Introduction

This is a tutorial on getting started using AWS Lambda’s SnapStart feature on Python.

References

Limitations

  • Best for when you have relatively large Import and Initialization code
  • Best for when your code doesn’t run frequently
    • Frequently ran function would not need initiation time and it would nullify the benefit of SnapStart
  • Must be attached to version of Function
    • This means all the calling services must reference the version (or Alias) and this must be updated accordingly whenever the Lambda is updated
    • This can be simplifed by using IAC such as Terraform
  • You cannot remove SnapStart setting on existing version, must delete version

Setup

Code Path

  • SNS Topic and Lambda
  • Add a Lambda layer for Pandas and openpyxl (just to add some overhead)
  • Create a version off of $LATEST
  • Assign a “Default” Alias off of latest version
  • Enable SnapStart on “Default” Alias
  • Subscribe both “$LATEST” and “Default” to SNS topic
  • This was deployed using Terraform

Python Code

import boto3
import pandas
import openpyxl
import time
import logging
import os
import json

## Initialization Code here
logger = logging.getLogger()
logger.setLevel(logging.INFO)

def reverse_print(s):
    reversed_str = ""
    for char in s:
        reversed_str = char + reversed_str
        print(reversed_str)

def lambda_handler(event, context):
    for item in event['Records']:
        for key, value in item.items():
            print(key, value)
            if key == 'Sns':
                for k, v in value.items():
                    print(k, v)
                    if k == 'Message':
                        reverse_print(v)
    
    
    return {
        'statusCode': 200,
        'headers': {
            'Content-Type': 'text/html; charset=utf-8'
        },
        'body': '<p>Hello world!</p>'
    }

Testing

  • Published a work of Shakespeare onto SNS topic and recorded duration as shown in CloudWatch
  • Cold Start (odd) followed immedately by a Warm Start (even)
Test SnapStart Restore SnapStart Duration Standard Init Standard Duration SnapStart Total Standard Total
1 431.95 2260.89 2811.36 2154.2 2692.84 4965.56
2 0 2094.38 0 2083.34 2094.38 2083.34
3 343.85 2296.17 2912.7 2034.59 2640.02 4947.29
4 0 2110.37 0 2062.88 2110.37 2062.88
5 606.46 2277 2839.19 2117.31 2883.46 4956.5
6 0 2139.26 0 1929.82 2139.26 1929.82
7 393.19 2019.49 3142.53 2288.43 2412.68 5430.96
8 0 1819.09 0 2115.96 1819.09 2115.96
9 422.25 2220.36 2918.62 2116.05 2642.61 5034.67
10 0 1981.42 0 2118.57 1981.42 2118.57

Conclusion

  • Storing snapshot of the Python initalization can increase the start to about 5 to 9 times regular start time
  • There is virtually no improvement if comparing against warmed function

Cost Consideration

AWS Lambda Snapstart is charged for the snapshot stored (cache) and time it takes to restore the snapshot.

Cache	$0.0000015046 per GB-second (minimum 3 hours)
Restore	$0.0001397998 for every GB restored

Whereas for non-snapstart enabled function, there is no charge for standard initialization. You are only charged for duration period and number of requests.