MLA-C01 Practice Question 98: High Memory Sporadic Inference
AWS Certified Machine Learning Engineer – Associate (MLA-C01) · Test 10 · question 8 of 10
A travel company's fare-prediction model runs on CPU but needs about 8 GB of memory at inference because it loads large lookup tables alongside the model. It is called a few times an hour at unpredictable times, and some nights not at all. Each caller waits synchronously and needs an answer within 60 seconds, even for the first request after hours of no traffic. The company wants to pay nothing for compute while the model is idle.
Which deployment target meets all of these requirements?
Show answer and explanation
Answer:
Lambda functions can be configured with up to 10,240 MB of memory and packaged as container images of up to 10 GB, enough for the model, the lookup tables, and 8 GB of working memory. Lambda bills only while code runs, so idle hours cost nothing for compute. After an idle period, a cold start has to load the image and tables, which takes seconds rather than the minutes needed to provision an instance, and callers still get a synchronous response (option A).
- B is wrong because serverless inference tops out at 6,144 MB (6 GB) of memory, less than the model needs. It would otherwise fit the traffic pattern. - C is wrong because asynchronous inference queues requests and writes results to Amazon S3 for callers to collect later, rather than answering synchronously. Scaling up from zero instances also takes minutes. - D is wrong because, after scaling in to zero instances, an endpoint cannot respond to requests until it provisions at least one instance again. That takes minutes, so the first request after idle hours would miss the 60-second limit.
Reference: Lambda quotas