Yeah, gotta go with C. SageMaker SSD requires RecordIO for object detection, and image classification (A or D) won't return weed locations, which is needed here. Parquet format (B, D) is a distractor. Pretty sure about this but let me know if I'm off.

Pretty sure it's C for this kind of weed location problem. Object detection like SSD can pinpoint where each weed is in the image, not just the type. RecordIO format is also what SageMaker object detection expects. Not 100% but makes sense with how SageMaker pipelines work.

Going with A here too. Using Lambda for transformation within Firehose avoids managing infrastructure, and AWS handles the scaling. The other choices involve running clusters or EC2, which is more to maintain. Pretty sure this is the least ops work, correct me if I missed something!

Don't think D is right. C locks down access with a VPC endpoint and bucket policy, so privacy/integrity are actually maintained. D's presigned URLs are more for temporary or external access, which is a trap here if you're thinking about security requirements.

I don't think opening up S3 access with ACLs or pre-signed URLs (A, D) really fits the security requirement. Also, B copies the data locally which doesn't actually secure S3 access itself. C's combo of VPC endpoint plus strict bucket policy is best for privacy and integrity. Pretty sure that's what AWS recommends. Open to other takes if anyone thinks differently about endpoint scope though.

C . Private workforce in SageMaker Ground Truth covers "authorized users only" since you control access, and using the built-in bounding box task means no custom labeling app needed. That keeps effort low compared to building your own tools or using Mechanical Turk (which isn't private enough for this kind of medical data). Unless there's some hidden requirement I missed, C fits best here.

I remember a similar scenario from labs and C was always the go-to. Setting up a private workforce in SageMaker Ground Truth with the bounding box task saves tons of setup compared to custom solutions, and it handles access for authorized users. Pretty sure that's what AWS expects here, but happy to hear if anyone has a real-world counterexample?

Option C, Had something like this in a mock, and SageMaker built-in algorithms are definitely optimized for RecordIO protobuf format. Using numpy arrays can slow down training since the built-ins expect RecordIO for efficiency. Not 100% but pretty sure C is the way to go here, agree?

Its C here. The built-in SageMaker algorithms are optimized for RecordIO protobuf, not numpy arrays or even Parquet, which is mostly about storage efficiency. D feels like a trap-hyperparameter optimization only tunes model params, not the raw data format. Pretty sure C, but if I missed something let me know.

I don’t think EMR (B) is right for "least dev effort". Glue (A) offers managed Spark, scales easily, and there’s basically no infrastructure to maintain. EMR’s more hands-on and can be overkill unless deep Spark tuning is needed. Athena and Lambda just don’t fit here. Open to being corrected if anyone has counter examples from recent AWS docs.

Ugh, AWS loves to push Glue for all ETL. I was thinking B at first since EMR gives you more direct Spark control, but the setup's a pain if you want to scale out later or just manage tons of jobs. Wouldn't Athena be easier here?

Option A looks right to me. Step Functions can coordinate the ETL workflow and wait for uploads, then Glue is ideal for joining large datasets in S3. CloudWatch with SNS covers notifications. Pretty sure this matches the scenario, unless I’m missing something.

Option B makes sense here. When features have very different scales, normalization is key since it evens out their influence on the cost function and helps gradients flow better during backprop. Regularization (C) helps with overfitting but doesn't fix scaling problems directly. I'm confident about B, but open if someone has another take.

I don't think it's A. Dimensionality reduction drops features but doesn't solve the scaling issue. I usually see B (data normalization) in practice for this, since it keeps big features from overpowering the others during backprop. C is tempting but more about overfitting. Anyone see it different?

Its B here, but here's a gotcha: if you were using something like tree-based models (say, XGBoost), feature scale actually wouldn't impact convergence much. Since this is about neural nets and backprop, normalization's necessary. Wouldn't pick B for every ML algorithm-depends on optimizer too.

A . Athena supports federated queries with connectors for S3, MySQL, and Redshift, so it can join across all three sources in a single query. B is tempting but Spectrum can’t directly query MySQL. Correct me if I missed something though!

Option A is what I'd pick here. Pretty sure Athena can query directly from S3, and it can also use federated queries to pull data from MySQL and Redshift. Not 100% because I get confused with Glue sometimes, but Athena seems best for actually joining across these sources and running SQL-type analysis. Anyone else agree?

A , had something like this in a mock. Athena supports federated queries across S3, MySQL, and Redshift, so you can join them directly for analysis. Not fully sure if Glue could do it too but Athena lines up with what I've seen.

C. AWS Glue could work if you wanted to join sources and transform data before analysis, right? It's got crawlers and jobs for ETL, so kinda fits when pulling from S3, MySQL, and Redshift. Not totally sure though-maybe that's overkill for just analysis.