AWS – Invoking Lambda Directly from S3 vs Going via SQS

If you want to have your Lambda-based microservice triggered both via REST and by SQS messages, then you can do that simply by using both API Gateway and also by publishing the SQS messages to an SNS topic. Lambda functions can be triggered from multiple event sources, see here: http://docs.aws.amazon.com/lambda/latest/dg/invoking-lambda-function.html

Let me know if this doesn't make sense. If you elect to run your own Python runtime on EC2 anyway, and are looking for a good Python framework for running async code, I'd recommend the Twisted Framework (https://twistedmatrix.com/trac/)

I don't think you are going to find any authoritative answers to this, although there are a fair number of articles about.

You should probably anticipate (in the first pass, anyway) that any mutable state in your solution could have more than one writer. So you should anticipate that all mutable writes use some sort of predicate/validator to support a conditional PUT.

Both DynamoDB and S3 support conditional puts, so they are options -- but they aren't necessarily free, in the sense that you'll need to think through your storage strategy, and implement the appropriate semantics on top of them

During the breakout on event sourcing at Re:Invent 2017, DynamoDB was the primary persistence choice in the discussion. After some discussion, the conclusion was that batch writes would not work in a multiple writer scenario - each conditional write would insert a single row.

You also need to think about what reliability guarantees that you need -- is it safe to broadcast events before you store them? Should L1 be reporting success before the persistent changes have been verified?

My best guess at an MVP would be to have your L1 endpoints reading and writing to Dynamo, and then chain other idempotent consumers off the back of that (ie a lambda that reads events out of dynamo and writes them to SNS, Kinesis, etc).

Depending on your guarantees, you might be able to simplify the writers somewhat via L1 -> SNS -> L2 -> Dynamo -> etc.

when you say mutable writes should use a predicate, do you mean I should write my events like "the state has been updated from 16 to 17", and then make L2 validate that the state is currently 16 before updating?

Yes -- think "compare and swap", If-Match, JSON Patch test operation, and so on. If multiple writers need to be enforcing an invariant, then you need some way to ensure that the state being overwritten is the correct one.