Skip to content

CONTAINER_IMAGE_CONFIG.md

Latest commit

 

History

History
201 lines (154 loc) · 6.11 KB

CONTAINER_IMAGE_CONFIG.md

File metadata and controls

201 lines (154 loc) · 6.11 KB

Container Image Configuration

Overview

The Lambda function is now configured to use Container Image deployment instead of ZIP deployment. This allows us to include FastEmbed and all dependencies without hitting the 250MB Lambda size limit.

Configuration Details

Lambda Function

  • Package Type: Image (Container)
  • Base Image: public.ecr.aws/lambda/python:3.14
  • Architecture: x86_64
  • Memory: 10GB (10240 MB)
  • Timeout: 900 seconds (15 minutes)
  • Max Image Size: 10GB (vs 250MB for ZIP)

Dependencies Included in Container

pandas>=2.0.0
numpy>=1.24.0
pyarrow>=12.0.0
boto3>=1.28.0
fastembed==0.7.4
onnxruntime
psutil>=5.9.0

Files Structure

customer_analytics/
├── Dockerfile          # Container build instructions
├── app.py             # Lambda handler with FastEmbed
└── requirements.txt   # All dependencies (no layers needed)

Deployment Process

Using Automated Scripts (Recommended)

Step 1: Data Setup

chmod +x setup-data.sh deploy-lambda.sh
./setup-data.sh

The script will:

  1. Create/verify S3 bucket
  2. Generate sample data
  3. Upload data to S3
  4. Display bucket name and file name for next step

Step 2: Lambda Deployment

./deploy-lambda.sh

The script will:

  1. Authenticate with ECR (Docker/Finch)
  2. Verify data exists in S3
  3. Build container image (using Docker/Finch)
  4. Push image to ECR (automatically created)
  5. Deploy Lambda function with container image

Manual Deployment

# Build container image
sam build

# Deploy (will create ECR repository automatically)
sam deploy \
    --stack-name lmi-customer-analytics-with-llm \
    --region us-east-1 \
    --parameter-overrides DataBucketName=<bucket> FileName=<file> \
    --capabilities CAPABILITY_IAM \
    --resolve-s3 \
    --resolve-image-repos

Key Differences from ZIP Deployment

Feature ZIP Deployment Container Image
Max Size 250MB 10GB
Layers Required Not needed
Build Tool sam build --use-container sam build
Dependencies Limited All included
FastEmbed ❌ Too large ✅ Fits easily
Cold Start ~5-10s ~10-20s
ECR Required No Yes

Build Time Expectations

  • First build: 10-15 minutes (downloads base image, installs all dependencies)
  • Subsequent builds: 2-5 minutes (uses cached layers)
  • Image size: ~1.5-2GB (compressed)

Container Build Process

  1. Base Image: Pulls public.ecr.aws/lambda/python:3.14
  2. Install Dependencies: Runs pip install -r requirements.txt
  3. Copy Code: Adds app.py to container
  4. Push to ECR: SAM automatically pushes to Amazon ECR
  5. Deploy: Lambda pulls image from ECR

ECR Repository

SAM will automatically create an ECR repository:

  • Name: lmi-customer-analytics-with-llm-customeranalyticsfunction-<hash>
  • Region: Same as deployment region
  • Lifecycle: Images retained (manual cleanup needed)
  • Authentication: Handled automatically by deploy-lambda.sh script

Cleanup

When deleting the stack, also delete the ECR repository:

# Delete stack
sam delete --stack-name lmi-customer-analytics-with-llm --region us-east-1

# List ECR repositories
aws ecr describe-repositories --region us-east-1

# Delete ECR repository
aws ecr delete-repository \
    --repository-name lmi-customer-analytics-with-llm-customeranalyticsfunction-<hash> \
    --region us-east-1 \
    --force

Troubleshooting

Issue: Docker/Finch not running

Solution: Start Docker Desktop or Finch before running sam build

Issue: ECR authentication failed

Solution: The deploy-lambda.sh script handles this automatically, but if needed manually:

# Get your AWS account ID
AWS_ACCOUNT_ID=$(aws sts get-caller-identity --query Account --output text)

# Authenticate Docker
aws ecr get-login-password --region us-east-1 | \
    docker login --username AWS --password-stdin ${AWS_ACCOUNT_ID}.dkr.ecr.us-east-1.amazonaws.com

# Or authenticate Finch
aws ecr get-login-password --region us-east-1 | \
    finch login --username AWS --password-stdin ${AWS_ACCOUNT_ID}.dkr.ecr.us-east-1.amazonaws.com

Issue: Build takes too long

Solution: First build is slow. Subsequent builds use cached layers and are much faster.

Issue: Image too large

Solution: Container images support up to 10GB, so this shouldn't be an issue with current dependencies.

Advantages of Container Image

✅ No size limits - Include FastEmbed, ONNX, and all dependencies ✅ Consistent environment - Same container locally and in Lambda ✅ No layers needed - All dependencies in one image ✅ Better for ML - Ideal for ML models and large libraries ✅ Version control - Image tags for versioning

Disadvantages

⚠️ Slower cold starts - 10-20s vs 5-10s for ZIP ⚠️ ECR costs - Storage costs for container images ($0.10/GB/month) ⚠️ Build time - Initial build takes longer ⚠️ Requires Docker/Finch - Need container runtime for local builds

Verification

After deployment, verify the container is working:

# Get API endpoint
aws cloudformation describe-stacks \
    --stack-name lmi-customer-analytics-with-llm \
    --query 'Stacks[0].Outputs[?OutputKey==`ApiEndpoint`].OutputValue' \
    --output text

# Test health endpoint
curl <api-endpoint>/health

# Test system info (should show FastEmbed loaded)
curl -X POST <api-endpoint> \
    -H 'Content-Type: application/json' \
    -d '{"requestType":"system_info"}'

The system_info response should show:

  • model_loaded: true
  • model_name: "sentence-transformers/all-MiniLM-L6-v2"
  • embedding_dimension: 384

Summary

Container image deployment enables the full AI-powered analytics with FastEmbed that was originally intended for this project. The trade-off is slightly longer cold starts and ECR storage costs, but you get the complete functionality with semantic search and AI-powered customer segmentation.

The deployment scripts (setup-data.sh and deploy-lambda.sh) handle all the complexity including ECR authentication, making the process straightforward and repeatable.