this option if CPU load significantly exceeds 60% because this form of scaling doesn't typically result in increased load on existing brokers. -
Option 3: Expand your cluster by adding brokers, then reassign existing partitions by using the partition reassignment tool named
kafka-reassign-partitions.sh. However, if you use this option, the cluster will need to spend resources to replicate data from broker to broker after partitions are reassigned. Compared to the two previous options, this can significantly increase the load on the cluster at first. As a result, Amazon MSK doesn't recommend using this option when CPU utilization is above 70% because replication causes additional CPU load and network traffic. Amazon MSK only recommends using this option if the two previous options aren't feasible.
Other recommendations:
-
Monitor total CPU utilization per broker as a proxy for load distribution. If brokers have consistently uneven CPU utilization it might be a sign that load isn't evenly distributed within the cluster. Amazon MSK recommends using Cruise Control to continuously manage load distribution via partition assignment.
-
Monitor produce and consume latency. Produce and consume latency can increase linearly with CPU utilization.
-
JMX scrape interval: If you enable open monitoring with the Prometheus feature, it is recommended that you use a 60 second or higher scrape interval (scrape_interval: 60s) for your Prometheus host configuration (prometheus.yml). Lowering the scrape interval can lead to high CPU usage on your cluster.
Monitor disk space
To avoid running out of disk space for messages, create a CloudWatch alarm that watches the
KafkaDataLogsDiskUsed metric. When the value of this metric reaches or
exceeds 85%, perform one or more of the following actions:
-
Use Automatic scaling. You can also manually increase broker storage as described in Manual scaling.
-
Reduce the message retention period or log size. For information on how to do that, see Adjust data retention parameters.
-
Delete unused topics.
For information on how to set up and use alarms, see Using Amazon CloudWatch Alarms. For a full list of Amazon MSK metrics, see Monitoring an Amazon MSK cluster.
Adjust data retention parameters
Consuming messages doesn't remove them from the log. To free up disk space regularly, you can explicitly specify a retention time period, which is how long messages stay in the log. You can also specify a retention log size. When either the retention time period or the retention log size are reached, Apache Kafka starts removing inactive segments from the log.
To specify a retention policy at the cluster level, set one or more of the following
parameters: log.retention.hours, log.retention.minutes,
log.retention.ms, or log.retention.bytes. For more
information, see Custom MSK configurations.
You can also specify retention parameters at the topic level:
-
To specify a retention time period per topic, use the following command.
kafka-configs.sh --bootstrap-server $bs --alter --entity-type topics --entity-nameTopicName--add-config retention.ms=DesiredRetentionTimePeriod -
To specify a retention log size per topic, use the following command.
kafka-configs.sh --bootstrap-server $bs --alter --entity-type topics --entity-nameTopicName--add-config retention.bytes=DesiredRetentionLogSize
The retention parameters that you specify at the topic level take precedence over cluster-level parameters.
Speeding up log recovery after unclean shutdown
After an unclean shutdown, a broker can take a while to restart as it does log recovery. By default, Kafka only uses a single thread per log directory to perform this recovery. For example, if you have thousands of partitions, log recovery can take hours to complete. To speed up log recovery, it's recommended to increase the number of threads using configuration property num.recovery.threads.per.data.dir. You can set it to the number of CPU cores.
Monitor Apache Kafka memory
We recommend that you monitor the memory that Apache Kafka uses. Otherwise, the cluster may become unavailable.
To determine how much memory Apache Kafka uses, you can monitor the
HeapMemoryAfterGC metric. HeapMemoryAfterGC is the
percentage of total heap memory that is in use after garbage collection. We recommend
that you create a CloudWatch alarm that takes action when HeapMemoryAfterGC
increases above 60%.
The steps that you can take to decrease memory usage vary. They depend on the way that
you configure Apache Kafka. For example, if you use transactional message delivery, you
can decrease the transactional.id.expiration.ms value in your Apache Kafka
configuration from 604800000 ms to 86400000 ms (from 7 days to
1 day). This decreases the memory footprint of each transaction.
Don't add non-MSK brokers
If you use Apache ZooKeeper commands to add brokers, these brokers don't get added to your MSK cluster, and your Apache ZooKeeper will contain incorrect information about the cluster. This might result in data loss. For supported cluster operations, see Amazon MSK: How it works.
Enable in-transit encryption
For information about encryption in transit and how to enable it, see Encryption in transit.
Reassign partitions
To move partitions to different brokers on the same cluster, you can use the partition reassignment tool
named kafka-reassign-partitions.sh. For example, after you add new brokers to expand a cluster, you can
rebalance that cluster by reassigning partitions to the new brokers. For information
about how to add brokers to a cluster, see Expanding an Amazon MSK cluster.
For information about the partition reassignment tool, see Expanding
your cluster
