Anatomy of an MLA question

Specimen A — multiple choice

An ML team stores 2 TB of Parquet training data in Amazon S3. Before each training cycle, a data scientist needs to run ad-hoc SQL queries to inspect feature distributions. The solution must require the LEAST operational overhead. Which solution meets these requirements?

A. Load the data into a provisioned Amazon Redshift cluster and query it there B. Query the data in place with Amazon Athena C. Launch an Amazon EMR cluster and use Spark SQL D. Import the data into Amazon RDS for PostgreSQL

The dissection

• Sentence 1 is setting: the data is already in S3, already Parquet. Any option that moves it somewhere else is fighting the scenario.
• The ask: ad-hoc SQL — occasional and interactive, not a production workload.
• The qualifier: LEAST operational overhead → serverless beats clusters.
• The kill: A, C, and D all work — that is deliberate. A and C run infrastructure for occasional queries; D moves the data and runs a database. B queries S3 in place with nothing provisioned.

The lesson: distractors are rarely impossible. They are feasible answers that lose on the qualifier — which is why you read it first.

Specimen B — Select TWO

A fraud-detection model runs behind a SageMaker AI real-time endpoint. During traffic spikes, requests time out; latency is fine the rest of the day. The team must handle the spikes while controlling cost. Select TWO actions that together meet the requirements.

A. Configure auto scaling on the endpoint's production variant B. Replace the endpoint with a nightly batch transform job C. Permanently provision enough instances for peak traffic D. Base the scaling policy on an invocations-per-instance metric E. Move inference into an AWS Lambda function with maximum memory

The dissection

• A + D is the set: A turns elasticity on; D makes it track the right signal, so capacity follows the spikes. The two correct answers cooperate — common in Select TWO items.
• B changes the interface: fraud scoring at request time cannot wait for a nightly job. Right service, wrong mode.
• C handles the spikes but fails the second requirement — "controlling cost" — by paying for peak around the clock.
• E relocates the problem: Lambda brings its own ceilings (15-minute timeout, no GPUs) and abandons the model-serving features the team already uses.
• Grading: exact set. A alone scores zero; A + C scores zero. Both picks, exactly right, or nothing.

The 20-second dissection protocol

1. Read the qualifier first

   Find the requirement sentence and underline MOST / LEAST / FASTEST — it is what decides among several feasible options.

2. Place it on the pipeline

   Which stretch of track is this — data prep, model development, deployment, or operations? The domain narrows the service shortlist before you read an option.

3. Kill the options that fight the scenario

   Wrong lifecycle stage, moves data that is already in place, changes the interface (batch for a request-time need) — those die before any comparison.

4. Compare survivors on the qualifier

   Several survivors will genuinely work; the key is the one that satisfies the requirement as qualified.

5. Select TWO (or more)?
6. Complete the set

   Exact-set grading, no partial credit: your sure pick plus the most plausible partner beats a lone certainty — and the correct answers usually cooperate.

7. Answer — never a blank

   No guessing penalty anywhere; flag it and sweep back if time allows.