Hey I can't find much about reservedConcurrency when you say limit to 1 thread, do lambdas share threads with more concurrency? Also if you limit the concurrency are the extra executions queued up for a bit or just fail? like if I call 100 lambda executions and there is a reservedConcurrency of 1, will it run 1 at a time until all 100 are up?
Greg, please have a look at this doc section "Throttling Behavior"
In short: when you exceed the reservedConcurrency limit, let's say call 100 lambdas while you set the limit to 1, you will experience throttling and the whether lambda will retry your call or not depends on the source of the event, e.g. if you call Lambda from API Gateway you will get the error from your API straight without any retry (you need to handle retries yourself in API called if you wish to)
After we experience throtlling, what happens to the messages that cannot be processed? are they going back to the queue? Does SQS considers them failed?
If your Lambda function is invoked asynchronously, e.g. from other function, AWS Lambda automatically retries the throttled event for up to six hours, with delays between retries.
In case of lambda invoked from sqs (poll-based) AWS Lambda will continue to retry until MessageRetentionPeriod set for the queue expires or if you have redrive policy set on the queue then until MaxReceiveCount after which the message goes to DLQ.
Yep got you, problem is that you cannot have a normal sequential processing by a certain number, say by 1, and fill up the queue. In a similar case, I ended up setting the MaxReceiveCount to 1000 in order to achieve that. Which worked alright but brought other issues, e.g. the message must not fail cause it will just keep retrying for a long time.
I'm not sure I get your problem. Would be nice if you could illustrate it with example project. By sequential you mean processing messages in order I suppose but for this you would need fifo queues or design the app differently, e.g. say you store the current processing state in db. When message2 arrives before message1 you check in db that message1 was not yet processed and you put message1 in another queue and have another lambda which will forward the message back to the original queue (then the receive count is 0 again), when message1 arrives next you process it and update db state so when next time you receive message2 you will know that you can process it already. You could also use the DLQ for that and the other lambda could pass messages from DLQ back to the original queue. Just an idea, don't know your particular case but maybe it works for you.
Ah no I wasn't talking about order but rather the regulation flow. So I want the messages to wait in the queue until the worker can process them but with SQS/Lambda the messages won't wait since they are experiencing throttling and will fail and will either be removed or go to the DLQ.
But I liked you technique on preserving order. I haven't tried it, maybe I will in the future.
Hmm.. I see. So maybe you could try the same technique here. When throttled message goes to DLQ you have a lambda function running in some scheduled interval which you set by experimenting with throttle rate, and this lambda will forward the message from DLQ back to original queue from which it could possibly be processed again and hopefully not throttled again.
I just created this repo and remembered of our discussion. This is how I managed to create a sequential processing pipeline with AWS SQS and Lambda on a Serverless project.
Let me know what you think! Thank you.
Nice :) I like it.
I tried to find out what happens to the CloudWatch event which is throttled, if it is retried or dropped but I could not see in docs, but I suppose it is more a problem of AWS what to do with high number of cumulated CloudWatch events which are continuously retried, rather than your problem, since setting of concurrency limit on Lambda to 1 provides that no more than 1 instance of this function will execute at certain time.
Nice wrap up. Thanks.
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.