Java's popularity contrasts with its low serverless adoption on AWS Lambda due to significant cold start and memory footprint issues.
A new execution environment is created for the first request or during scaling, involving steps like code download and JVM initialization that cause latency.
AWS SnapStart improves Java startup performance by creating a snapshot of an initialized microVM during deployment and restoring it on invocation.
Use the `beforeCheckpoint` CRaC hook to pre-initialize expensive components like HTTP clients and JSON marshallers, moving their setup cost into the deployment phase.
Performance benchmarks show that SnapStart with priming significantly reduces cold starts, while factors like memory allocation and package size also impact latency.
Adjusting Java compilation options, such as using tiered compilation or stopping at level one, and sticking with the default G1 garbage collector can optimize performance.
The AWS CRT HTTP client outperforms the default Apache client for cold starts, and the ARM64 architecture offers a cost-effective alternative to x86.
SnapStart uses a tiered low-latency cache that makes subsequent cold starts progressively faster after the initial invocations, but be aware of its limitations like a 14-day inactivity timeout.
GraalVM Native Image provides the best performance by creating a pre-compiled native executable, resulting in the fastest cold starts and lowest memory footprint.
Adopting GraalVM requires managing its closed-world assumption through reflection configuration and accommodating longer, more resource-intensive build processes in your CI/CD pipeline.
Start with the fully managed AWS SnapStart for significant improvements with minimal effort, and adopt GraalVM Native Image when absolute peak performance is a critical business requirement.
26:54Vadym Kazulkin
37:55Adam Bien
41:04Alex Soto
27:11Kevin Dubois & Daniel Oh
49:30Moritz Kammerer
50:06Jens Happe
48:07Sebastien Stormacq
23:19Nočnica Fee
.gif?w=240&auto=compress,format)
Jobs that call for the skills explored in this talk.
wunschlösung GmbH
Jena, Germany
wunschlösung GmbH
Jena, Germany
doinstruct Software GmbH
Berlin, Germany
