How To Reduce Cold Starts for Java Serverless Applications in AWS • Vadym Kazulkin • GOTO 2024

Cold starts remain a significant challenge for Java serverless applications, with initial startup times of 2-3.5 seconds without optimizations

AWS SnapStart can reduce cold start times significantly:

• Creates snapshots during deployment
• Restores snapshots during execution
• Free to use with managed Java on AWS Lambda
• Can reduce cold starts to under 1 second

GraalVM native image compilation offers the best cold start performance:

• Produces smaller deployment packages
• Requires additional configuration for reflection
• Build process needs 6-10GB memory
• Can achieve sub-500ms cold starts

Memory allocation impacts cold start performance:

• 1GB is optimal for most cases
• Increasing beyond 1GB shows diminishing returns
• CPU allocation is tied to memory settings

Priming technique further improves cold start times:

• Initialize resources in static blocks
• Pre-warm JSON marshalling
• Initialize AWS clients during deployment
• Can reduce cold starts below 100ms when combined with SnapStart

Best practices for reducing cold starts:

• Use Lambda layers for dependencies
• Remove unnecessary dependencies
• Initialize clients and resources early
• Consider async programming patterns
• Package only what’s needed

Cold start frequency considerations:

• Typically affects ~1% of invocations
• AWS recycles containers periodically
• Cache invalidation can trigger new cold starts
• Multi-region deployments need separate optimization

HTTP client choice impacts performance:

• AWS CRT client performs best for serverless
• Apache client has more features but slower startup
• Default URL connection client is not optimal

Java serverless adoption is growing:

• From 4% to 10% of Lambda functions in past 3 years
• Major frameworks now support serverless deployment
• AWS actively investing in Java serverless tooling

Monitoring and measurement is crucial:

• Test with production-like workloads
• Consider P90/P99 latencies
• Re-measure after Java/AWS updates
• Account for regional differences