Data-driven & cost-aware performance tuning for AWS Lambda

How can we combine cost optimization and performance tuning into a single data-driven process, and even integrate it into our CI/CD pipelines? This talk will showcase an open-source tool to optimize Lambda functions resource allocation without requiring complex refactors or architectural changes, while showcasing charts and cost/performance patterns of real-world use cases.

Transcript

1. Alex Casalboni
   Sr. Developer Advocate, AWS
   @alex_casalboni
   Data-driven & cost-aware
   performance tuning for AWS Lambda
   @ 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved
   

   View Slide

2. About me
   • Software Engineer & Web Developer
   • Data science background
   • Worked in a US startup for 4.5 years
   • ServerlessDays Milan co-organizer
   • AWS customer since 2013
   

   View Slide

3. Fundamentals (very quickly!)
   Optimization best practices
   Idea: data-driven approach
   Real-world examples & patterns
   Agenda
   

   View Slide

4. Fundamentals
   

   View Slide

5. View Slide

6. AWS operational responsibility models
   On-premises Cloud
   Less More
   Compute Virtual machine
   Amazon EC2 AWS Elastic Beanstalk AWS Lambda
   AWS Fargate
   Databases MySQL
   MySQL on EC2 Amazon RDS MySQL Amazon Aurora Aurora Serverless Amazon QLDB/DynamoDB
   Storage Storage
   Amazon S3
   Messaging ESBs
   Amazon MQ Amazon Kinesis Amazon EventBridge/SNS/SQS
   Analytics
   Hadoop Hadoop on EC2 Amazon EMR Amazon Elasticsearch Service Amazon Athena
   

   View Slide

7. Anatomy of a Lambda function
   Handler() function
   Function to be executed
   upon invocation
   Event object
   Data sent during Lambda
   function Invocation
   Context object
   Methods available to
   interact with runtime
   information (request ID,
   log group, more)
   import json
   def lambda_handler(event, context):
   # TODO implement
   return {
   'statusCode': 200,
   'body': json.dumps('Hello World!')
   }
   

   View Slide

8. The request lifecycle
   Bootstrap
   the runtime
   Start your
   code
   Cold
   start
   Warm
   start
   Download
   your code
   Start new
   micro vm
   AWS optimization Your optimization
   

   View Slide

9. Lambda + VPC =
   ß Before: 14.8 sec duration
   

   View Slide

10. Lambda + VPC =
    ß Before: 14.8 sec duration
    After: 933ms duration à
    

    View Slide

11. Optimization best practices
    

    View Slide

12. Avoid «monolithic» functions
    Reduce deployment package size
    Micro/Nano services
    Optimize dependencies (and imports)
    E.g. up to 120ms faster with Node.js SDK
    Optimization best practices (cold start)
    Minify/uglify production code
    Browserify/Webpack
    Lazy initialization of shared libs/objs
    Helps if multiple functions per file
    

    View Slide

13. Provisioned Concurrency for AWS Lambda
    New!
    Managed solution to avoid “cold starts”
    You set provisioned concurrency for individual functions
    No code changes required
    Integrated with AWS Auto Scaling
    Provisioning rampup (500/min)
    Wide support (CloudFormation, Terraform, Serverless
    Framework, Datadog, Epsagon, Lumigo, Thundra, etc.)
    

    View Slide

14. Externalize orchestration
    Avoid idle/sleep – delegate to Step Functions
    Transform, not Transport
    Minimize data transfer (S3 Select, advanced
    filtering, etc.)
    Discard uninteresting events asap
    Trigger config (S3 prefix, SNS filter, etc.)
    Fine-tune resources allocation
    Don’t guesstimate function memory
    Lambda Destinations
    Simplified chaining (async) and DLQ
    Keep in mind retry policies
    Very powerful (and configurable), but not free
    Optimization best practices (performance/cost)
    

    View Slide

15. Externalize orchestration
    Avoid idle/sleep – delegate to Step Functions
    Transform, not Transport
    Minimize data transfer (S3 Select, advanced
    filtering, etc.)
    Discard uninteresting events asap
    Trigger config (S3 prefix, SNS filter, etc.)
    Fine-tune resources allocation
    Don’t guesstimate function memory
    Lambda Destinations
    Simplified chaining (async) and DLQ
    Keep in mind retry policies
    Very powerful (and configurable), but not free
    Optimization best practices (performance/cost)
    

    View Slide

16. Resources allocation
    Memory Power
    

    View Slide

17. CPU-bound example
    “Compute 1,000 times all prime numbers <= 1M”
    128 MB 11.722 sec $0.024628
    256 MB 6.678 sec $0.028035
    512 MB 3.194 sec $0.026830
    1024 MB 1.465 sec $0.024638
    

    View Slide

18. CPU-bound example
    “Compute 1,000 times all prime numbers <= 1M”
    128 MB 11.722 sec $0.024628
    256 MB 6.678 sec $0.028035
    512 MB 3.194 sec $0.026830
    1024 MB 1.465 sec $0.024638
    

    View Slide

19. CPU-bound example
    

    View Slide

20. Cost-aware performance optimization
    310ms 400ms
    294ms 300ms
    5% performance optimization
    25% cost optimization
    480ms 500ms
    408ms 500ms
    15% performance optimization
    0% cost optimization
    A B
    

    View Slide

21. Idea: data-driven approach
    

    View Slide

22. Data-driven cost & performance
    optimization for AWS Lambda
    Available as a SAR app
    Easy to integrate with CI/CD
    Meet AWS Lambda Power Tuning
    github.com/alexcasalboni/aws-lambda-power-tuning
    

    View Slide

23. github.com/alexcasalboni/aws-lambda-power-tuning
    v1 v2 v3
    

    View Slide

24. AWS Lambda Power Tuning (input)
    github.com/alexcasalboni/aws-lambda-power-tuning
    

    View Slide

25. AWS Lambda Power Tuning (output)
    github.com/alexcasalboni/aws-lambda-power-tuning
    

    View Slide

26. AWS Lambda Power Tuning (visualization)
    github.com/alexcasalboni/aws-lambda-power-tuning
    

    View Slide

27. Real-world examples
    and patterns
    

    View Slide

28. No-Op (trivial data manipulation <100ms)
    

    View Slide

29. CPU-bound (numpy: inverting 1500x1500 matrix)
    

    View Slide

30. CPU-bound (prime numbers)
    

    View Slide

31. CPU-bound (prime numbers – more granularity)
    

    View Slide

32. Network-bound (third-party API call)
    

    View Slide

33. Network-bound (3x DDB queries)
    

    View Slide

34. Network-bound (S3 download – 150MB)
    

    View Slide

35. Network-bound (S3 download multithread – 150MB)
    

    View Slide

36. Cost/Performance patterns
    github.com/alexcasalboni/aws-lambda-power-tuning
    

    View Slide

37. Takeaways
    

    View Slide

38. Takeaways
    Optimization best practices (what are you trying to optimize?)
    Avoid cold starts with Provisioned Concurrency
    Memory Power
    Optimal resources allocation can be automated (CI/CD)
    Think in terms of workload categories and cost/performance patterns
    Visualize optimal trade-offs with /alexcasalboni/aws-lambda-power-tuning
    

    View Slide

39. Go Build!
    Here to help you build
    

    View Slide

40. Alex Casalboni
    Sr. Developer Advocate, AWS
    @alex_casalboni
    @ 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved
    Thank you!
    

    View Slide