How to choose the right Amazon MSK cluster type for you

Amazon Managed Streaming for Apache Kafka (Amazon MSK) is an AWS streaming data service that manages Apache Kafka infrastructure and operations, making it easy for developers and DevOps managers to run Apache Kafka applications and Kafka Connect connectors on AWS, without the need to become experts in operating Apache Kafka. Amazon MSK operates, maintains, and scales Apache Kafka clusters, provides enterprise-grade security features out of the box, and has built-in AWS integrations that accelerate development of streaming data applications. You can easily get started by creating an MSK cluster using the AWS Management Console with a few clicks.

When creating a cluster, you must choose a cluster type from two options: provisioned or serverless. Choosing the best cluster type for each workload depends on the type of workload and your DevOps preferences. Amazon MSK provisioned clusters offer more flexibility in how you scale, configure, and optimize your cluster. Amazon MSK Serverless, on the other hand, makes scaling, load management, and operation of the cluster easier for you. With MSK Serverless, you can run your applications without having to configure, manage the infrastructure, or optimize clusters, and you pay for the data volume you stream and retain. MSK Serverless fully manages partitions, including monitoring as well as ensuring an even balance of partition distribution across brokers in the cluster (auto-balancing).

In this post, I examine a use case with the fictitious company AnyCompany, who plans to use Amazon MSK for two applications. They must decide between provisioned or serverless cluster types. I describe a process by which they work backward from the applications’ requirements to find the best MSK cluster type for their workloads, including how the organizational structure and application requirements are relevant in finding the best offering. Lastly, I examine the requirements and their relationship to Amazon MSK features.

Use case

AnyCompany is an enterprise organization that is ready to move two of their Kafka applications to Amazon MSK.

The first is a large ecommerce platform, which is a legacy application that currently uses a self-managed Apache Kafka cluster run in their data centers. AnyCompany wants to migrate this application to the AWS Cloud and use Amazon MSK to reduce maintenance and operations overhead. AnyCompany has a DataOps team that has been operating self-managed Kafka clusters in their data centers for years. AnyCompany wants to continue using the DataOps team to manage the MSK cluster on behalf of the development team. There is very little flexibility for code changes. For example, a few modules of the application require plaintext communication and access to the Apache ZooKeeper cluster that comes with an MSK cluster. The ingress throughput for this application doesn’t fluctuate often. The ecommerce platform only experiences a surge in user activity during special sales events. The DataOps team has a good understanding of this application’s traffic pattern, and are confident that they can optimize an MSK cluster by setting some custom broker-level configurations.

The second application is a new cloud-native gaming application currently in development. AnyCompany hopes to launch this gaming application soon followed by a marketing campaign. Throughput needs for this application are unknown. The application is expected to receive high traffic initially, then user activity should decline gradually. Because the application is going to launch first in the US, traffic during the day is expected to be higher than at night. This application offers a lot of flexibility in terms of Kafka client version, encryption in transit, and authentication. Because this is a cloud-native application, AnyCompany hopes they can delegate full ownership of its infrastructure to the development team.

Solution overview

Let’s examine a process that helps AnyCompany decide between the two Amazon MSK offerings. The following diagram shows this process at a high level.

In the following sections, I explain each step in detail and the relevant information that AnyCompany needs to collect before they make a decision.

Competency in Apache Kafka

AWS recommends a list of best practices to follow when using the Amazon MSK provisioned offering. Amazon MSK provisioned, offers more flexibility so you make scaling decisions based on what’s best for your workloads. For example, you can save on cost by consolidating a group of workloads into a single cluster. You can decide which metrics are important to monitor and optimize your cluster through applying custom configurations to your brokers. You can choose your Apache Kafka version, among different supported versions, and decide when to upgrade to a new version. Amazon MSK takes care of applying your configuration and upgrading each broker in a rolling fashion.

With more flexibility, you have more responsibilities. You need to make sure your cluster is right-sized at any time. You can achieve this by monitoring a set of cluster-level, broker-level, and topic-level metrics to ensure you have enough resources that are needed for your throughput. You also need to make sure the number of partitions assigned to each broker doesn’t exceed the numbers suggested by Amazon MSK. If partitions are unbalanced, you need to even-load them across all brokers. If you have more partitions than recommended, you need to either upgrade brokers to a larger size or increase the number of brokers in your cluster. There are also best practices for the number of TCP connections when using AWS Identity and Access Management (IAM) authentication.

An MSK Serverless cluster takes away the complexity of right-sizing clusters and balancing partitions across brokers. This makes it easy for developers to focus on writing application code.

AnyCompany has an experienced DataOps team who are familiar with scaling operations and best practices for the MSK provisioned cluster type. AnyCompany can use their DataOps team’s Kafka expertise for building automations and easy-to-follow standard procedures on behalf of the ecommerce application team. The gaming development team is an exception, because they are expected to take the full ownership of the infrastructure.

In the following sections, I discuss other steps in the process before deciding which cluster type is right for each application.

Custom configuration

In certain use cases, you need to configure your MSK cluster differently from its default settings. This could be due to your application requirements. For example, AnyCompany’s ecommerce platform requires setting up brokers such that the default retention period for all topics is set to 72 hours. Also, topics should be or auto-created when they are requested and don’t exist.

The Amazon MSK provisioned offering provides a default configuration for brokers, topics, and Apache ZooKeeper nodes. It also allows you to create custom configurations and use them to create new MSK clusters or update existing clusters. An MSK cluster configuration consists of a set of properties and their corresponding values.

MSK Serverless doesn’t allow applying broker-level configuration. This is because AWS takes care of configuring and managing the backend nodes. It takes away the heavy lifting of configuring the broker nodes. You only need to manage your applications’ topics. To learn more, refer to the list of topic-level configurations that MSK Serverless allows you to change.

Unlike the ecommerce platform, AnyCompany’s gaming application doesn’t need broker-level custom configuration. The developers want to set the and max.message.bytes per each topic only.

Application requirements

Apache Kafka applications differ in terms of their security; the way they connect, write, or read data; data retention period; and scaling patterns. For example, some applications can only scale vertically, whereas other applications can scale only horizontally. Although a flexible application can work with encryption in transit, a legacy application may only be able to communicate in plaintext format.

Cluster-level quotas

Amazon MSK enforces some quotas to ensure the performance, reliability, and availability of the service for all customers. These quotas are subject to change at any time. To access the latest values for each dimension, refer to Amazon MSK quota. Note that some of the quotas are soft limits and can be increased using a support ticket.

When choosing a cluster type in Amazon MSK, it’s important to understand your application requirements and compare those against quotas in relation with each offering. This makes sure you choose the best cluster type that meets your goals and application’s needs. Let’s examine how you can calculate the throughput you need and other important dimensions you need to compare with Amazon MSK quotas:

Number of clusters per account – Amazon MSK may have quotas for how many clusters you can create in a single AWS account. If this is limiting your ability to create more clusters, you can consider creating those in multiple AWS accounts and using secure connectivity patterns to provide access to your applications.

Message size – You need to make sure the maximum message size that your producer writes for a single message is lower than the configured size in the MSK cluster. MSK provisioned clusters allow you to change the default value in a custom configuration. If you choose MSK Serverless, check this value in Amazon MSK quota. The average message size is helpful when calculating the total ingress or egress throughput of the cluster, which I demonstrate later in this post.

Message rate per second – This directly influences total ingress and egress throughput of the cluster. Total ingress throughput equals the message rate per second multiplied by message size. You need to make sure your producer is configured for optimal throughput by adjusting batch.size and properties. If you’re choosing MSK Serverless, you need to make sure you configure your producer to optimal batches with the rate that is lower than its request rate quota.

Number of consumer groups – This directly influences the total egress throughput of the cluster. Total egress throughput equals the ingress throughput multiplied by the number of consumer groups. If you’re choosing MSK Serverless, you need to make sure your application can work with these quotas.

Maximum number of partitions – Amazon MSK provisioned recommends not exceeding certain limits per broker (depending the broker size). If the number of partitions per broker exceeds the maximum value specified in the previous table, you can’t perform certain upgrade or update operations. MSK Serverless also has a quota of maximum number of partitions per cluster. You can request to increase the quota by creating a support case.

Partition-level quotas

Apache Kafka organizes data in structures called topics. Each topic consists of a single or many partitions. Partitions are the degree of parallelism in Apache Kafka. The data is distributed across brokers using data partitioning. Let’s examine a few important Amazon MSK requirements, and how you can make sure which cluster type works better for your application:

Maximum throughput per partition – MSK Serverless automatically balances the partitions of your topic between the backend nodes. It instantly scales when your ingress throughput increases. However, each partition has a quota of how much data it accepts. This is to ensure the data is distributed evenly across all partitions and backend nodes. In an MSK Serverless cluster, you need to create your topic with enough partitions such that the aggregated throughput is equal to the maximum throughput your application requires. You also need to make sure your consumers read data with a rate that is below the maximum egress throughput per partition quota. If you’re using Amazon MSK provisioned, there is no partition-level quota for write and read operations. However, AWS recommends you monitor and detect hot partitions and control how partitions should balance among the broker nodes.

Data storage – The amount of time each message is kept in a particular topic directly influences the total amount of storage needed for your cluster. Amazon MSK allows you to manage the retention period at the topic level. MSK provisioned clusters allow broker-level configuration to set the default data retention period. MSK Serverless clusters allow unlimited data retention, but there is a separate quota for the maximum data that can be stored in each partition.


Amazon MSK recommends that you secure your data in the following ways. Availability of the security features varies depending on the cluster type. Before making a decision about your cluster type, check if your preferred security options are supported by your choice of cluster type.

Encryption at rest – Amazon MSK integrates with AWS Key Management Service (AWS KMS) to offer transparent server-side encryption. Amazon MSK always encrypts your data at rest. When you create an MSK cluster, you can specify the KMS key that you want Amazon MSK to use to encrypt your data at rest.

Encryption in transit – Amazon MSK uses TLS 1.2. By default, it encrypts data in transit between the brokers of your MSK cluster. You can override this default when you create the cluster. For communication between clients and brokers, you must specify one of the following settings:

Only allow TLS encrypted data. This is the default setting.
Allow both plaintext and TLS encrypted data.
Only allow plaintext data.

Authentication and authorization – Use IAM to authenticate clients and allow or deny Apache Kafka actions. Alternatively, you can use TLS or SASL/SCRAM to authenticate clients, and Apache Kafka ACLs to allow or deny actions.

Cost of ownership

Amazon MSK helps you avoid spending countless hours and significant resources just managing your Apache Kafka cluster, adding little or no value to your business. With a few clicks on the Amazon MSK console, you can create highly available Apache Kafka clusters with settings and configuration based on Apache Kafka’s deployment best practices. Amazon MSK automatically provisions and runs Apache Kafka clusters. Amazon MSK continuously monitors cluster health and automatically replaces unhealthy nodes with no application downtime. In addition, Amazon MSK secures Apache Kafka clusters by encrypting data at rest and in transit. These capabilities can significantly reduce your Total Cost of Ownership (TCO).

With MSK provisioned clusters, you can specify and then scale cluster capacity to meet your needs. With MSK Serverless clusters, you don’t need to specify or scale cluster capacity. MSK Serverless automatically scales the cluster capacity based on the throughput, and you only pay per GB of data that your producers write to and your consumers read from the topics in your cluster. Additionally, you pay an hourly rate for your serverless clusters and an hourly rate for each partition that you create. The MSK Serverless cluster type generally offers a lower cost of ownership by taking away the cost of engineering resources needed for monitoring, capacity planning, and scaling MSK clusters. However, if your organization has a DataOps team with Kafka competency, you can use this competency to operate optimized MSK provisioned clusters. This allows you to save on Amazon MSK costs by consolidating several Kafka applications into a single cluster. There are a few critical considerations to decide when and how to split your workloads between multiple MSK clusters.

Apache ZooKeeper

Apache ZooKeeper is a service included in Amazon MSK when you create a cluster. It manages the Apache Kafka metadata and acts as a quorum controller for leader elections. Although interacting with ZooKeeper is not a recommended pattern, some Kafka applications have a dependency to connect directly to ZooKeeper. During the migration to Amazon MSK, you may find a few of these applications in your organization. This could be because they use an older version of the Kafka client library or other reasons. For example, applications that help with Apache Kafka admin operations or visibility such as Cruise Control usually need this kind of access.

Before you choose your cluster type, you first need to check which offering provides direct access to the ZooKeeper cluster. As of writing this post, only Amazon MSK provisioned provides direct access to ZooKeeper.

How AnyCompany chooses their cluster types

AnyCompany first needs to collect some important requirements about each of their applications. The following table shows these requirements. The rows marked with an asterisk (*) are calculated based on the values in previous rows.

Ecommerce Platform
Gaming Application

Message rate per second

Maximum message size
15 MB
1 MB

Average message size
30 KB
15 KB

* Ingress throughput (average message size * message rate per second)

Number of consumer groups

* Outgress throughput (ingress throughput * number of consumer groups)
9 GBps
15 MBps

Number of topics

Average partition per topic

* Total number of partitions (number of topics * average partition per topic)

* Ingress per partition (ingress throughput / total number of partitions)
450 KBps
300 KBps

* Outgress per partition (outgress throughput / total number of partitions)
900 KBps
300 KBps

Data retention
72 hours
168 hours

* Total storage needed (ingress throughput * retention period in seconds)
1,139.06 TB
1.3 TB

Plaintext and SASL/SCRAM

Need ZooKeeper access

For the gaming application, AnyCompany doesn’t want to use their in-house Kafka competency to support an MSK provisioned cluster. Also, the gaming application doesn’t need custom configuration, and its throughput needs are below the quotas set by the MSK Serverless cluster type. In this scenario, an MSK Serverless cluster makes more sense.

For the e-commerce platform, AnyCompany wants to use their Kafka competency. Moreover, their throughput needs exceed the MSK Serverless quotas, and the application requires some broker-level custom configuration. The ecommerce platform also can’t split between multiple clusters. Because of these reasons, AnyCompany chooses the MSK provisioned cluster type in this scenario. Additionally, AnyCompany can save more on cost with the Amazon MSK provisioned pricing model. Their throughput is consistent at most times and AnyCompany wants to use their DataOps team to optimize a provisioned MSK cluster and make scaling decisions based on their own expertise.


Choosing the best cluster type for your applications may seem complicated at first. In this post, I showed a process that helps you work backward from your application’s requirement and the resources available to you. MSK provisioned clusters offer more flexibility in how you scale, configure, and optimize your cluster. MSK Serverless, on the other hand, is a cluster type that makes it easier for you to run Apache Kafka clusters without having to manage compute and storage capacity. I generally recommend you begin with MSK Serverless if your application doesn’t require broker-level custom configurations, and your application throughput needs don’t exceed the quotas for the MSK Serverless cluster type. Sometimes it’s best to split your workloads between multiple MSK Serverless clusters, but if that isn’t possible, you may need to consider an MSK provisioned cluster. To operate an optimized MSK provisioned cluster, you need to have Kafka competency within your organization.

For further reading on Amazon MSK, visit the official product page.

About the author

Ali Alemi is a Streaming Specialist Solutions Architect at AWS. Ali advises AWS customers with architectural best practices and helps them design real-time analytics data systems that are reliable, secure, efficient, and cost-effective. He works backward from customers’ use cases and designs data solutions to solve their business problems. Prior to joining AWS, Ali supported several public sector customers and AWS consulting partners in their application modernization journey and migration to the cloud.