Lambda Cold Start Latency

Overview

Overview

A Lambda cold start happens when AWS must provision a new execution environment for your function. Your handler only receives the first event after three phases finish: download, runtime start and your init code.

Cold start by runtime

• Java and Kotlin — Often adds 1–10 seconds — The JVM takes time to start and class-load
• Node.js and Python — 50–500 milliseconds — Fast enough that users rarely notice
• Go — Under 100 milliseconds — Runtime compiled into a single binary, no separate interpreter startup

VPC attachment

> [!WARNING] Verify your function uses the newer Hyperplane model: functions created before 2020 may not.

Tools and what each tells you

• CloudWatch Logs Insights, INIT_DURATION in the REPORT line — Which invocations paid the initialization tax: the field only appears on cold starts
• AWS X-Ray — Initialization as its own segment before the Invocation segment; the gap is your cold start
• Lambda Insights (managed layer) — cold_start metrics in CloudWatch, by function name and version

Fixes, by cost and effectiveness

• Reduce package size — Shortens download: remove dev dependencies, use Lambda layers for shared libraries, prefer tree-shaking bundlers like esbuild — Always worth doing, rarely eliminates the problem
• Avoid unnecessary VPC attachment — Removes network-interface setup — Only where the function does not need the VPC
• SnapStart (Java) — Snapshots the initialized JVM after init code runs and restores from it, bringing Java cold starts under 1 second — Free, but only helps Java
• Provisioned concurrency — Pre-warmed environments that are always ready — Most reliable; charges for idle time, so fits latency-sensitive production, not batch jobs or dev
• Scheduled pings — Keep an environment warm — Unreliable: Lambda may still start fresh environments

> [!IMPORTANT] The core trade-off is cost versus a latency guarantee.

Key concepts

INIT_DURATION

The time Lambda spent initializing the execution environment, printed in the REPORT log line only on cold starts. Includes download, runtime startup, and init code execution.

Provisioned Concurrency

A Lambda configuration that keeps a specified number of execution environments pre-initialized and ready to respond instantly, eliminating cold starts for those environments at an ongoing hourly cost.

SnapStart

A Java-specific Lambda feature that takes a snapshot of the initialized execution environment after init code completes, then restores from that snapshot on cold starts instead of re-running initialization.

Hyperplane ENI

The shared elastic network interface model AWS introduced to fix VPC cold start latency. Hyperplane ENIs are pre-created and reused across function invocations, removing the per-cold-start ENI attachment delay.

Execution Environment Lifecycle

Lambda environments go through Init, Invoke (one or more times), and Shutdown phases. A cold start only occurs during Init; subsequent invocations reuse the same environment and pay only Invoke time.

When to use it

• Diagnosing p99 latency spikes that correlate with traffic lulls
• Explaining Lambda execution model and initialization phases in a system design interview
• Designing a Lambda-based service that must meet strict latency SLAs

Interview questions

• A customer-facing Lambda API has p99 latency of 800ms during normal traffic but spikes to 8 seconds after a 5-minute quiet period. How do you diagnose and fix this?
• Why does enabling VPC for a Lambda function affect cold start times, and how do you determine if the Hyperplane fix applies to your function?

More Troubleshooting topics

• ELB 5xx Errors
• Flaky CI/CD Pipeline Failures
• HPA Not Scaling
• IAM Permission Denied Errors
• JVM Garbage Collection Pauses
• Kubernetes Service Not Reachable
• Log Volume Explosion and Cost Runaway
• Memory Leak Diagnosis
• Metrics Gaps in Dashboards
• Missing Traces in Distributed Tracing
• Multi-Region High Latency
• NAT Gateway Bandwidth and Throttling