Lambda Cold Start Troubleshooting

What is a Lambda cold start?

A cold start occurs when AWS Lambda needs to create a new execution environment to handle a request. This happens when there are no warm containers available - either because the function hasn't been invoked recently, traffic has spiked beyond the current pool of containers, or the function was just deployed.

During a cold start, Lambda performs three steps before your code runs:

1. Downloads your deployment package (or pulls the container image)
2. Starts the runtime (Node.js, Python, Java, Go, etc.)
3. Runs your initialization code (module-level code outside the handler)

The combined time for these steps shows up as Init Duration in CloudWatch logs. This is the latency your users experience on top of the normal function execution time.

Identifying cold starts in CloudWatch logs

Lambda logs cold starts with a specific line format. Look for these patterns:

INIT_START Runtime Version: nodejs20.x  Runtime Version ARN: arn:aws:lambda:us-east-1::runtime:...

And in the REPORT line, cold starts include an extra Init Duration field:

REPORT RequestId: abc-123  Duration: 45.21 ms  Billed Duration: 946 ms  Memory Size: 512 MB  Max Memory Used: 128 MB  Init Duration: 891.33 ms

The Init Duration field only appears on cold start invocations. If the REPORT line doesn't have it, the invocation reused a warm container.

Measuring cold start impact

To understand how cold starts affect your users, you need three numbers from your logs:

• Cold start rate What percentage of invocations have Init Duration? Above 10-15% means cold starts are frequent enough to affect user experience.
• Avg init duration How long is the initialization? Under 200ms is generally acceptable. Over 500ms is noticeable. Over 1 second needs attention.
• P99 impact Cold starts disproportionately affect P99 duration. Compare P99 vs P50 - if the gap is large, cold starts are likely the cause.

Cold start times by runtime

Runtime choice is the single biggest factor in cold start duration. These are typical observed ranges across standard Lambda configurations (128–512 MB memory, non-VPC, package under 10 MB):

Figures are typical observed ranges, not AWS-guaranteed SLAs. Actual durations vary with memory allocation, package size, and initialization complexity. Higher memory allocations get proportionally more CPU, which can reduce init time for CPU-bound startup work.

How smplogs detects cold starts

When you analyze a Lambda log file with smplogs, the WASM engine automatically parses every REPORT line, extracts Init Duration values, and produces a finding like this:

smplogs also correlates cold starts with P99 duration spikes in the root cause analysis, so you can see exactly how much cold starts contribute to tail latency.

Common causes of high cold start rates

How to reduce cold start latency

// Bad - creates a new client every invocation
exports.handler = async (event) => {
  const dynamo = new DynamoDBClient({});
  // ...
};

// Good - client created once per container
const dynamo = new DynamoDBClient({});
exports.handler = async (event) => {
  // reuses the client from above
};

Related guides