AWS DevOps Blog

Strategies to optimize the costs of your builds on AWS CodeBuild

AWS CodeBuild is a fully managed continuous integration service that compiles source code, runs tests, and produces ready-to-deploy software packages. With CodeBuild, you don’t need to provision, manage, and scale your own build servers. You just specify the location of your source code and choose your build settings, and CodeBuild will run your build scripts for compiling, testing, and packaging your code.

CodeBuild uses simple pay-as-you-go pricing. There are no upfront costs or minimum fees. You pay only for the resources you use. You are charged for compute resources based on the duration it takes for your build to execute.

There are three main factors that contribute to build costs with CodeBuild:

  • Build duration
  • Compute types
  • Additional services

Understanding how to balance these factors is key to optimizing costs on AWS and this blog post will take a look at each.

Compute Types

CodeBuild offers three compute instance types with different amounts of memory and CPU, for example the Linux GPU Large compute type has 255GB of memory and 32 vCPUs and enables you to execute CI/CD workflow for deep learning purpose (ML/AI) with AWS CodePipeline. Incremental changes in your code, data, and ML models can now be tested for accuracy, before the changes are released through your pipeline.

The Linux 2XLarge instance type is another instance type with 145GB of memory and 72 vCPUs and is suitable for building large and complex applications that require high memory and CPU resources. It can help reduce build time, speed up delivery, and support multiple build environments.

The GPU and 2XLarge compute types are powerful but are also the most expensive compute types per minute. For most build tasks the small, medium or large instance compute types are more than adequate. Using the pricing listed in US East (Ohio) we can see the price variance between the small, medium and large Linux instance types in Figure 1 below.

Visual comparison of small, medium and large compute types vs cost per minute

Figure 1. AWS CodeBuild small, medium and large compute types vs cost per minute

Analyzing the CodeBuild compute costs leads us to a number of cost optimization considerations.

Right Sizing AWS CodeBuild Compute Types to Match Workloads

Right sizing is the process of matching instance types and sizes to your workload performance and capacity requirements at the lowest possible cost. It’s also the process of looking at deployed instances and identifying opportunities to eliminate or downsize without compromising capacity or other requirements, which results in lower costs.

Right sizing is a key mechanism for optimizing AWS costs, but it is often ignored by organizations when they first move to the AWS Cloud. They lift and shift their environments and expect to right size later. Speed and performance are often prioritized over cost, which results in oversized instances and a lot of wasted spend on unused resources.

CodeBuild monitors build resource utilization on your behalf and reports metrics through Amazon CloudWatch. These include metrics such as

  • CPU
  • Memory
  • Disk I/O

These metrics can be seen within the CodeBuild console, for an example see Figure 2 below:

CloudWatch Resource utilization metrics

Figure 2. Resource utilization metrics

Leveraging observability to measuring build resource usage is key to understanding how to rightsize and CodeBuild makes this easy with CloudWatch metrics readily available through the CodeBuild console.

Consider ARM / Graviton

If we compare the costs of arm1.small and general1.small over a ten minute period we can see that the arm based compute type is 32% less expensive.

Visual comparison of small arm and general compute types vs cost per minute

Figure 3. Comparison of small arm and general compute types

But cost per minute while building is not the only benefit here, ARM processors are known for their energy efficiency and high performance. Compiling code directly on an ARM processor can potentially lead to faster execution times and improved overall system performance.

The ideal workload to migrate to ARM is one that doesn’t use architecture-specific dependencies or binaries, already runs on Linux and uses open-source components, for example: Migrating AWS Lambda functions to Arm-based AWS Graviton2 processors.

AWS Graviton processors are custom built by Amazon Web Services using 64-bit Arm Neoverse cores to deliver the best price performance for your cloud workloads. The AWS Graviton Fast Start program helps you quickly and easily move your workloads to AWS Graviton in as little as four hours for applications such as serverless, containerized, database, and caching.

Consider migrating Windows workloads to Linux

If we compare the cost of a general1.medium Windows vs Linux compute type we can see that the Linux Compute type is 43% less expensive over ten minutes:

Comparison of build times on Windows compared to Linux

Figure 4. Build times on Windows compared to Linux

Migrating to Linux is one strategy to not only reduce the costs of building and testing code in CodeBuild but also the cost of running in production.

The effort required to re-platform from Windows to Linux varies depending on how the application was implemented. The key is to identify and target workloads with the right characteristics, balancing strategic importance and implementation effort.

For example, older .Net applications may be able to be migrated to later versions of .NET (previously named .Net Core) first before deploying to Linux. AWS have a Porting Assistant for .NET that is an analysis tool that scans .NET Framework applications and generates a cross-platform compatibility assessment, helping you port your applications to Linux faster.

See our guide on how to escape unfriendly licensing practices by migrating Windows workloads to Linux.

Build duration

One of the dimensions of the CodeBuild pricing is the duration of each build. This is calculated in minutes, from the time you submit your build until your build is terminated, rounded up to the nearest minute. For example: if your build takes a total of 35 seconds using one arm1.small Linux instance on US East (Ohio), each build will cost the price of the full minute, which is $0.0034 in that case. Similarly, if your build takes a total of 5 minutes and 20 seconds, you’ll be charged for 6 minutes.

When you define your CodeBuild project, within a buildspec file, you can specify some of the phases of your builds. The phases you can specify are install, pre-build, build, and post-build. See the documentation to learn more about what each of those phases represent. Besides that, you can define how and where to upload reports and artifacts each build generates. It means that on each of those steps, you should do only what is necessary for the task you want to achieve. Installing dependencies that you won’t need, running commands that aren’t related to your task, or performing tests that aren’t necessary will affect your build time and unnecessarily increase your costs. Packaging and uploading target artifacts with unnecessary large files would cause a similar result.

On top of the CodeBuild phases and steps that you are directly in control, each time you start a build, it takes additional time to queue the task, provision the environment, download the source code (if applicable), and finalize. See Figure 5 below a breakdown of a succeeded build:

AWS CodeBuild phase details

Figure 5. AWS CodeBuild Phase details

In the above example, for each build, it takes approximately 42 seconds on top of what is specified in the buildspec file. Considering this overhead, having several smaller builds instead of fewer larger builds can potentially increase your costs. With this in mind, you have to keep your builds as short as possible, by doing only what is necessary, so that you can minimize the costs. Furthermore, you have to find a good balance between the duration and the frequency of your builds, so that the overhead doesn’t take a large proportion of your build time. Let’s explore some approaches you can factor in to find this balance.

Build caching

A common way to save time and cost on your CodeBuild builds is with build caching. With build caching, you are able to store reusable pieces of your build environment, so that you can save time next time you start a new build. There are two types of caching:

  • Amazon S3 — Stores the cache in an Amazon S3 bucket that is available across multiple build hosts. If you have build artifacts that are more expensive to build than to download, this is a good option for you. For large build artifacts, this may not be the best option, because it can take longer to transfer over your network.
  • Local caching — Stores a cache locally on a build host that is available to that build host only. When you choose this option, the cache is immediately available on the build host, making it a good option for large build artifacts that would take long network transfer time. If you choose local caching, there are multiple cache modes you can choose including source cache mode, docker layer cache mode and custom cache mode.

Docker specific optimizations

Another strategy to optimize your build time and reduce your costs is using custom Docker images. When you specify your CodeBuild project, you can either use one of the default Docker images provided by CodeBuild, or use your own build environment packaged as a Docker image. When you create your own build environment as a Docker image, you can pre-package it with all the tools, test assets, and required dependencies. This can potentially save a significant amount of time, because on the install phase you won’t need to download packages from the internet, and on the build phase, when applicable, you won’t need to download e.g., large static test datasets.

To achieve that, you must specify the image value on the environment configuration when creating or updating your CodeBuild project. See Docker in custom image sample for CodeBuild to learn more about how to configure that. Keep in mind that larger Docker images can negatively affect your build time, therefore you should aim to keep your custom build environment as lean as possible, with only the mandatory contents. Another aspect to consider is to use Amazon Elastic Container Registry (ECR) to store your Docker images. Downloading the image from within the AWS network will be, in most of the cases, faster than downloading it from the public internet and can avoid bottlenecks from public repositories.

Consider which tests to run on the feature branch

If you are using a feature-branch approach, consider carefully which build steps and tests you are going to run on your branches. Running unit tests is a good example of what you should run on the feature branches, but unless you have very specific requirements, you probably don’t need penetration or integration tests at this point. Usually the feature branch changes often, hence running all types of tests all the time is a potential waste. Prefer to have your complex, long-running tests at a later stage of your CI/CD pipeline, as you build confidence on the version that you are to release.

Build once, deploy everywhere

It’s widely considered a best practice to avoid environment-specific code builds, therefore consider a build once, deploy everywhere strategy. There are many benefits to separating environment configuration from the build including reducing build costs, improve maintainability, scalability, and reduce the risk of errors.

Build once, deploy everywhere can be seen in the AWS Deployment Pipeline Reference Architecture where the Beta, Gamma and Prod stages are created from a single artifact created in the Build Stage:

Application Pipeline reference architecture

Figure 6. Application Pipeline reference architecture

Additional Services

CloudWatch Logs and Metrics

Amazon CloudWatch can be used to monitor your builds, report when something goes wrong, take automatic actions when appropriate or simply keep logs of your builds.

CloudWatch metrics show the behavior of your builds over time. For example, you can monitor:

  • How many builds were attempted in a build project or an AWS account over time.
  • How many builds were successful in a build project or an AWS account over time.
  • How many builds failed in a build project or an AWS account over time.
  • How much time CodeBuild spent running builds in a build project or an AWS account over time.
  • Build resource utilization for a build or an entire build project. Build resource utilization metrics include metrics such as CPU, memory, and storage utilization.

However, you may incur charges from Amazon CloudWatch Logs for build log streams. For more information, see Monitoring AWS Codebuild in the CodeBuild User Guide and the CloudWatch pricing page.

Storage Costs

You can create an CodeBuild build project with a set of output artifacts and publish then to S3 buckets. Using S3 as a repository for your artifacts, you only pay for what you use. Check the S3 pricing page.

Encryption

Cloud security at AWS is the highest priority and encryption is an important part of CodeBuild security. Some encryption, such as for data in-transit, is provided by default and does not require you to do anything. Other encryption, such as for data at-rest, you can configure when you create your project or build. Codebuild uses Amazon KMS to encrypt the data at-rest.

Build artifacts, such as a cache, logs, exported raw test report data files, and build results, are encrypted by default using AWS managed keys and are free of charge. Consider using these keys if you don’t need to create your own key.

If you do not want to use these KMS keys, you can create and configure a customer managed key. For more information, see the documentation on creating KMS Keys and AWS Key Management Service concepts in the AWS Key Management Service User Guide.

Check the KMS pricing page.

Data transfer costs

You may incur additional charges if your builds transfer data, for example:

  • Avoid routing traffic over the internet when connecting to AWS services from within AWS by using VPC endpoints
  • Traffic that crosses an Availability Zone boundary typically incurs a data transfer charge. Use resources from the local Availability Zone whenever possible.
  • Traffic that crosses a Regional boundary will typically incur a data transfer charge. Avoid cross-Region data transfer unless your business case requires it
  • Use the AWS Pricing Calculator to help estimate the data transfer costs for your solution.
  • Use a dashboard to better visualize data transfer charges – this workshop will show how.

Here’s an Overview of Data Transfer Costs for Common Architectures on AWS.

Conclusion

In this blog post we discussed how compute types; build duration and use of additional services contribute to build costs with AWS CodeBuild.

We highlighted how right sizing compute types is an important practice for teams that want to reduce their build costs while still achieving optimal performance. The key to optimizing is by measuring and observing the workload and selecting the most appropriate compute instance based on requirements.

Further compute type cost optimizations can be found by targeting AWS Graviton processors and Linux environments. AWS Graviton Processors in particular offer several advantages over traditional x86-based instances and are designed by AWS to deliver the best price performance for your cloud workloads.

For further reading, see the summary of CI/CD best practices from the Practicing Continuous Integration and Continuous Delivery on AWS whitepaper,  my CI/CD pipeline is my release captain and also the cost optimization pillar from the AWS Well-Architected Framework which focuses on avoiding unnecessary costs.

About the authors:

Leandro Cavalcante Damascena

Leandro Damascena is a Solutions Architect at AWS, focused on Application Modernization and Serverless. Prior to that, he spent 20 years in various roles across the software delivery lifecycle, from development to architecture and operations. Outside of work, he loves spending time with his family and enjoying water sports like kitesurfing and wakeboarding.

Rafael Ramos

Rafael is a Solutions Architect at AWS, where he helps ISVs on their journey to the cloud. He spent over 13 years working as a software developer, and is passionate about DevOps and serverless. Outside of work, he enjoys playing tabletop RPG, cooking and running marathons.

Matt Laver

Matt Laver is a Solutions Architect at AWS working with SMB customers in the UK. He is passionate about DevOps and loves helping customers find simple solutions to difficult problems.