Lambda

Introduction

Serverless, virtual functions
Short executions up to 15 minutes
Run on-demand
Pay for number of invocations and compute time
Works with many programming languages
- Node.js, python, Java, c#, Go, Powershell, Ruby, and Custom Runtime API (which can run practically any language)
You can provision up to 10GB of RAM per function

Pay per call:
- First 1,000,000 requests are free
- .20 per 1 million requests after the first million
Pay per duration
- 400,000 GB-seconds of compute time per month for free

When invoking the function from the CLI, SDK, API Gateway, or ALB, the call is synchronous, meaning the result is returned right away
Error handling must happen on the client side (retires, exponential backoff, etc.)

S3, SNS, CloudWatch Events are all processed asynchronously
The events are placed in an internal event queue
The lambda function will read from the event queue and attempt to process the events
Lambda will attempt to retry failures up to 3 times
- This means that event may be processed multiple times, so make sure the lambda function is idempotent
- If the function is retried, you will see duplicate entries in CloudWatch Logs
- You can define a DLQ (dead-letter queue) (SNS or SQS) for failed processing
Async invocations allow you to speed up the processing if you don’t need to wait for the result

Lambda will poll from the sources and be invoked synchronously
- Kinesis Data Streams
- SQS or SQS FIFO
- DynamoDB Streams
Two categories of Event Source Mapping:
- Streams
  - Kinesis or DynamoDB Streams
  - One Lambda invokation per stream shard
  - If you use parallelization, up to 10 batches processed per shard simultaneously
- Queues
  - Poll SQS using Long Polling

By default, Lambda functions are launched outside of your VPC. Therefore, it cannot access resources in your VPC.
Lambda can create an Elastic Network Interface inside your VPC
- You must define the VPC ID, subnets, and security groups
- Lambda requires the AWSLambdaVPCAccessExecutionRole
By default, a Lambda function in your VPC does not have internet access
- Deploying a Lambda function in a public subnet does not give it internet access
- Instead, you can deploy the Lambda function in a private subnet and give it internet access via a NAT Gateway / NAT Instance

Concurrency limit up to 1000 concurrent executions
each invocation over the concurrency limit will respond with a HTTP 429
Cold starts and provisioned concurrency
- If the init is large, cold start could take a long time. This may cause the first request to have high latency than the rest
- To resolve the cold start issue, you can use Provisioned concurrency
  - With Provisioned Concurrency, concurrency is allocated before the function is invoked