Serverless Architectures on AWS in practice - OSCON 2018

From functions to containers to databases, serverless is a huge paradigm shift. The ability to only pay for what we use and not worry about underlying infrastructure is very tempting to developers and DevOps engineers, and the rate of innovation in this area has been very rapid across all major public cloud providers. Serverless architectures are the natural evolution of microservices design. While Lambda has become synonymous with serverless in AWS, there are several new and upcoming patterns that take serverless architectures to the next level.

Manish Pandit explains how to identify these patterns and put them to use. Using Marqeta’s efforts to move its payments infrastructure to the public cloud as an example, Manish explores the services that Marqeta considered, customized, hacked around, and successfully implemented as a part of this move.

Transcript

1. Serverless Architectures on AWS in practice Manish Pandit 07/19/2018
6. If I had asked people what they wanted, they would have said: “faster horses” -Henry Ford
7. The Journey
8. About Me Manish Pandit Director of Platform Engineering @marqeta @lobster1234 lobster1234.github.io
12. Why? Scale technology to support the growing business Use public cloud capabilities so we can focus on product innovation
13. Break down the monolith
14. Shrink the Infrastructure
15. Cloud-Native-First Mindset ..and many more
16. Microservices Independent development, deployment, and scaling of functionally isolated services.
17. Microservices “Loosely coupled, yet highly aligned”
18. Microservices You can have microservices without containers and serverless! Containers and Serverless are the most optimal way to package and deploy microservices.
19. Abstractions Bare Metal O/S Application Fn Serverless Containers VMs Hardware
20. Containers Distributed Applications/Services Economics of Scale - Optimal Resource Utilization Being lightweight, they’re fast to deploy and run
21. Serverless Function as a unit of deployment (FaaS) Pay for what you use, never for idle resources 100% abstraction from underlying resources
22. Serverless Components 1. Functions themselves 2. Events that trigger these functions, or handlers Event driven programming - Functions are event handlers You only pay when the handler runs
23. Lambda Amazon’s FaaS Supports Node, Python, Java 8, C#, Go Billed as GB-seconds
24. Lambda Pricing First 1M requests per month are free 400,000 GB-seconds per month are free START RequestId: 6bd9cf60-7e2b-11e8-9f4a-ebdea81e794d Version: $LATEST END RequestId: 6bd9cf60-7e2b-11e8-9f4a-ebdea81e794d REPORT RequestId: 6bd9cf60-7e2b-11e8-9f4a-ebdea81e794d Duration: 646.26 ms Billed Duration: 700 ms Memory Size: 128 MB Max Memory Used: 31 MB
25. Use Cases
26. Consul Backups GET /v1/catalog/service/mqapi PutObject trigger
27. Database Replication Jobs Get updates PutObject trigger apply
28. Async Events publish publish trigger trigger Service A Service B Service C ECS
29. File Processing Pipeline Split PGP & Transfer Payment Networks Banks trigger trigger N 1 Object + Metadata
30. Lambda limits Memory : 128-3008 MB with 64 MB increments Time : 300 seconds
31. Batch Jobs
32. AWS Batch : Managed Compute https://.. Submitter Job Job Job Batch Cloudwatch Event
33. Containers not Serverless Enough? = ECS minus the Cluster
34. Logging For Lambdas, Cloudwatch logs For ECS, filebeat agent on the host For Fargate, LogStashTcpSocketAppender
35. Monitoring and Alerting
36. Testing Docker Compose Localstack Use environment variables for configuration (12-Factor App)
37. Gotchas.. - Lambda Cold start on both Python and Java - Due to VPC and ENI - Be aware of the retries - Async invocation gets two retries - Sync invocation gets no retries - Apply the Principle of Least Privilege to the Lambda and ECS Roles - S3 triggers - source and destination buckets
39. Tips - Declare connection pools outside of the handler’s scope - Write tests! - Test locally - the only error you should be expecting is IAM Permissions error. <role> is not authorized to perform: <operation> on resource: <resource> - Do not create Microliths - Set up billing alerts
40. Tips - Pick the best tool for the job - Do not pick the tool and try to fit the job into it - Beware of the shiny object syndrome - Take measured risks - Production is (mostly) unforgiving
42. “You’re free to choose, but you’re not free from the consequences of your choice” - Universal Paradox
43. Fin Manish Pandit Director of Platform Engineering @marqeta @lobster1234 lobster1234.github.io