Exam AWS Certified Solutions Architect - Associate SAA-C02 topic 1 question 280 discussion

C. AWS just update Lambda to support 10G memory and helping compute intensive applications like machine learning... No disk access, lowest cost Posted On: Dec 1, 2020 https://aws.amazon.com/about-aws/whats-new/2020/12/aws-lambda-supports-10gb-memory-6-vcpu-cores-lambda-functions/#:~:text=Customer%20Enablement-,AWS%20Lambda%20now%20supports%20up%20to%2010%20GB%20of%20memory,vCPU%20cores%20for%20Lambda%20Functions&text=AWS%20Lambda%20customers%20can%20now,previous%20limit%20of%203%2C008%20MB. P/s: ELB has limit of 100 target groups

D To meet the requirements (such as "Most of the models are used sporadically") and to be "MOST cost-effective

A is the answer

Not sure why the answer is C, Labda is just 15min and we are not sure how much time takes for model. I will go for D option

Path based routing is for ALB, not API gateway. https://aws.amazon.com/blogs/containers/using-aws-application-load-balancer-path-based-routing-to-combine-amazon-ecs-launch-types/

Anyone who passed the exam, let us know what they chose?

Answer is D as ECS is cheaper than Lambda

C https://aws.amazon.com/es/about-aws/whats-new/2020/12/aws-lambda-supports-10gb-memory-6-vcpu-cores-lambda-functions/

I thought D at first. But C for the win. Anyone noticed the key words "... the majority of models are used seldom ..." - Lambda saves costs here compared to running EC2 instances.

C - correct (better fit to the entire demand than A); it also tackles no disk requirements, scalability & latency. B or D can not be right as for ECS you need EC2 instances (and allocated disk storage) and it really does not even mention Fargate in order to make sense. Also each EC2 instances take load time and cost allocated resources more than the variable requirements mentioned in the demand. (PS: Do not forget that API GW also has a free tier of 1M runs/month for first 12m)

IMO: C has lower cost with pas as you go: https://aws.amazon.com/blogs/compute/pay-as-you-go-machine-learning-inference-with-aws-lambda/; ECS has underlying instance costs: https://aws.amazon.com/blogs/containers/using-aws-application-load-balancer-path-based-routing-to-combine-amazon-ecs-launch-types/

D seems correct because ALB supports path based routing. API gateway creates stages for method request, integration request, method response, integration response. API gateway enables caching and set throttling limits.

definitely not C. Because every time Lambda function runs it needs to retrieve the model from S3 which causes delay. Lambda is just a runtime env you can not save anything before running it or after it is run.

another anoying question, who writes these things? all answers seem to meet the requirements so the focus is on "MOST cost-effective" (a) and (c) are the same with the exception of (a) using more API Gateways and so being more expensive. so it can't be (a) (b) and (d) are also the same with the exception of (b) using more ALBs and so being more expensive, so it can't be (b) between (c) and (d) is a tough one, lambda vs ecs. 0GB of data into RAM and doesn't need disk access suggests (c) but robertomartinez comment below calculates ECS is cheaper than lambda for similar compute power, so (d) but (d) uses a single ALB and they are limited to 100 target group instances and the question says "hosts hundreds of machine learning models" so (c) but "hundreds of machine learning models" doesn't specify if they are all different learning models or not (you could have 200 learning models run but they are only of 5 types, so you'd only want 5 target groups), so (d) and for (c) can you do serious machine learning if a lambda function can only run for 15 minutes at a time? to conclude, i have no idea

Not an expert but Lambda needs to load 10G worth of data and max run time for Lambda is 15min. Doesn't sound right. Maybe D.

In my opinion answer is D, based on what follows. Solution is between C and D. Question says "during startup each model requires to load 10GB of data from S3 into memory". Can you do that with lambda functions (C) ?. I think no. You would require each called function read 10GB from S3 before processing the request. Using containers this is possible, that's why I go with D. (please clarifications about if lambda functions can do that or not are wellcome)

check this link. It says that AWS have a way to reduce cost using lambda for machine learning models with sporadic inference requests throughout the day. https://aws.amazon.com/blogs/compute/pay-as-you-go-machine-learning-inference-with-aws-lambda/ Ans C