Data Preparation for Machine Learning (ML)

MLA-C01 · 84 questions

• A municipal transfer station lands 80 daily weigh-ticket columns on Amazon S3. Analysts later scan only five columns with Amazon Athena before the set is handed to Amazon SageMaker. Which file layout should they choose?
• A municipal water-taxi kiosk emits nested JSON boarding events that operations staff still open in a text editor. Next quarter the same objects will be scanned as flat columns for a no-show model. When should the team keep JSON, and when should they convert?
• A municipal grain-silo co-op already runs Apache Hive-style tables on Amazon EMR and wants a columnar, type-aware file layout that those jobs already speak. Which format fits that access pattern?
• A community shortwave logger keeps adding fields to show-log records while older files must still be readable. Writers need a compact row format that carries the schema with the data. Which format supports that evolution better than CSV?
• A municipal curling-club attendance team wants to feed an Amazon SageMaker built-in algorithm that expects the RecordIO-wrapped layout rather than raw CSV lines. Which ingest convention should they apply?
• A municipal arboretum sensor desk can either enforce a declared schema as files land or drop whatever bytes arrive and discover columns later. Which pair names those two ingest paths?
• A county plat office dumps every parcel photo and roll row under one Amazon S3 prefix. Training jobs time out listing objects. Which change matches the access pattern without buying a bigger instance?
• A historic sawmill finishes each night’s kiln logs and needs cheap, durable object storage that AWS Glue, SageMaker Data Wrangler, and later training jobs can all read. Where should that batch lake land?
• Three Amazon SageMaker training instances must open the same directory tree of bat-detector spectrogram files with ordinary file-path reads, not object GETs. Which store should they share?
• A county fry-hatchery’s existing on-premises NetApp share already holds labeled fry photos, and the ML team wants those same NFS semantics in AWS without rewriting the read path. Which store should they choose?
• A municipal rail-yard vibration set is tens of tebibytes of sequential files that training jobs will stream repeatedly. Object GET overhead dominates. Which initial storage decision fits that I/O pattern?
• A municipal trolley farebox database lives in Amazon RDS. The ML engineer needs those tables in the S3 lake without a hand-rolled nightly dump script if a managed extract exists. Which approach is correct?
• A community archive-card tap table in Amazon DynamoDB is hot during opening hours. The team needs a point-in-time extract for a hold-prediction model. Which extract path should they prefer?
• A municipal materials-recovery workstation stages local shards before upload; the default general-purpose EBS volume stalls under random reads. Which lever matches that bottleneck?
• A community hop-yard field office in a far Region uploads harvest CSVs into a lake bucket in another Region and watches WAN transfers crawl. Which Domain 1 extract option speeds those long-distance uploads?
• A municipal water-plant still writes lab CSVs to an on-premises NFS export and needs a scheduled, checksummed copy into Amazon S3 for later SageMaker Data Wrangler work. Which in-scope service should they use?
• Municipal tug-dispatch radios emit position pings that must keep per-vessel order and be available to multiple consumers within seconds. Nightly S3 drops miss the freshness bar. Which ingest service fits?
• A municipal sled-hill grooming desk wants click-and-sensor records to land as objects in Amazon S3 every few minutes without the team managing stream shards or consumer applications. Which service fits?
• A municipal aquarium gate needs five-minute rolling visitor counts with state, late events, and windows before those aggregates land for a staffing model. Which ingest path should they use?
• A community credit-union core already publishes ledger-change messages to Apache Kafka and will not rewrite producers this quarter. The ML lake must subscribe to that protocol. What should the team keep as the streaming source?
• A community hoop-house climate model must see new sensor rows within ten minutes, but today’s pipeline is a 2 a.m. Amazon S3 drop. Which ingest change meets that freshness bar?
• Several teams must read the same Amazon S3 lake tables for feature work, and the data engineer wants a catalog-level permission model rather than a thicket of raw bucket policies. Which Domain 1 service pair fits?
• Orchard-harvest CSVs sit in Amazon S3, but SageMaker Data Wrangler and Amazon Athena show no tables. Which missing step should the team take?
• A county harvest-fair needs ticket scans from Amazon S3, vendor payments from Amazon RDS, and volunteer shifts from Amazon DynamoDB as one training table. Which merge path matches the official skill?
• After maple-sugar CSVs reach Amazon S3, a small ML team wants a visual flow to profile and start transforms without standing up Amazon EMR first. Which ingest-into-prep step should they take?
• A municipal canal-ferry operator wants reused numeric features written once and read later by both training and, later, serving. Which Domain 1 ingest target should they load?
• Cold kiln archives are rarely reread; hot daily logs are scanned every training run; a third pile is a sparse wide table. How should the team map those piles, and what must they not do?
• An AWS Glue job and later training listings stall because a kiosk writes millions of 4 KB objects. Which ingest fix matches that scale issue?
• A municipal grain-scale dataset has a few tickets recorded in pounds instead of tons, which pull a linear scaler sideways. What should the team do before scaling or training?
• County forage-plot moisture probes drop overnight readings. Dropping every incomplete row would erase frost nights. Which missing-value strategy is appropriate?
• Community festival volunteer logs leave the role field blank on weekends. Filling those blanks with the weekday mode would invent a role that never happens on Saturday. Which missing-categorical treatment is appropriate?
• A municipal water-taxi farebox extract double-wrote the same tap when a job retried. Training row counts jumped but unique riders did not. Which cleaning step should run before feature builds?
• Combining night-market ticket scans with a many-row vendor catalog on a loose key multiplies rows and fabricates extra purchases. How should the team combine those sources?
• A 12-column berry-harvest table needs histograms, a few imputes, and an exported flow. Standing up Amazon EMR Spark is overkill. Which tool matches that modest SageMaker-bound set?
• A county permit clerk who does not live in SageMaker Studio still needs a visual recipe to clean permit CSVs in the lake. Which official no-code prep tool fits that path?
• Nightly merges of three 200-million-row sources need a repeatable Spark job, job bookmarks, and a catalog update. Which service should run that transform?
• A historic lime-works transform needs a specific Spark runtime and native libraries that the team already packages on Amazon EMR, not the Glue-only path. Which compute should they choose?
• Each trail-register ping needs a unit conversion and a field drop before it lands. The payload is tiny and stateless. Which official lightweight transform should run on the stream?
• Gondola boarding events need sessionization across minutes of out-of-order records before features are written. A single AWS Lambda timeout is the wrong ceiling. Which transform path fits?
• Aquarium gate-wait minutes are roughly bell-shaped; another column is a hard 0–100 saturation percent. How should those two numeric columns be scaled?
• Municipal bulk-trash surcharge dollars run from pocket change to a few extreme commercial hauls. A linear model treats the tail as the whole story. Which transform should they apply to that positive amount first?
• Water-taxi rider age is messy at the edges, but planning only needs child, adult, and senior bands. Which feature step matches that need?
• A single gondola departureat string hides hour-of-day and weekday that a no-show model needs. Which feature-engineering technique should they apply?
• Municipal tug vessel class has four unordered labels. Feeding those strings into a linear trainer is wrong. Which encoding should they apply?
• Forage-plot soil-grade is poor, fair, good, excellent. Integer ranks preserve that order. Using the same integers on unordered berth names would invent a false rank. When is label encoding correct?
• Municipal trolley stop-ID has hundreds of values; one-hot would explode width. Which official alternative encoding should they choose?
• Festival volunteer incident notes need to become tokens before any bag-of-words or embedding step. Which Domain 1 encoding technique is that, and what should the item not become?
• A tree ensemble on the same bulk-trash table is insensitive to column magnitude, but a k-nearest or linear path on that table is not. When should the team scale?
• A notebook fits a scaler on the entire hop-harvest file, then splits. Validation rows have already leaked into the transform. What did the team do wrong?
• Nightly lock-ferry training reads a historical feature snapshot; a later low-latency lookup needs the current row. Which two SageMaker Feature Store access patterns is that?
• Two gondola notebooks independently recompute mean wait over 7 days and drift from each other. Which Feature Store practice keeps train and serve aligned?
• Orchard botanists must tag leaf-damage classes on a private workforce with a labeling UI and audit trail. Which official labeling service should they use?
• A one-off batch of 8,000 short choir-rehearsal clips needs human tags and the agency has no private labelers. Which official public workforce should they use?
• A historic brick-kiln has ten years of unlabeled traces and a smaller Ground Truth-tagged defect set. Which pile is supervised training data?
• After humans tag a seed batch of marina berth photos, the team wants the service to propose labels on the rest for human confirmation. Which Ground Truth pattern is that?
• A night-market ticket parse is usually fine, but low-confidence rows should pause for a person before they enter the training table. Which in-scope service routes that review?
• Failed potash-kiln batches are 2 percent of rows. Accuracy looks fine if the model always predicts pass. Which official pre-training metric should the team compute first?
• Priority inspection labels appear far more often for one tug harbor than another even when traffic volume is similar. Which official pre-training metric surfaces that gap?
• Festival volunteer incident notes are almost all routine; the hazard class is a handful of paragraphs. Which data-prep problem is that?
• An orchard leaf-photo set has thousands of healthy leaves and a few blight shots. Training on the raw folders will ignore blight. Which data-prep problem is that?
• A historic brick-kiln’s 2 percent defect class needs more presence in the training slice. Which official strategy should they apply, and where?
• There are too few vineyard blight photos to undersample healthy ones without starving the set. Which official CI strategy can supply more minority examples?
• Transfer-station tickets were only recorded on weekday mornings, so weekend commercial haulers never appear. Which pre-training bias is that?
• One grain-scale weighbridge reads 8 percent heavy, and one botanist tags damage much more strictly than peers. Which bias is that, and what is it not?
• Before any trainer starts, the team needs CI, DPL, and related pre-training metrics on the berry-harvest table. Which official AWS tool should they run?
• Null rates on forage-plot moisture and impossible negative weights should fail the nightly job before a training dataset is published. Which official lake-scale validator can block that publish?
• A clerk wants an interactive quality look at a 20 MB permit file; the lake job needs a scheduled ruleset on 200 million rows. Which official tool fits which moment?
• Independent grain-scale tickets with no time leak need a held-out test slice and a validation slice. Which official prep split should they create?
• Daily trolley ridership rows must not let next month’s fares appear in this month’s training fold. Which split should they use, and what is the trap?
• Plat files were written district-by-district, so the last 20 percent is only one district. Splitting in file order creates a biased holdout. Which official prep step should they run first?
• A random split on a 2 percent brick-kiln defect table can put almost no defects in validation. Which split keeps the rare class in each slice?
• The vineyard blight folder is tiny. Simple crops and flips of existing blight photos enlarge that class without new field visits. Which official prep step is that?
• A resolutionhours field is filled only after a defect is closed, but it sits in the training table used to predict whether a defect will occur. What should the team do with that column?
• The berry-harvest lake holds grower identifiers and must sit encrypted under a customer-managed key before any SageMaker job reads it. Which Domain 1 control is that?
• Hop-yard field offices still upload over plain HTTP to a public bucket URL. Which integrity control belongs on the way into the lake?
• Water-taxi rider names and card numbers are still in the transform output that data scientists pull. Where should those columns be handled?
• A credit-union ML table still needs a stable customer key for joins but must not carry the raw account number. Which official technique should they apply?
• A clinic’s visit-risk table includes protected health information and must remain in a single approved Region. Copying it to a cheaper distant Region for just prep is not allowed. Which handling is correct?
• Nobody is sure whether old S3 prefixes still hold national IDs. Which in-scope service should inventory that PII before the objects become features?
• Public tide tables, internal ferry logs, and confidential manifest notes sit in the same bucket prefix. What should drive encryption, masking, and who may ingest each pile?
• Coastal lidar tiles already sit on Amazon EFS. Recopying that tree onto each SageMaker instance root volume would move several terabytes before training starts. How should the training resource read the tiles?
• A high-throughput Amazon FSx for Lustre file system already holds sequential rolling-mill acoustic traces. The team needs the SageMaker trainer to read that volume instead of restaging the traces. What should they configure?
• A 4 TB S3 prefix of orchard drone frames is read sequentially. The algorithm does not need every object on local disk before the first epoch, and a full copy delays the job. Which SageMaker channel input mode fits?
• A city bus-ridership table still fails Glue Data Quality checks, and SageMaker Clarify has not reported pre-training bias. The lead wants to launch the SageMaker training job tonight and clean the table later. What should the team do first?