crunr

Product Introduction

1. Definition: crunr is a free and open-source command-line interface (CLI) tool designed for on-demand, serverless-style compute orchestration on a user's own AWS (Amazon Web Services) infrastructure. It falls under the technical categories of DevOps automation, cloud cost optimization, and machine learning operations (MLOps).
2. Core Value Proposition: crunr exists to eliminate the hidden costs and operational overhead of running GPU and high-compute workloads in the cloud. Its primary value is delivering pure, pay-per-second compute by automatically provisioning and terminating cloud instances, thereby solving the critical problems of idle time billing, forgotten instances, and DevOps complexity.

Main Features

1. Zero-Idle, On-Demand Instance Lifecycle: crunr's core function is to manage the complete lifecycle of an AWS EC2 instance. Upon running crunr run [script], it automatically selects and launches the cheapest suitable spot instance, executes the user's script, streams the output back to the local terminal, and then immediately terminates the instance upon job completion or failure. This ensures the user's AWS bill stops the moment compute ends.
2. Local-First, Zero-Backend Architecture: crunr operates with a strict local-first philosophy. It has no servers, no dashboard, and no persistent backend infrastructure. The CLI binary communicates directly with AWS APIs. User AWS credentials are stored and used locally from the standard ~/.aws/credentials file and are never transmitted to any crunr-controlled service, ensuring maximum security and privacy.
3. Polyglot Script Execution & Dependency Management: crunr supports running scripts in multiple languages including Python, Node.js, bash, R, and Go without requiring cloud-specific rewrites or SDKs. It automatically syncs the local code directory to the launched instance via rsync and handles dependency installation (e.g., via pip, apt, npm) based on the project's standard files before executing the user's command.
4. Automatic Cost-Optimized Instance Selection: The tool polls AWS pricing in real-time to select the most cost-effective spot instance type that matches the user's specified requirements (e.g., --gpu). It abstracts away the complexity of region and availability zone selection, instance type comparison, and spot market bidding.
5. Secure Output Synchronization & Automatic Cleanup: Upon job completion, crunr securely syncs output files and directories (e.g., outputs/) back to the user's local machine. It provides the crunr clean command to manually terminate any running jobs. All instances use AWS IAM roles for permissions, meaning the user's personal AWS access keys are never exposed on the remote machines.

Problems Solved

1. Pain Point: Uncontrolled Cloud Spend from Idle Resources. The primary problem crunr solves is the "idle time tax," where users pay for expensive GPU instances (like AWS g5 or p4d) that run unused between jobs, during debugging, or after being forgotten, leading to bills hundreds or thousands of dollars higher than expected.
2. Pain Point: DevOps Overhead for Episodic Compute. It removes the need for users—especially researchers, data scientists, and small teams—to become experts in AWS EC2, VPC, security groups, IAM roles, and AMI management just to run a batch job or training script, saving days lost to failed setups and configuration.
3. Target Audience: The product is built for ML/AI engineers and researchers, indie AI builders, data scientists running heavy ETL or batch jobs, startup engineers without dedicated DevOps, students, and anyone currently renting GPUs from platforms like Lambda Labs or RunPod who face similar idle billing problems.
4. Use Cases:
   • Machine Learning Training/Fine-Tuning: Running a train.py script on a specific GPU (A100, g5.xlarge) for exactly the duration needed.
   • Batch Data Processing: Executing a heavy data transformation or model inference job on a high-memory instance without managing a server.
   • Academic Research & Experimentation: Allowing students and researchers to run reproducible experiments on powerful hardware, paying only for the compute minutes used.
   • CI/CD for Compute-Intensive Tasks: Automating one-off build or test jobs that require specialized hardware.

Unique Advantages

1. Differentiation from Traditional Cloud Consoles & CLI: Unlike manual AWS Console or AWS CLI usage, crunr fully automates the tedious setup, execution, and cleanup workflow into a single command, guaranteeing termination and cost control—a step often missed manually.
2. Differentiation from Managed MLOps Platforms: Compared to full-featured (and expensive) MLOps platforms, crunr is minimalist, open-source, and uses the user's own AWS account. There is no monthly platform fee, no vendor lock-in, and no data passes through a third-party service.
3. Differentiation from Other "Serverless" GPU Services: While some services offer serverless GPU containers, they often have cold starts, limited customizability, and their own markup. crunr provides direct access to the full spectrum of AWS EC2 instances at standard AWS spot rates, with no crunr markup.
4. Key Innovation: The "Ghosting" Guarantee: The fundamental innovation is the architectural guarantee of termination. The system is designed with no persistent components, forcing a $0 cost state when idle. This shifts the cloud cost model from "always-on reserve" to "truly on-demand," matching the mental model of researchers and developers running local scripts.

Frequently Asked Questions (FAQ)

1. How does crunr save money on AWS GPU costs? crunr saves money by exclusively using AWS spot instances (the cheapest pricing model) and, most critically, by automatically terminating the instance the millisecond your job finishes or fails. This eliminates all costs from idle time, which is the largest contributor to unexpected cloud bills for episodic compute workloads.
2. Is crunr secure for my AWS credentials and data? Yes, crunr is designed with a security-first, local-only approach. Your AWS credentials never leave your machine, stored only in the standard ~/.aws/credentials file. The launched EC2 instances use an IAM role for permissions, so your secret key is not present on them. All code and data transfer occurs via direct rsync/ssh between your laptop and your instance.
3. What happens if my crunr job crashes or my laptop disconnects? crunr is built for resilience. If your script crashes with a non-zero exit code, the remote instance will capture this and terminate itself automatically, stopping the bill. If your local laptop disconnects from the internet, the remote job continues to run on AWS until completion, after which the instance will still self-terminate as designed.
4. Can I use crunr with existing cloud setups or VPC configurations? Currently, crunr is optimized for simplicity and defaults. It manages its own security groups and network settings to "just work." Advanced users who require specific VPCs, subnets, or security group rules may find its configuration options limited compared to manual infrastructure-as-code tools.
5. How is crunr different from using AWS Batch or SageMaker Training Jobs? AWS Batch and SageMaker are fully managed, heavyweight services with more features but also more complexity and potential for higher cost (e.g., managed storage, API costs). crunr is a lightweight, open-source alternative that gives you direct, bare-metal EC2 access with zero management overhead and no extra fees, ideal for users who want maximum control and minimum cost for straightforward script execution.