SNS to Lambda vs SNS to SQS to Lambda

Question

I'm trying to understand whether I need SQS in my workflow if someone can help explain. In my app, when an action is taken, it submits info to SNS topic which invokes Lambda to do some processing. This is working great as it is.

When I do research online, it seems that people are using SQS in this stack as well where SNS would put info on SQS and then SQS would then invoke Lambda.

I guess what I'm trying to understand is the need for SQS in this. What value does that add? In other words, what am I losing by invoking my Lambda directly from SNS?

Solution 1

Primary advantage of having a SQS in between SNS and Lambda is Reprocessing. Assume that the Lambda fails to process certain event for some reason (e.g. timeout or lack of memory footprint), you can increase the timeout (to max 5 minutes) or memory (to max of 1.5GB) and restart your polling and you can reprocess the older events.

This would not be possible in case of SNS to Lambda, wherein if Lambda fails the event is lost. And even if you configure DLQ you would still have to make provisions for reading that separately and processing the message

So if your events are critical and you don't want to miss out on them, then go for SNS - SQS - Lambda

The other advantage of having SQS is cost saving on Lambda invocations (Thanks @codesinthedark for bringing this up). You can have much better scaling and less cost, as it allows you to process messages in batches. So one lambda can be executed for a batch of 10 messages while in case of direct SNS each message would trigger a lambda invocation.

Solution 2

I think couple of things changed in 2019 and SQS can trigger lambda via event source mapping which is mentioned by @alexs. Related blog post: https://aws.amazon.com/about-aws/whats-new/2018/04/aws-lambda-now-supports-amazon-sqs-as-event-source/

To summarize, you can use SQS to lambda with following benefits:

• Reprocessing events in case of failure and configure how many times a message should be retried before you give up (receive count)
• Longer retention period
• Typically chosen in the scenarios where there is a long running job and the lambda polls one by one from job queue.

you can choose to use SNS:

• If you need to fanout a single message to multiple destinations, say X message should be processed by Y and Z applications. I feel this is the biggest advantage, and if you want reliability in this, you can couple SNS and SQS together.
• You do not care about lost messages. Remember that there are still retry stratergies when using SNS(linear, geometric, exponential etc)
• Typically used in the cases where you can ingest/process messages faster. This can sometimes be a problem as well; imagine a scenario where there is a SNS notification for every email that your business receives and you dont have enough lambda concurrency to process all of them. You can solve this by putting an SQS to consume at your own pace.

In both the cases, there can be duplicate messages(in the cases of retry) and there cannot be order guarantees. If you need one, consider Kinesis streams.

Solution 3

You can now use SQS as en event source

AWS Lambda Adds Amazon Simple Queue Service to Supported Event Sources

Solution 4

Adding to @Arafat Nalkhande's answer here are few benefits of SQS's lambda

1. In SQS we can put a delay, so that message gets processed after some time, it may be useful in the scenario where data takes time to be available.

2. SQS can serve as a contingency store, lets say downstream services are unavailable, message can be retained in sqs for 15 days.

View more solutions

Comments

[dmulter]

Repeating as a comment here instead of as a different answer below... AWS Lambda Adds Amazon Simple Queue Service to Supported Event Sources

[Arafat Nalkhande]

Right, @dmulter, SQS triggering Lambda support was added on 28 JUN 2018. However, the question mentions that "In my app, when an action is taken, it submits info to SNS topic which invokes LAMBDA to do some processing", so this new feature would still not answer the question.

[user3311298]

Any Lambda function invoked asynchronously is retried twice and then can be sent to a DLQ which can be a SNS or SQS docs.aws.amazon.com/lambda/latest/dg/dlq.html So the reason given in this answer as to why add SQS between SNS and lambda does not make sense

[Arafat Nalkhande]

I do not agree that this post does not answer the question, because assume that you have configured a DLQ instead of going the direct route of adopting SQS, then for all the failures you will have to go to the DLQ. If for some reason the failures were simply timeouts or due to insufficient memory footprints then I'll have to make separate provisions to read the DLQ and process the messages, whereas if the SQS was used in between, then I simply have to reconfigure the lambda with larger memory footprint and rest if taken care off automatically

[Arafat Nalkhande]

Thanks for pointing out the DLQ. I've made slight adjustments to the answer and added that part

[lfk]

This is assuming that SQS never goes down, otherwise events can still get lost. But I guess that's a reasonably safe assumption?

[Jacek Kościesza]

Things changed again in 2020 with SNS FIFO - providing strict message ordering and deduplicated message delivery: aws.amazon.com/blogs/aws/…

[Payam]

Lambda doesn't "fail" easily and "lose" the data if there is no external component such as http/s microservice, email, or sms is involved. More info here: aws.amazon.com/lambda/faqs/#Scalability_and_availability

[Arafat Nalkhande]

@Payam, The link that you posted says "On failure, Lambda functions being invoked synchronously will respond with an exception. Lambda functions being invoked asynchronously are retried at least 3 times"... What if it fails even after 3 times

[Payam]

@ArafatNalkhande, If the process needs high reliability, then the Lambda should forward the results to another AWS service. For example, we use SNS->Lambda->Elasticache/Redis a lot. Amazon complies with Six-Sigma (99.9999% Availability), and you can increase the number of retires or eventually forward the failed processes to a DLQ SQS

[Arafat Nalkhande]

Lets say I forward that to Redis I would then need to add some compute tier to process it from Redis and I would assume the SNS to SQS stuff will serve the same purpose and would be much simpler approach

[CodesInTheDark]

I think that you forgot one huge benefit of using lambda with SQS instead of SNS. You can have much better scaling and less cost, as it allows you to process messages in batches so one lambda can be executed for 10 messages while each SNS message would trigger lambda and it would fail or throttle for over 1000msg/s. You would pay one lambda execution and one SQS request instead of 10 requests for SNS. You can update your answer.

[Scott Simontis]

The retry policy for SNS only applies to situations where the message is not received because the consumer's service is unavailable. If your Lambda fails to process the message correctly, it can be configured to send its message to a DLQ (after 2 retries by default), but it will never be sent back to SNS for re-transmission.