From reading AWS documentation understand you can trigger a Lambda from a SNS. If the Lambda fails AWS will retry it twice some time later. You can configure the SNS/Lambda to send the failed messages (after the 3 attempts) to a dead-letter queue. Why/When should I put a SQS between the SNS and the Lambda?
When putting SQS between SNS and Lambda you don't need a dead letter queue (DLQ). In my opinion it depends on what you are doing with events in the DLQ. If those events are treated differently then you can use a DLQ. But if you are feeding them back into the input queue because you want to keep trying to process them until they succeed, just put SQS in front in the first place.
Does that make sense?
my thinking is that if a message failed to be processed 3 times (with delays between retries), must be something wrong with the message and needs to be verified kind of manually or will be going around the SQS forever. Another case can be that an external component (3rd party service, database...) is down but then do we want to immediately put the message back in the SQS and try to reprocess again? Better to move to the DLQ and wait little bit longer.
Thanks for your input.
The message processing might fail because there is something wrong with the message, yes. It might also fail however because a downstream system is not available and the Lambda fails because of this. Then it's a useful retry mechanism.
In the end it's an architecture decision and there are no right solutions, only trade-offs. But you're asking the right questions!
I read decoupling is very important in serverless architectures, only do minimal work in synchronous Lambda calls etc.
And I heard AWS works on SQS Lambda triggers, which will ease a lot of pain in that regard :)
I think putting a queue between S3 and Lambda has yet another benefit.
In case, your lambda failed on processing the event from S3 directly connected to it, the event would be lost after second retry. Putting a queue in-between helps to persist the message for longer time.
An alternative would be to add a DLQ to Lambda itself (docs.aws.amazon.com/lambda/latest/...).
Not sure which one is better.
I totally agree with you. I also find it hard to stay on top of all the different limits and behaviours. How many retries you will get depending on the invocation type and resource... Definitely something to think about in advance.
We're a place where coders share, stay up-to-date and grow their careers.
We strive for transparency and don't collect excess data.