Linux containers on Fargate container image pull behavior for Amazon ECS
Every Fargate task runs on its own single use, single tenant instance. When you run Linux containers on Fargate, container images or container image layers are not cached on the instance. Therefore, for each container image defined in the task, the whole container image needs to be pulled from the container image registry for each Fargate task. The time it takes to pull the images is directly correlated to the time taken to start an Fargate task.
Take the following into account to optimize the image pull time.
- Container image proximity
-
To reduce the time it takes to download container images, locate the data as close to the compute as possible. Pulling a container image over the internet or across AWS Regions might impact the download time. We recommend that you store the container image in the same Region where the task will run. If you store the container image in Amazon ECR, use a VPC interface endpoint to further reduce the image pull time. For more information, see Amazon ECR interface VPC endpoints (AWS PrivateLink) in the Amazon ECR User Guide.
- Container image size reduction
-
The size of a container image directly impacts the download time. Reducing the size of the container image or the number of container image layers, can reduce the time it takes for an image to download. Lightweight base images (such as the minimal Amazon Linux 2023 container image) can be significantly smaller than those based on traditional operating system base images. For more information about the minimal image, see AL2023 Minimal container image in the Amazon Linux 2023 User Guide.
- Alternative compression algorithms
-
Container image layers are often compressed when pushed to a container image registry. Compressing the container image layer reduces the amount of data that has to be transferred across the network and stored in the container image registry. After a container image layer has been downloaded to an instance by the container runtime, that layer is decompressed. The compression algorithm used and the amount of vCPUs available to the runtime impact the time it takes to decompress the container image. On Fargate, you can increase the size of the task or leverage the more performant zstd compression algorithm to reduce the time taken for decompression. For more information, see zstd
on GitHub. For information about how to implement the images for Fargate, see Reducing AWS Fargate Startup Times with zstd Compressed Container Images . - Lazy Loading container images
-
For large container images (> 250mb), it might be optimal to lazy load a container image rather than downloading all of the container image. On Fargate, you can use Seekable OCI (SOCI) to lazy load a container image from a container image registry. For more information, see soci-snapshotter
on GitHub and Lazy loading container images using Seekable OCI (SOCI).
