cold-stat

A CLI tool to statically analyze the initialization time, known as cold-start, of AWS Lambda functions.

Prerequisites

• Rust toolchain is required to install the tool
• AWS credential is required to access AWS
• Following actions must be granted in IAM policy
  • lambda:GetFunctionConfiguration to fetch the current env vars of the function
  • Lambda.UpdateFunctionConfiguration to update the env vars of the function, to force cold-start
  • lambda:InvokeFunction to invoke tha function
  • logs:StartQuery to query CloudWatch Logs for cold start analysis

Installation

cargo install cold-stat

Usage

cold-stat [OPTIONS] --function <FUNCTION> --payload <PAYLOAD>

# Example
cold-stat --verbose --function=YOUR-FUNC-NAME --iterations=300 \
          --payload='{"foo": "bar"}'

Options:

• -f, --function <FUNCTION>
  • Name or ARN of function to invoke
• -p, --payload <PAYLOAD>
  • JSON payload to send to the function
• --log-group-name <LOG_GROUP_NAME>
  • Name of CloudWatch log group to analyze
  • [default: /aws/lambda/FUNCTION]
• --log-stream-filter <LOG_STREAM_FILTER>
  • Regex to filter CloudWatch log group stream. Useful when log group is shared by multiple functions
  • [example: /YOUR-FUNCTION-NAME/] when log streams are named like 2021/01/01/YOUR-FUNCTION-NAME[$LATEST]
• -i, --iterations <ITERATIONS>
  • Number of iterations to invoke the function
  • It is recommended to set 30 at least. Because the number of collected cold starts often is a bit shorter than the specified ITERATIONS due to eventual consistency of CloudWatch Logs
  • [default: 100]
• -v, --verbose
  • Print debug logs if enabled
• -h, --help
  • Print help
• -V, --version
  • Print version

Result

• mem
  • Memory size of the function
• count
  • Number of cold starts collected
• stddev, min, max
  • Standard deviation, minimum, and maximum of cold start time respectively
• p50, p90, p95, p99, p995, p999
  • Percentiles of cold start time
  • For example, p50 is 50 percentile, also known as the median
  • p995 and p999 are 99.5 and 99.9 percentiles, respectively

Development

Build

git clone https://github.com/exoego/cold-stat
cd cold-stat
cargo build

Run

cargo run -- \
  --function=YOUR-FUNC-NAME \
  --iterations=10 \
  --verbose \
  --payload='{"foo": "bar"}'

Acknowledgement

• This tool is highly inspired by lumigo-io/SAR-measure-cold-start