AWS Big Data Blog

Real-time cost savings for Amazon Managed Service for Apache Flink

When running Apache Flink applications on Amazon Managed Service for Apache Flink, you have the unique benefit of taking advantage of its serverless nature. This means that cost-optimization exercises can happen at any time—they no longer need to happen in the planning phase. With Managed Service for Apache Flink, you can add and remove compute with the click of a button.

Apache Flink is an open source stream processing framework used by hundreds of companies in critical business applications, and by thousands of developers who have stream-processing needs for their workloads. It is highly available and scalable, offering high throughput and low latency for the most demanding stream-processing applications. These scalable properties of Apache Flink can be key to optimizing your cost in the cloud.

Managed Service for Apache Flink is a fully managed service that reduces the complexity of building and managing Apache Flink applications. Managed Service for Apache Flink manages the underlying infrastructure and Apache Flink components that provide durable application state, metrics, logs, and more.

In this post, you can learn about the Managed Service for Apache Flink cost model, areas to save on cost in your Apache Flink applications, and overall gain a better understanding of your data processing pipelines. We dive deep into understanding your costs, understanding whether your application is overprovisioned, how to think about scaling automatically, and ways to optimize your Apache Flink applications to save on cost. Lastly, we ask important questions about your workload to determine if Apache Flink is the right technology for your use case.

How costs are calculated on Managed Service for Apache Flink

To optimize for costs with regards to your Managed Service for Apache Flink application, it can help to have a good idea of what goes into the pricing for the managed service.

Managed Service for Apache Flink applications are comprised of Kinesis Processing Units (KPUs), which are compute instances composed of 1 virtual CPU and 4 GB of memory. The total number of KPUs assigned to the application is determined by multiplying two parameters that you control directly:

  • Parallelism – The level of parallel processing in the Apache Flink application
  • Parallelism per KPU – The number of resources dedicated to each parallelism

The number of KPUs is determined by the simple formula: KPU = Parallelism / ParallelismPerKPU, rounded up to the next integer.

An additional KPU per application is also charged for orchestration and not directly used for data processing.

The total number of KPUs determines the number of resources, CPU, memory, and application storage allocated to the application. For each KPU, the application receives 1 vCPU and 4 GB of memory, of which 3 GB are allocated by default to the running application and the remaining 1 GB is used for application state store management. Each KPU also comes with 50 GB of storage attached to the application. Apache Flink retains application state in-memory to a configurable limit, and spillover to the attached storage.

The third cost component is durable application backups, or snapshots. This is entirely optional and its impact on the overall cost is small, unless you retain a very large number of snapshots.

At the time of writing, each KPU in the US East (Ohio) AWS Region costs $0.11 per hour, and attached application storage costs $0.10 per GB per month. The cost of durable application backup (snapshots) is $0.023 per GB per month. Refer to Amazon Managed Service for Apache Flink Pricing for up-to-date pricing and different Regions.

The following diagram illustrates the relative proportions of cost components for a running application on Managed Service for Apache Flink. You control the number of KPUs via the parallelism and parallelism per KPU parameters. Durable application backup storage is not represented.

pricing model

In the following sections, we examine how to monitor your costs, optimize the usage of application resources, and find the required number of KPUs to handle your throughput profile.

AWS Cost Explorer and understanding your bill

To see what your current Managed Service for Apache Flink spend is, you can use AWS Cost Explorer.

On the Cost Explorer console, you can filter by date range, usage type, and service to isolate your spend for Managed Service for Apache Flink applications. The following screenshot shows the past 12 months of cost broken down into the price categories described in the previous section. The majority of spend in many of these months was from interactive KPUs from Amazon Managed Service for Apache Flink Studio.

Analyse the cost of your Apache Flink application with AWS Cost Explorer

Using Cost Explorer can not only help you understand your bill, but help further optimize particular applications that may have scaled beyond expectations automatically or due to throughput requirements. With proper application tagging, you could also break this spend down by application to see which applications account for the cost.

Signs of overprovisioning or inefficient use of resources

To minimize costs associated with Managed Service for Apache Flink applications, a straightforward approach involves reducing the number of KPUs your applications use. However, it’s crucial to recognize that this reduction could adversely affect performance if not thoroughly assessed and tested. To quickly gauge whether your applications might be overprovisioned, examine key indicators such as CPU and memory usage, application functionality, and data distribution. However, although these indicators can suggest potential overprovisioning, it’s essential to conduct performance testing and validate your scaling patterns before making any adjustments to the number of KPUs.


Analyzing metrics for your application on Amazon CloudWatch can reveal clear signals of overprovisioning. If the containerCPUUtilization and containerMemoryUtilization metrics consistently remain below 20% over a statistically significant period for your application’s traffic patterns, it might be viable to scale down and allocate more data to fewer machines. Generally, we consider applications appropriately sized when containerCPUUtilization hovers between 50–75%. Although containerMemoryUtilization can fluctuate throughout the day and be influenced by code optimization, a consistently low value for a substantial duration could indicate potential overprovisioning.

Parallelism per KPU underutilized

Another subtle sign that your application is overprovisioned is if your application is purely I/O bound, or only does simple call-outs to databases and non-CPU intensive operations. If this is the case, you can use the parallelism per KPU parameter within Managed Service for Apache Flink to load more tasks onto a single processing unit.

You can view the parallelism per KPU parameter as a measure of density of workload per unit of compute and memory resources (the KPU). Increasing parallelism per KPU above the default value of 1 makes the processing more dense, allocating more parallel processes on a single KPU.

The following diagram illustrates how, by keeping the application parallelism constant (for example, 4) and increasing parallelism per KPU (for example, from 1 to 2), your application uses fewer resources with the same level of parallel runs.

How KPUs are calculated

The decision of increasing parallelism per KPU, like all recommendations in this post, should be taken with great care. Increasing the parallelism per KPU value can put more load on a single KPU, and it must be willing to tolerate that load. I/O-bound operations will not increase CPU or memory utilization in any meaningful way, but a process function that calculates many complex operations against the data would not be an ideal operation to collate onto a single KPU, because it could overwhelm the resources. Performance test and evaluate if this is a good option for your applications.

How to approach sizing

Before you stand up a Managed Service for Apache Flink application, it can be difficult to estimate the number of KPUs you should allocate for your application. In general, you should have a good sense of your traffic patterns before estimating. Understanding your traffic patterns on a megabyte-per-second ingestion rate basis can help you approximate a starting point.

As a general rule, you can start with one KPU per 1 MB/s that your application will process. For example, if your application processes 10 MB/s (on average), you would allocate 10 KPUs as a starting point for your application. Keep in mind that this is a very high-level approximation that we have seen effective for a general estimate. However, you also need to performance test and evaluate whether or not this is an appropriate sizing in the long term based on metrics (CPU, memory, latency, overall job performance) over a long period of time.

To find the appropriate sizing for your application, you need to scale up and down the Apache Flink application. As mentioned, in Managed Service for Apache Flink you have two separate controls: parallelism and parallelism per KPU. Together, these parameters determine the level of parallel processing within the application and the overall compute, memory, and storage resources available.

The recommended testing methodology is to change parallelism or parallelism per KPU separately, while experimenting to find the right sizing. In general, only change parallelism per KPU to increase the number of parallel I/O-bound operations, without increasing the overall resources. For all other cases, only change parallelism—KPU will change consequentially—to find the right sizing for your workload.

You can also set parallelism at the operator level to restrict sources, sinks, or any other operator that might need to be restricted and independent of scaling mechanisms. You could use this for an Apache Flink application that reads from an Apache Kafka topic that has 10 partitions. With the setParallelism() method, you could restrict the KafkaSource to 10, but scale the Managed Service for Apache Flink application to a parallelism higher than 10 without creating idle tasks for the Kafka source. It is recommended for other data processing cases to not statically set operator parallelism to a static value, but rather a function of the application parallelism so that it scales when the overall application scales.

Scaling and auto scaling

In Managed Service for Apache Flink, modifying parallelism or parallelism per KPU is an update of the application configuration. It causes the application to automatically take a snapshot (unless disabled), stop the application, and restart it with the new sizing, restoring the state from the snapshot. Scaling operations don’t cause data loss or inconsistencies, but it does pause data processing for a short period of time while infrastructure is added or removed. This is something you need to consider when rescaling in a production environment.

During the testing and optimization process, we recommend disabling automatic scaling and modifying parallelism and parallelism per KPU to find the optimal values. As mentioned, manual scaling is just an update of the application configuration, and can be run via the AWS Management Console or API with the UpdateApplication action.

When you have found the optimal sizing, if you expect your ingested throughput to vary considerably, you may decide to enable auto scaling.

In Managed Service for Apache Flink, you can use multiple types of automatic scaling:

  • Out-of-the-box automatic scaling – You can enable this to adjust the application parallelism automatically based on the containerCPUUtilization metric. Automatic scaling is enabled by default on new applications. For details about the automatic scaling algorithm, refer to Automatic Scaling.
  • Fine-grained, metric-based automatic scaling – This is straightforward to implement. The automation can be based on virtually any metrics, including custom metrics your application exposes.
  • Scheduled scaling – This may be useful if you expect peaks of workload at given times of the day or days of the week.

Out-of-the-box automatic scaling and fine-grained metric-based scaling are mutually exclusive. For more details about fine-grained metric-based auto scaling and scheduled scaling, and a fully working code example, refer to Enable metric-based and scheduled scaling for Amazon Managed Service for Apache Flink.

Code optimizations

Another way to approach cost savings for your Managed Service for Apache Flink applications is through code optimization. Un-optimized code will require more machines to perform the same computations. Optimizing the code could allow for lower overall resource utilization, which in turn could allow for scaling down and cost savings accordingly.

The first step to understanding your code performance is through the built-in utility within Apache Flink called Flame Graphs.

Flame graph

Flame Graphs, which are accessible via the Apache Flink dashboard, give you a visual representation of your stack trace. Each time a method is called, the bar that represents that method call in the stack trace gets larger proportional to the total sample count. This means that if you have an inefficient piece of code with a very long bar in the flame graph, this could be cause for investigation as to how to make this code more efficient. Additionally, you can use Amazon CodeGuru Profiler to monitor and optimize your Apache Flink applications running on Managed Service for Apache Flink.

When designing your applications, it is recommended to use the highest-level API that is required for a particular operation at a given time. Apache Flink offers four levels of API support: Flink SQL, Table API, Datastream API, and ProcessFunction APIs, with increasing levels of complexity and responsibility. If your application can be written entirely in the Flink SQL or Table API, using this can help take advantage of the Apache Flink framework rather than managing state and computations manually.

Data skew

On the Apache Flink dashboard, you can gather other useful information about your Managed Service for Apache Flink jobs.

Open the Flink Dashboard

On the dashboard, you can inspect individual tasks within your job application graph. Each blue box represents a task, and each task is composed of subtasks, or distributed units of work for that task. You can identify data skew among subtasks this way.

Flink dashboard

Data skew is an indicator that more data is being sent to one subtask than another, and that a subtask receiving more data is doing more work than the other. If you have such symptoms of data skew, you can work to eliminate it by identifying the source. For example, a GroupBy or KeyedStream could have a skew in the key. This would mean that data is not evenly spread among keys, resulting in an uneven distribution of work across Apache Flink compute instances. Imagine a scenario where you are grouping by userId, but your application receives data from one user significantly more than the rest. This can result in data skew. To eliminate this, you can choose a different grouping key to evenly distribute the data across subtasks. Keep in mind that this will require code modification to choose a different key.

When the data skew is eliminated, you can return to the containerCPUUtilization and containerMemoryUtilization metrics to reduce the number of KPUs.

Other areas for code optimization include making sure that you’re accessing external systems via the Async I/O API or via a data stream join, because a synchronous query out to a data store can create slowdowns and issues in checkpointing. Additionally, refer to Troubleshooting Performance for issues you might experience with slow checkpoints or logging, which can cause application backpressure.

How to determine if Apache Flink is the right technology

If your application doesn’t use any of the powerful capabilities behind the Apache Flink framework and Managed Service for Apache Flink, you could potentially save on cost by using something simpler.

Apache Flink’s tagline is “Stateful Computations over Data Streams.” Stateful, in this context, means that you are using the Apache Flink state construct. State, in Apache Flink, allows you to remember messages you have seen in the past for longer periods of time, making things like streaming joins, deduplication, exactly-once processing, windowing, and late-data handling possible. It does so by using an in-memory state store. On Managed Service for Apache Flink, it uses RocksDB to maintain its state.

If your application doesn’t involve stateful operations, you may consider alternatives such as AWS Lambda, containerized applications, or an Amazon Elastic Compute Cloud (Amazon EC2) instance running your application. The complexity of Apache Flink may not be necessary in such cases. Stateful computations, including cached data or enrichment procedures requiring independent stream position memory, may warrant Apache Flink’s stateful capabilities. If there’s a potential for your application to become stateful in the future, whether through prolonged data retention or other stateful requirements, continuing to use Apache Flink could be more straightforward. Organizations emphasizing Apache Flink for stream processing capabilities may prefer to stick with Apache Flink for stateful and stateless applications so all their applications process data in the same way. You should also factor in its orchestration features like exactly-once processing, fan-out capabilities, and distributed computation before transitioning from Apache Flink to alternatives.

Another consideration is your latency requirements. Because Apache Flink excels at real-time data processing, using it for an application with a 6-hour or 1-day latency requirement does not make sense. The cost savings by switching to a temporal batch process out of Amazon Simple Storage Service (Amazon S3), for example, would be significant.


In this post, we covered some aspects to consider when attempting cost-savings measures for Managed Service for Apache Flink. We discussed how to identify your overall spend on the managed service, some useful metrics to monitor when scaling down your KPUs, how to optimize your code for scaling down, and how to determine if Apache Flink is right for your use case.

Implementing these cost-saving strategies not only enhances your cost efficiency but also provides a streamlined and well-optimized Apache Flink deployment. By staying mindful of your overall spend, using key metrics, and making informed decisions about scaling down resources, you can achieve a cost-effective operation without compromising performance. As you navigate the landscape of Apache Flink, constantly evaluating whether it aligns with your specific use case becomes pivotal, so you can achieve a tailored and efficient solution for your data processing needs.

If any of the recommendations discussed in this post resonate with your workloads, we encourage you to try them out. With the metrics specified, and the tips on how to understand your workloads better, you should now have what you need to efficiently optimize your Apache Flink workloads on Managed Service for Apache Flink. The following are some helpful resources you can use to supplement this post:

About the Authors

Jeremy BerJeremy Ber has been working in the telemetry data space for the past 10 years as a Software Engineer, Machine Learning Engineer, and most recently a Data Engineer. At AWS, he is a Streaming Specialist Solutions Architect, supporting both Amazon Managed Streaming for Apache Kafka (Amazon MSK) and Amazon Managed Service for Apache Flink.

Lorenzo NicoraLorenzo Nicora works as Senior Streaming Solution Architect at AWS, helping customers across EMEA. He has been building cloud-native, data-intensive systems for over 25 years, working in the finance industry both through consultancies and for FinTech product companies. He has leveraged open-source technologies extensively and contributed to several projects, including Apache Flink.