DOCSEARCHbyAlgolia
Submit Docs

Kubeflow on AWS

website·Indexed 30 days ago

https://awslabs.github.io/kubeflow-manifests/docs/

An open source distribution of Kubeflow designed for building machine learning systems on Amazon EKS. This documentation covers deployment guides (including vanilla, Cognito, RDS, and S3 configurations), component guides for Notebooks, Pipelines, and KServe, and AWS-specific integrations such as Amazon EFS, FSx for Lustre, and CloudWatch.

Tokens
145.8K
Snippets
657
Records
920
Agent score
62%

What's inside Kubeflow on AWS

  1. Overview of Kubeflow on AWS Deployment Add-ons

    Kubeflow on AWS provides several add-ons to integrate additional AWS services into the deployment, regardless of the chosen deployment option. Available integrations include:

    • AWS Application Load Balancer (ALB): Used to expose the Kubeflow dashboard to external traffic with secure traffic management.
    • Amazon Elastic File System (EFS): Provides persistent, scalable, and shareable volumes for Kubeflow Notebooks, training, and inference workloads (e.g., Jupyter).
    • Amazon FSx for Lustre: High-performance volumes used to cache training data with direct S3 connectivity, suitable for distributed training and Jupyter notebooks requiring sub-millisecond latency.
    • Amazon CloudWatch: Used for persistent log management to overcome default Kubernetes log limits and improve monitoring.
    • Amazon Managed Prometheus and Grafana: A scalable managed platform for monitoring Kubeflow and cluster metrics.

  2. Overview of SageMaker Components for Kubeflow Pipelines

    SageMaker Components for Kubeflow Pipelines allow developers to move data processing and training jobs from a local Kubernetes cluster to Amazon SageMaker's machine learning-optimized managed service. This integration enables the creation and monitoring of native SageMaker training, tuning, endpoint deployment, and batch transform jobs directly from the Kubeflow Pipelines UI, while maintaining the orchestration capabilities of Kubeflow.

  3. Overview of Kubeflow on AWS

    Kubeflow on AWS is an open-source distribution of Kubeflow designed to run on Amazon Elastic Kubernetes Service (Amazon EKS). It provides ready-made integrations with AWS services to streamline the building, training, tuning, and deployment of machine learning models for use cases such as computer vision, natural language processing, and financial modeling. The project is maintained in the AWS Labs repository.
    # AWS Labs repository for Kubeflow on AWS manifests
git clone https://github.com/awslabs/kubeflow-manifests

  4. Overview of Kubeflow on AWS

    Kubeflow on AWS is an open-source distribution of Kubeflow designed to streamline the building of machine learning systems on Amazon Elastic Kubernetes Service (Amazon EKS). It provides ready-made integrations with AWS services to reduce operational overhead and improve the reliability, security, and scalability of ML workflows. It supports various ML use cases including computer vision, natural language processing, speech translation, and financial modeling.

  5. Overview of Kubeflow on AWS

    Kubeflow on AWS is an open-source distribution of Kubeflow designed for Amazon Elastic Kubernetes Service (Amazon EKS). It provides ready-made integrations with AWS services to streamline data science tasks and reduce operational overhead when building, training, tuning, and deploying scalable machine learning systems for use cases such as computer vision, natural language processing, and financial modeling.

  6. Overview of Kubeflow on AWS Deployment Add-ons

    Kubeflow on AWS provides several optional add-ons to integrate with AWS services regardless of the deployment option used. Available integrations include:

    • AWS Application Load Balancer (ALB): Used to expose the Kubeflow dashboard to external traffic with secure traffic management.
    • Amazon Elastic File System (EFS): Provides persistent, scalable, and shareable volumes for Kubeflow Notebooks, training, and inference workloads.
    • Amazon FSx for Lustre: High-performance storage used to cache training data with direct S3 connectivity, suitable for distributed training and Jupyter notebooks.
    • Amazon CloudWatch: Used for persistent log management to overcome default Kubernetes log limits and improve monitoring.
    • Amazon Managed Prometheus and Grafana: Provides a scalable, managed platform for monitoring Kubeflow and cluster metrics.

  7. Overview of Kubeflow on AWS

    Kubeflow on AWS is an open-source distribution of the Kubeflow machine learning toolkit specifically optimized for Amazon Elastic Kubernetes Service (Amazon EKS). It provides ready-made AWS service integrations to streamline data science tasks and reduce operational overhead when building, training, tuning, and deploying scalable machine learning systems for use cases such as computer vision, natural language processing, and financial modeling.

  8. Overview of Kubeflow on AWS

    Kubeflow on AWS is an open-source distribution of the Kubeflow machine learning toolkit specifically designed for Amazon Elastic Kubernetes Service (Amazon EKS). It provides ready-made AWS service integrations to streamline data science tasks and reduce operational overhead when building scalable, secure machine learning systems for use cases such as computer vision, natural language processing, and financial modeling.

  9. Overview of Kubeflow on AWS

    Kubeflow on AWS is an open source distribution of Kubeflow designed to help users build machine learning systems using ready-made AWS service integrations. It aims to streamline data science tasks and enable the creation of reliable, secure, and scalable ML systems while reducing operational overhead.

  10. Overview of Kubeflow Profiles and AWS IAM Integration

    A Kubeflow Profile is a user-specific configuration that defines access privileges and ensures consistency across policies, roles, and bindings. Resources for a Profile are isolated within a Profile namespace. To manage permissions for external AWS resources, Kubeflow uses the AwsIamForServiceAccount plugin, which leverages IAM roles for service accounts (IRSA). This allows a Kubernetes service account (SA) to be associated with an AWS IAM role, granting pods in that namespace specific AWS permissions (e.g., S3 access).

  11. Overview of using Amazon RDS and S3 with Kubeflow

    By default, Kubeflow Pipelines (KFP) and Katib use local MySQL pods for data persistence and MinIO for object storage. Replacing these with Amazon RDS and Amazon S3 provides several advantages:

    Amazon RDS (for KFP and Katib):

    • Availability: Supports Multi-AZ deployments for high availability and failover.
    • Scalability: Easier scaling compared to EBS-hosted Persistent Volume Claims.
    • Persistence: Decouples datastores from the Kubeflow deployment, allowing multiple installations to share the same RDS instance.
    • Management: Simplifies database patching, version updates, and instance type changes.

    Amazon S3 (for KFP):

    • Scalability & Durability: Offers higher durability and scalability than local MinIO.
    • Persistent Artifacts: Decouples the artifact store from the deployment, allowing multiple installations to access the same pipeline architectures and run artifacts.
    • Management: Provides advanced tools for data organization and compliance.

  12. Overview of using Amazon RDS and S3 with Kubeflow

    Replacing the default local MySQL and MinIO installations with Amazon RDS and S3 provides several advantages for Kubeflow Pipelines (KFP) and Katib:

    Amazon RDS (Database):

    • Availability: Multi-AZ deployments provide high availability and failover support.
    • Scalability: Supports automatic online resizing and flexible instance types.
    • Persistence: Decouples data from the cluster, allowing multiple Kubeflow installations to share the same datastore.

    Amazon S3 (Artifact Store):

    • Durability: Offers higher scalability and availability than local MinIO.
    • Persistence: KFP artifacts (pipeline architectures and run artifacts) persist beyond the lifecycle of a single cluster installation.
    • Management: Provides advanced AWS tools for data organization and compliance.