MLA-C01 AWS Certified Machine Learning Engineer - Associate Study Guide¶

Independent AI-assisted resource — SOURCES + OBJECTIVES CHECKED; HUMAN REVIEW PENDING. Objective coverage, citations, transition warnings, links, and exam-integrity compliance were reviewed again on September 28, 2026 against the available official PDF; current HTML blueprint access remains unresolved. This is not a guarantee that the guide is error-free or current after that date. See the sources-and-objectives record. The official MLA-C01 exam guide is authoritative.

Study guide at a glance

Start here

Report an issue with MLA-C01

Blueprint checked2026-10-28

Review stateSources + objectives checked — human review pending

Guide-only active reading1 hr–1 hr 30 min

Prerequisites: No formal prerequisite; AWS targets candidates with at least one year using SageMaker and other ML engineering services plus experience in backend development, DevOps, data engineering, MLOps or data science. Practice data contracts/quality/features, traditional ML training/tuning/evaluation, SageMaker deployment and pipelines, monitoring/drift, cost, networking, IAM/KMS and incident recovery. New learners should normally target MLA-C02 because English MLA-C01 delivery ends September 28, 2026.

Exam essentialsRead the objective map, key distinctions, and readiness checklist.

Deep understandingWork through every domain, including decisions, failure modes, and related items.

Hands-on labsBegin with Hands-on lab path, then use the remaining labs to produce evidence you can inspect and explain.

Official objective domains

See the objective map in this guide

The active-reading range uses 90–130 words per minute for technical material. Labs, troubleshooting, note-taking, spaced review, and prerequisite gaps add time; it is not a total preparation promise.

Current baseline: MLA-C01, four scored domains, 50 scored plus 15 unidentified unscored questions, and a 720 minimum scaled score
Upcoming blueprint change: Retirement announced. September 28, 2026 is the last day to take MLA-C01 in English. Korean, Japanese, and Simplified Chinese remain available until MLA-C02 general availability. The September announcement gives January 14, 2027 as the planned date, while the credential page still says TBD; recheck the language-specific appointment before relying on that date. Registration for MLA-C02 opened September 1; its English-only beta delivery code is ME1-C02, with delivery beginning September 29. Verify the live certification page before booking.
Replacement: MLA-C02 study guide, which links its official blueprint. New learners should normally prepare for MLA-C02. This page exists for candidates already scheduled for C01 and for historical/reference use; passing C01 earns the certification, not an automatic exam-version conversion.
Important freshness boundary: MLA-C01 emphasizes traditional ML engineering and MLOps, with limited foundation-model selection/fine-tuning context. MLA-C02 explicitly adds generative AI, Amazon Bedrock, foundation-model development, agentic workflows, and their operations. Do not assume a C01 course covers C02 merely because the credential name is unchanged.
Official source: AWS Certified Machine Learning Engineer - Associate exam guide

Living-guide watch — September 28, 2026¶

Transition evidence: The September announcement gives a January 14, 2027 plan, but the credential page still leaves GA registration and delivery as TBD. English C01 ends September 28; the other three languages follow GA. Existing credentials retain their original validity.

Blueprint access gap: The C01 HTML landing page and all four domain pages returned 404 during this review. The official static PDF, version 1.0, remained accessible: all 70 detailed skill bullets were mapped, with knowledge requirements read alongside them. This is a historical baseline, not confirmation that no later revision exists. Keep the automated HTML check unresolved until AWS restores or identifies its replacement; do not replace C01 objectives with C02's. The review report records the evidence.

Confirmed product lifecycle; blueprint still names the concepts: AWS announced that several services and features moved to maintenance/no-new-customer status on July 30, including SageMaker Clarify, Debugger, and Model Monitor. Existing Model Monitor customers can continue using it, but AWS says there will be no new features and points new customers to an open-source SageMaker monitoring approach with QuickSight and CloudWatch; see the service availability announcement and Model Monitor availability notice.

Concern and study decision: Keep learning drift, bias, explainability, data-quality, alerting, and monitoring decisions because they remain in the C01 blueprint. Treat Model Monitor/Clarify/Debugger product-specific details as historical-current context for this retiring exam, not a recommendation for a new production design. Recheck official availability before choosing an implementation.

Current guide watch — October 28, 2026¶

English delivery of MLA-C01 ended September 28, 2026. AWS says Japanese, Korean and Simplified Chinese delivery continues during the MLA-C02 beta until the standard MLA-C02 launch on January 14, 2027. This guide preserves the C01 blueprint for historical/current non-English candidates; English candidates should prepare against MLA-C02 instead. The retired C01 guide URL now returns 404, so its preserved objective snapshot is the reference for this guide. AWS September 2026 certification update

How to use this guide¶

MLA-C01 validates the ability to build, operationalize, deploy, monitor, maintain, and secure ML solutions and pipelines on AWS. AWS targets someone with at least one year using SageMaker and related AWS services plus experience in a role such as backend development, DevOps, data engineering, MLOps, or data science. The target is an implementer working within an architecture—not the person defining an enterprise-wide ML strategy or deeply specializing in multiple ML fields.

The live page lists a 130-minute, 65-question, USD 150 exam. The detailed guide identifies multiple-choice, multiple-response, ordering, and matching items; 50 are scored and 15 unscored items are not identified. Recheck delivery, price, language, scheduling, and transition details immediately before booking.

Use this page in one of two ways:

Already booked for C01: map each weak area to the objective table, do the labs, then use original practice questions to test decisions rather than definitions.
Not yet booked: read the transition section, move to MLA-C02, and use only the enduring C01 material—data quality, traditional ML, SageMaker lifecycle, deployment, monitoring, security, and cost—as foundation.

For every scenario, identify:

the business prediction, latency, scale, explainability, cost, and compliance contract;
data source, format, arrival pattern, label quality, leakage, bias, and train/serve consistency;
algorithm or managed AI service, training strategy, metric, baseline, and approval evidence;
batch, asynchronous, serverless, real-time, multi-model, container, or edge delivery target;
orchestration, versioning, tests, deployment safety, monitoring, retraining, rollback, and ownership;
identities, network paths, encryption boundaries, logs, quotas, and cost controls.

About related items: A Related item: callout adds prerequisite, operational, architectural, or adjacent context that makes the objective easier to reason about. It is supporting knowledge, not a claim that the item appears verbatim in the official outline.

Objective map¶

Published domain	Weight	Central question
Data Preparation for Machine Learning	28%	How should data be ingested, stored, transformed, validated, protected, and made reproducible for training?
ML Model Development	26%	Which approach, algorithm, training method, tuning process, metric, and evidence fit the problem?
Deployment and Orchestration of ML Workflows	22%	How should a model be packaged, provisioned, released, scaled, automated, tested, and rolled back?
ML Solution Monitoring, Maintenance, and Security	24%	How are drift, service health, cost, access, networking, audit, and retraining controlled in production?

1. Data Preparation for Machine Learning — 28%¶

The official PDF Domain 1 page covers ingestion and storage, transformation and feature engineering, and data integrity/readiness.

Frame the data contract before choosing a service¶

A useful design begins with grain: what does one row, record, image, sequence, or event represent? Record the prediction timestamp, label definition, entity key, source-of-truth owner, arrival delay, retention, region, classification, and expected volume. Then decide batch versus stream and choose storage/processing from the access pattern.

Need	Common fit	Decision boundary
Durable data lake and training artifacts	Amazon S3	Object storage; design prefixes/partitions, formats, lifecycle, versioning, encryption, and access separately
Shared POSIX file access	EFS	Elastic shared file system; useful only when file semantics are required
High-performance managed file system	FSx family	Select the exact file system from workload/protocol/performance requirements
Relational operational source	RDS/Aurora	Preserve transactional workload; extract deliberately rather than training against production indiscriminately
Key-value operational source	DynamoDB	Export/stream from access patterns; do not scan a hot production table casually
Batch discovery and transformation	Glue, EMR/Spark, SageMaker Processing/Data Wrangler	Choose visual versus code, serverless versus cluster, scale, libraries, lineage, and operational ownership
Ordered, replayable event stream	Kinesis or managed Kafka path	Define partitioning, retention, checkpoint, late/duplicate events, schema, and backpressure
Reusable online/offline features	SageMaker Feature Store	Define event time, record identifier, online/offline need, ownership, freshness, and point-in-time behavior

CSV and JSON are easy to inspect and exchange but can be verbose and weakly typed. Parquet and ORC are columnar, compressed, and efficient when analytics or training reads a subset of columns. Avro is row-oriented with schema-evolution strengths for event/data exchange. RecordIO variants can optimize specific training paths. Compression saves storage and transfer but changes splittability and CPU cost. A format answer is correct only when it matches producer, reader, schema, query, and parallelism requirements.

For batch ingestion, make each run idempotent: immutable landing objects, a run or snapshot identifier, schema validation, checkpoints/manifests, deterministic outputs, and atomic publication where possible. For streams, define event time versus processing time, partition key, duplicate/out-of-order policy, watermark/window, checkpoint, replay, and poison-record destination. Capacity failures are not just a service-sizing problem; small files, skew, hot keys, inefficient serialization, and unbounded consumers can create them.

Related item: A data lake is not automatically a governed training dataset. Raw, curated, feature, label, and prediction data need explicit contracts, lineage, access, retention, and reproducibility.

For extraction bottlenecks, measure source read rate, network transfer, serialization and destination write rate separately. S3 Transfer Acceleration addresses a different bottleneck from provisioned EBS IOPS; neither repairs a skewed join or malformed records. File-based training also needs compatible file-system mounts, network paths, permissions and sufficient throughput, not just a readable source URI.

Transform without creating leakage or skew¶

Data preparation includes type coercion, deduplication, missing-value treatment, outlier decisions, normalization/standardization, binning, transformations, categorical encoding, tokenization, image/audio processing, and feature creation. Fit stateful transformations on training data only, persist their parameters, and apply the same transformation artifact to validation, test, and inference data.

Data leakage occurs when training uses information unavailable at prediction time or learns from the evaluation set. Common forms include future events in a random time split, target-derived features, normalization fitted on the entire dataset, duplicate entities across train/test, and joins that select a later record. Choose random, stratified, grouped, or time-based splits from the deployment reality. Keep a final test set isolated from iterative tuning.

Feature Store can help maintain training/serving consistency, but it does not repair a bad definition. For every feature, record entity, event time, computation version, source, owner, freshness SLA, default/null semantics, sensitive classification, online/offline availability, and point-in-time retrieval rule. “Latest value” is often wrong for historical training because it leaks future information.

Use Glue Data Quality or equivalent assertions to test schema, completeness, uniqueness, validity, distribution, freshness, and referential expectations. Quarantine failing data instead of silently coercing it. Publish counts and distributions by meaningful slices so a globally healthy average cannot hide a broken source, region, or customer segment.

Protect integrity, privacy, and fairness¶

Class imbalance means labels are uneven; selection bias means the sample does not represent the target population; measurement bias means collection/labeling systematically distorts reality. These require different responses. Resampling, class weights, synthetic data, threshold changes, and targeted collection can address imbalance, but none proves fairness. Evaluate impact by relevant subgroups and connect every mitigation to the intended use and error cost.

For existing customers, SageMaker Clarify can calculate pre-training bias metrics and later analyze model bias/explanations. A metric is evidence, not a verdict: confirm label meaning, favorable outcome, facet definition, sample size, and acceptable thresholds with domain owners. Maintain approval and limitation records.

Classify PII, PHI, secrets, confidential business data, licensed data, and residency constraints before moving it. Minimize collection; mask, tokenize, anonymize, or aggregate only with a threat model. Encrypt objects, volumes, databases, file systems, channels, and artifacts as required, and ensure principals can use both the data resource and KMS key. Restrict cross-account and training-role access; retain lineage and audit events without leaking sensitive payloads into logs.

Related item: De-identification is contextual. Removing names does not make a high-dimensional dataset anonymous if combinations of remaining fields can re-identify people.

2. ML Model Development — 26%¶

The official PDF Domain 2 page covers approach selection, training/refinement, and performance analysis.

Choose the least complex approach that meets the contract¶

First ask whether ML is warranted. Deterministic rules, search, analytics, optimization, or a managed AI API may meet the need with less data and operational risk. If ML fits, frame the target:

classification predicts a class or probability;
regression predicts a continuous value;
ranking orders alternatives;
forecasting predicts a time-indexed quantity;
clustering/embedding/anomaly detection discovers structure without conventional labels;
computer vision, NLP, speech, translation, recommendations, and foundation-model use cases have specialized services and evaluation needs.

Managed services such as Transcribe, Translate, Rekognition, Comprehend, Personalize, or Bedrock can reduce model ownership when their contract fits. SageMaker built-in algorithms provide managed containers and documented input/hyperparameter behavior. Framework containers support TensorFlow, PyTorch, scikit-learn, XGBoost, and other supported stacks. Script mode supplies training code to a managed framework image; bring-your-own-container is appropriate when runtime, system libraries, or serving behavior require control. JumpStart provides models and solution starting points. Verify licenses, regions, data handling, and current support.

Algorithm choice depends on target, data type and size, linear/nonlinear relationships, sparsity, class balance, missing values, explainability, training/inference cost, latency, update frequency, and team skill. Begin with a meaningful baseline—business rule, naive predictor, simple linear/tree model—so complexity has something to beat.

Related item: Interpretability includes global behavior, local explanation, uncertainty, and operational transparency. A technically explainable model can still be unsuitable if stakeholders cannot act on its outputs.

Train reproducibly and tune deliberately¶

A reproducible training run pins input snapshot/query, split, transformation artifact, code commit, container digest, dependency versions, random seed where meaningful, algorithm/hyperparameters, instance type/count, distributed strategy, environment, metric definitions, and output locations. Store model artifacts and evaluation evidence together, then register versions with approval status and lineage.

Batch size changes gradient noise, memory, throughput, and sometimes generalization. Learning rate changes update size; epochs/steps determine exposure to data. L1/L2 regularization, dropout, early stopping, feature selection, augmentation, and more data can control overfitting depending on the model. Underfitting calls for suitable features/model capacity/training; overfitting calls for stronger generalization controls—not simply more epochs.

Distributed data parallelism divides batches among workers; model parallelism divides a model that does not fit or perform on one device. Distribution adds communication and convergence costs. Managed Spot Training can reduce interruptible training cost when checkpointing and job tolerance make interruption acceptable. Warm pools can reduce repeated startup overhead. Match CPU/GPU/accelerator, memory, network, storage, and instance count to measured bottlenecks.

Automatic Model Tuning searches a declared hyperparameter space against an objective metric. Random search is a broad baseline; Bayesian approaches use prior results to choose promising trials; Hyperband-style early stopping can allocate less work to weak configurations. Prevent tuning from optimizing on the final test set. Bound jobs, ranges, concurrency, time, and spend, and inspect whether the “best” trial is stable and deployable.

Fine-tuning a pretrained model requires representative data, base-model/license review, adaptation method, catastrophic-forgetting controls, evaluation against the unchanged baseline, cost, artifact ownership, and safety/privacy assessment. C01 mentions foundation models, but this is not a substitute for MLA-C02’s expanded FM/GenAI scope.

Ensembling combines predictions; stacking learns a combiner from out-of-fold predictions to avoid leakage; boosting adds models that address previous errors. Each adds inference, explainability and versioning costs. Reduce model size through feature selection, pruning or compatible numeric representations only after rechecking quality and runtime compatibility. When fine-tuning, keep a regression set for earlier capabilities so catastrophic forgetting is measurable.

Evaluate the decision, not a vanity metric¶

For classification:

precision asks how many predicted positives are correct;
recall asks how many actual positives were found;
F1 balances precision and recall through their harmonic mean;
specificity measures true-negative rate;
ROC-AUC ranks positives above negatives across thresholds but can look optimistic in highly imbalanced problems;
precision-recall curves focus on positive-class tradeoffs;
a confusion matrix exposes actual error types.

For regression, MAE is interpretable and less sensitive to large errors; MSE/RMSE penalize large errors more; percentage metrics can fail around zero. Ranking, forecasting, anomaly, vision, and language tasks need metrics aligned to their output and business decision. Select a threshold from error costs and capacity, not from 0.5 by habit.

Compare train versus validation behavior to diagnose bias/variance. Use cross-validation when suitable, but preserve group/time constraints. Evaluate by region, cohort, device, product, class, and other relevant slices. Check calibration when probabilities drive decisions. Compare performance, latency, throughput, resource use, and cost—not accuracy alone.

For existing customers, SageMaker Debugger can capture tensors/metrics and identify training issues; Clarify can analyze bias and feature attribution; experiments and lineage make comparisons reproducible. A shadow variant receives production-like traffic without driving the response, enabling realistic comparison; an A/B test exposes variants to users and measures outcomes. Both need assignment, duration, safety, rollback, and statistical interpretation.

Related item: Model approval is a risk decision. The winning experiment should not enter production without data, metric, fairness, security, latency, cost, and rollback evidence.

3. Deployment and Orchestration of ML Workflows — 22%¶

The official PDF Domain 3 page covers deployment selection, scripted infrastructure, and CI/CD/orchestration.

Match inference mode to traffic and latency¶

Mode	Best fit	Watch
Batch Transform	Bounded offline datasets; no persistent endpoint	Input partitioning, job duration, failed records, output reconciliation, per-job provisioning
Real-time endpoint	Predictable low-latency interactive traffic	Always-on cost, instance selection, scaling, health, variants, quotas
Serverless inference	Intermittent traffic that tolerates supported payload/latency constraints	Cold starts, limits, feature/region support, concurrency
Asynchronous inference	Large payloads or longer processing with queued response	S3 input/output, queue age, notification, expiration, failure path
Multi-model endpoint	Many compatible models with shared fleet and variable use	Model loading/caching, noisy neighbors, cold latency, container compatibility
ECS/EKS/custom target	Platform requirements exceed managed SageMaker hosting fit	Cluster, scaling, networking, patching, observability, deployment ownership
Lambda	Small, short, supported inference or orchestration	Package/runtime/memory/time/concurrency limits and model loading
Edge optimization	Local/offline/low-latency constraints	Supported hardware/operators, accuracy change, update/security fleet lifecycle

Evaluate p50/p95/p99 latency, throughput, payload size, request duration, concurrency, burstiness, availability, data locality, hardware, explainability, cost, and operational ownership. CPU may be cheaper for small classical models; GPU/accelerators help only when model and batching use them. Benchmark the actual container and traffic shape.

Provided containers reduce maintenance; custom containers control dependencies and inference logic but add base-image, patch, scan, user, secret, logging, health, and compatibility responsibilities. Record image digest with model artifact, inference code, schema, dependencies, and evaluation report. Never use an unversioned “latest” tag as release identity.

VPC-attached endpoints and jobs require subnet address capacity, security groups, DNS, routing, VPC endpoints or controlled egress, endpoint policies, and least-privilege roles. A private subnet alone does not guarantee private data flow.

For edge deployment, learn the compilation/target-compatibility problem represented by SageMaker Neo in the C01 baseline. Verify device architecture, supported operators, model package, offline update/rollback and measured accuracy/latency; the old service name is not evidence of present onboarding or device support.

Provision, scale, and release safely¶

Use CloudFormation or CDK to define repeatable infrastructure, roles, encryption, networking, alarms, repositories, build projects, pipelines, and endpoints. Parameterize genuine environment differences; avoid uncontrolled console drift. Change sets, policy checks, tests, and rollback make IaC a release mechanism rather than a template archive.

Target-tracking scaling follows a metric such as invocations per instance; step scaling reacts in defined increments; scheduled scaling prepares for known demand. Min/max capacity, cooldown, metric delay, instance startup, downstream quotas, and failure response matter. Scaling cannot repair a slow model or a saturated dependency. Provisioned resources serve steady load; on-demand/serverless/Spot choices trade readiness, interruption, and cost.

Deployment patterns include all-at-once, rolling, blue/green, canary, linear, and shadow. Define routing percentage, bake time, success metrics, alarm thresholds, minimum sample, rollback trigger, and artifact/config/database compatibility. Version and approve models in Model Registry; separate registration from production deployment so evidence and authorization remain visible.

Orchestrate the complete lifecycle¶

SageMaker Pipelines models processing, training, evaluation, conditional approval/registration, and related ML steps as a DAG. Step Functions is useful for broader service orchestration and explicit branching/retry/catch behavior. Managed Workflows for Apache Airflow fits teams already using Airflow DAGs. EventBridge can trigger workflows from schedules or events. Choose based on scope, state model, integrations, retry semantics, observability, and operator skill—not by diagram preference.

A production path commonly follows:

source change or schedule → data validation → processing/features → training → evaluation gates → bias/security checks → registration → approval → staging deployment → integration/load tests → canary → monitor → promote or rollback

CodePipeline coordinates release stages; CodeBuild executes builds/tests; CodeDeploy supports deployment patterns for compatible compute. Git branch/flow conventions should protect mainline and connect commits to artifacts. GitHub Flow typically uses short-lived branches and reviewed mainline integration; Gitflow introduces dedicated release/development branches and therefore additional promotion/merge responsibilities. In either case, bind pipeline triggers to approved commits and prevent pull-request code from receiving deployment secrets. Unit tests cover transformation and inference logic; integration tests exercise storage, permissions, containers, endpoints, and orchestration; end-to-end tests validate representative requests and rollback. Include malformed input, schema change, missing feature, duplicate event, timeout, throttling, unauthorized access, corrupt artifact, and failed evaluation.

Retraining triggers can be schedule-, data-, drift-, performance-, or event-based. A trigger starts evaluation, not automatic promotion. Reuse the approved pipeline, compare to the production baseline, require gates, and preserve rollback. Keep data, feature, code, container, and model versions distinct so an incident can locate the changed component.

Related item: ML CI validates code and infrastructure; ML continuous training validates data and model behavior; CD releases an approved artifact. Treating all three as one opaque pipeline weakens controls.

4. ML Solution Monitoring, Maintenance, and Security — 24%¶

The official PDF Domain 4 page covers model inference, infrastructure/cost, and security.

Monitor four different systems¶

Data: schema, types, missingness, ranges, categories, freshness, volume, label availability, sensitive-field leakage, and feature distribution.
Model: prediction distribution, confidence/calibration, bias and feature-attribution drift, ground-truth performance, threshold/business outcome, and model version.
Service: request count, errors, model latency, overhead latency, concurrency, queue/backlog, CPU/GPU/memory, disk/network, endpoint/variant health, and dependency behavior.
Workflow/business: pipeline duration/failure, skipped/late inputs, approval state, retraining result, deployment result, cost, user impact, and ownership/SLA.

Data drift is a statistical input change; concept drift changes the relationship between input and outcome; label drift changes outcome distribution; model-quality degradation is measured after ground truth arrives. For existing customers, Model Monitor supports data quality, model quality, bias drift, and feature-attribution drift workflows. Establish a representative baseline, capture inference data safely, schedule monitoring, publish constraints/statistics, alarm on meaningful violations, and connect alarms to investigation—not blind retraining.

An A/B test measures competing variants on assigned live traffic; shadow testing observes a candidate without affecting responses. Define hypotheses and guardrails before traffic begins. Correlation is not causation when marketing campaigns, seasonality, or population shifts coincide with deployment.

CloudWatch metrics, logs, alarms, dashboards, and Logs Insights show service/workflow behavior. CloudTrail records API activity and can support event-driven response. EventBridge routes operational events. X-Ray helps trace supported request paths. Use correlation IDs across data, workflow, model, endpoint, and application logs while excluding sensitive payloads. An alert needs owner, severity, runbook, evidence query, safe mitigation, and closure criteria.

Choose an available implementation¶

The blueprint's named service and an available new-account lab can differ. Clarify's migration guidance describes direct bias calculations, SHAP and evaluation alternatives. Debugger's guidance points to MLflow for run evidence, TensorBoard for model internals, and CloudWatch for resource telemetry. Neither feature is open to new customers. The June 30 announcement of July changes also includes Ground Truth, Mechanical Turk and A2I; use synthetic, manually adjudicated labels for introductory labs.

For Model Monitor, follow the linked availability notice and assess AWS's open-source monitoring designs before choosing a new implementation. A reference architecture still needs account permissions, cost estimates, dependency/version checks and validation. Practice the portable contract first: prediction ID, model version, timestamp, permitted input statistics, delayed label join, slice metrics, alert owner and recovery action. Existing customers should preserve evidence and prove replacement coverage before disabling schedules or deleting artifacts.

Optimize from evidence¶

Investigate latency by separating client/network, queueing, container overhead, model inference, serialization, and dependency time. Test instance family/size, concurrency, model size, precision, batching, compilation, caching, and autoscaling against representative traffic. SageMaker Inference Recommender can assist instance/config comparison; Compute Optimizer applies only to supported resources and is not a substitute for model-specific benchmarking.

Tag resources by application, environment, owner, model, and cost center. Use Cost Explorer for analysis, Budgets for thresholds/forecasts/actions, and current pricing tools for estimates. Common levers include right-sizing, shutting down idle endpoints/notebooks, batch or async delivery, autoscaling, multi-model hosting, managed Spot Training with checkpoints, suitable storage/lifecycle, efficient data format, and avoiding repeated transformations. Optimize dollars per useful outcome, not one resource line in isolation.

Service quotas can present as throttling, failed scale-out, queued jobs, or deployment failure. Monitor quota use and request increases before events. Provisioned concurrency/capacity and reservations trade guaranteed readiness or discount against utilization commitment; verify which purchase options apply to the exact resource.

Secure every principal, path, and artifact¶

Separate human, notebook, processing, training, pipeline, build, deployment, and endpoint execution roles. Grant only required actions on scoped resources and constrain iam:PassRole. Align identity policies, resource policies, KMS key policies/grants, bucket policies, endpoint policies, SCPs, and permissions boundaries. Do not solve access errors by adding wildcard administrator rights.

Protect data and artifacts with TLS and appropriate at-rest encryption. KMS permissions must permit the actual service principal or execution role along the full path. Store credentials in Secrets Manager or appropriate managed configuration, rotate them, and prevent exposure in notebooks, environment variables, images, logs, model artifacts, and pipeline parameters.

Use private networking where required: approved subnets/security groups, interface/gateway endpoints, controlled egress, DNS, and no-public-internet settings supported by SageMaker. ECR image scanning, signed/provenanced artifacts, dependency review, patched base images, non-root containers, and immutable digests protect the software supply chain. CloudTrail, Config, logs, model lineage, and deployment evidence support audit and incident response.

Threats include poisoned training data, unauthorized artifact replacement, dependency compromise, over-broad notebook access, secrets in code, malicious serialized models, inference abuse, sensitive prediction logging, and cross-tenant access. Controls should prevent, detect, and recover: source validation, checksums/signatures, approval gates, isolation, least privilege, input/output validation, rate limits, monitoring, backup, and rollback.

Related item: Shared responsibility applies inside the ML lifecycle. AWS secures underlying managed infrastructure; the customer still owns data, identities, network configuration, code, containers, model behavior, monitoring, and use.

Integrated scenarios¶

Scenario 1: Fraud scoring with delayed labels¶

A payment platform needs sub-100-ms predictions, receives labels weeks later, and has severe class imbalance. Land immutable events in S3 and stream operational features with explicit event time. Use point-in-time-correct offline features and a low-latency online path. Split evaluation by time and customer/entity to avoid leakage. Compare a simple baseline and tree model; tune against a metric tied to review capacity and false-negative cost, not accuracy. Register data/code/model evidence. Deploy to a real-time endpoint with a canary, idempotent callers, and rollback. Monitor service latency immediately; monitor data/prediction distribution continuously; calculate model quality only when labels mature. Investigate drift before retraining and require the candidate to beat the production model on performance, bias, latency, and cost gates.

Scenario 2: Nightly demand forecast for many products¶

The output is consumed once daily, so a permanent endpoint is unnecessary. Build time-aware features without future sales/promotions, preserve product groups in evaluation, and compare naive seasonal forecasts before complex models. Orchestrate processing, training/tuning, evaluation, registration, approval, and Batch Transform. Partition work and output by date/product, reconcile record counts, and publish atomically. Use Spot training only with supported checkpoint/retry tolerance. Monitor source freshness, pipeline duration, missing products, forecast error by product group, and total business outcome. Retraining on schedule is justified only when new labels and stable data are available.

Scenario 3: Regulated document classification¶

Documents contain sensitive data and require explainability. Classify/minimize data before landing it in an encrypted S3 bucket. Use private paths, scoped roles, KMS controls, and auditable processing. Validate labels and evaluate precision/recall by document/customer group. Compare a managed AI service, pretrained/fine-tuned model, and custom model against accuracy, explainability, residency, cost, and ownership. Store lineage and approval. Deploy asynchronously for large documents; capture only permitted inference metadata. Monitor input schema/distribution, failure queue, latency, model performance after adjudication, and subgroup behavior. Quarantine rather than log malformed sensitive content.

Worked decisions¶

These small examples are original exercises. The arithmetic was checked locally; no AWS resources were created.

Example 1: A historical feature can still arrive too late¶

At 10:00 a prediction runs. Feature A has event time 09:55 and was available to inference at 09:56; feature B describes 09:58 but arrives at 10:05. A historical replay of what the model could know must exclude B. Filter both event time and actual availability time before selecting the latest eligible value. Define how late corrections, deletes and stale values are treated.

The AWS/GoDaddy feature-history article explains event, API and offline-write timestamps. Its example selects corrected history using write time as a tie-breaker. Our availability cutoff is an additional requirement for reproducing the original decision: a later correction may be appropriate for a corrected-history experiment, but must not silently enter an as-served replay. Preserve the experiment's chosen semantics.

Example 2: Choose a threshold from review capacity¶

Across 1,000 cases, 100 are actually positive. Threshold A produces TP=80, FP=40, FN=20, TN=860: precision is 80/120 = 66.7%, recall is 80%, F1 is 72.7%, and 120 cases require review. Threshold B produces TP=65, FP=15, FN=35, TN=885: precision is 81.25%, recall 65%, F1 72.2%, and 80 cases require review.

With capacity for 100 reviews, A exceeds capacity and B fits. B is not automatically acceptable: compare the cost of 15 additional missed positives against 25 fewer false alarms, or redesign capacity/routing. Neither accuracy nor F1 alone settles the business decision. Check slice performance and sampling uncertainty before promotion.

Example 3: Provisioning cannot repair a release contract¶

An endpoint can sustain 25 requests/second per instance at the required latency. A forecast of 80 requests/second needs ceil(80/25) = 4 instances under this measured model. Losing one leaves only 75 requests/second; 5 are needed to preserve the target after one instance loss. Benchmark again under realistic payloads, quotas and startup delay; this arithmetic is not an AWS capacity guarantee.

A canary receiving 10% of 10,000 requests sees only 1,000 requests. If a critical population is 1% of traffic, it may see about 10 examples, too few to establish safety. Add targeted evaluation and an explicit sample/time gate. Keep a deployable prior manifest containing preprocessing, model artifact, image digest, endpoint configuration and permissions; verify rollback against dependencies.

Example 4: Delayed labels change the denominator¶

Of 200 predictions, only 50 have matured labels; 45 are correct. Observed accuracy is 90% on labeled cases, with 25% label coverage. It is not 90% accuracy across all 200, and 150 unknown outcomes are not automatically failures. If slow labels cluster in harder cases, the early score is biased. Report label age/coverage and compare matched, mature cohorts while separately monitoring schema, distribution, latency and errors.

Hands-on lab path¶

Use a sandbox account and budgets. Destroy billable resources after each lab; do not use production or sensitive data.

Data contract and format: Create synthetic tabular events with schema versions and event timestamps. Compare CSV and Parquet size/read behavior in S3; document partition and replay rules.
Leakage-safe preprocessing: Build train/validation/test splits, fit transforms only on training data, persist preprocessing, inject schema/null/range failures, and publish a quality report.
Feature consistency: Define two time-aware features, create offline historical values and a simulated online lookup, and prove point-in-time retrieval avoids future leakage.
Train, tune, and evaluate: Train a simple baseline and XGBoost or supported equivalent, run bounded tuning, compare metrics/confusion matrices/slices, and register the winning evidence.
Inference modes: Run batch inference and a temporary real-time or serverless endpoint. Measure payload, latency, throughput, cold/warm behavior, and estimated cost; delete the endpoint.
Pipeline gate: Build a SageMaker Pipeline with processing, training, evaluation, conditional threshold, and registration. Make a failed metric prevent registration.
Canary and monitoring: Route limited synthetic traffic to a candidate variant, capture permitted data, create baseline/monitoring statistics, inject drift, alarm, and execute rollback.
Security and incident drill: Scope S3/KMS/SageMaker roles, deny an unauthorized path, trace the failure with CloudTrail/CloudWatch, repair least privilege, rotate a test secret, and record cleanup evidence.

Availability adjustment: For monitoring, bias and convergence labs, use the alternatives above if the account cannot access the legacy features. Start with a local CSV containing prediction IDs, labels and subgroup columns; reproduce Examples 2 and 4 and inject missing/duplicate labels. Proposed cloud labs have not been executed in this review.

For each lab, preserve architecture, assumptions, commands/code, policy, dataset/model versions, expected and actual results, failure injection, cost, cleanup, and what you would change for production.

Original knowledge checks¶

These are original blueprint-aligned prompts, not recalled exam questions.

Why can a random split leak future information in a forecasting or churn dataset?
When is Parquet preferable to JSON, and what producer/consumer constraint could reverse the choice?
What makes a batch ingestion job idempotent?
Why is a high-cardinality stream partition key operationally important?
What is the distinction between event time and processing time?
Why can “latest feature value” be wrong for historical training?
Which checks distinguish schema validity, completeness, uniqueness, and freshness?
How do class imbalance, selection bias, and measurement bias differ?
Why does encryption require both resource and KMS authorization reasoning?
What evidence is needed before labeling a dataset de-identified?
When should a managed AI service be chosen over a custom model?
What baseline would make a complex model improvement credible?
How do underfitting and overfitting appear in training and validation metrics?
Why must hyperparameter tuning not optimize against the final test set?
When is data parallelism more appropriate than model parallelism?
What makes Spot training safe or unsafe for a job?
Why can ROC-AUC be misleading for a rare-positive problem?
How should an operating threshold be selected from precision and recall?
What additional evidence beyond accuracy belongs in model approval?
How does a shadow test differ from an A/B test?
Which inference mode fits a nightly complete-file job, and why?
When is asynchronous inference a better fit than a real-time endpoint?
What risk does a multi-model endpoint introduce even when it saves cost?
Why is an image tag such as latest insufficient release identity?
Which network dependencies can break a VPC-attached training job?
How can endpoint autoscaling lag behind a sudden burst?
What must be defined before a canary can be called safe?
How do SageMaker Pipelines and Step Functions differ in primary scope?
Why should model registration be separate from production deployment?
What negative tests belong in an ML delivery pipeline?
Why should a drift alarm not automatically promote a retrained model?
How do data drift, concept drift, label drift, and model degradation differ?
Why can a global metric hide a production failure?
What is the purpose of an inference-data baseline?
Which signals separate model latency from queue or dependency latency?
Why might a larger instance be cheaper per prediction?
What tags and dimensions make ML cost explainable?
Which distinct roles should exist across an ML lifecycle?
Why does private-subnet placement not by itself guarantee private traffic?
Which controls reduce training-data poisoning and model-artifact replacement risk?

Answer explanations¶

Random splitting can place later observations or the same entity in training and test; use a split that matches future deployment and fit transforms only on training.
Parquet reduces column scans and carries a schema; streaming producers or consumers requiring JSON may outweigh that benefit.
A stable input/run identity and deterministic, deduplicated output make replay produce the same committed result.
A well-distributed key reduces hot partitions; high cardinality alone does not fix a single disproportionately busy key.
Event time describes when the event happened; processing time describes when a system handled it. Late arrivals require an explicit window/replay policy.
The latest value may include future events or corrections unavailable at the original prediction time; apply the appropriate time contract.
Schema checks types/structure; completeness checks required values/records; uniqueness checks duplicate keys; freshness checks acceptable age.
Imbalance concerns label counts; selection bias concerns who enters the dataset; measurement bias concerns how observations or labels are recorded.
Permission to read an S3 object does not necessarily grant permission to use its KMS key; evaluate the relevant identity, resource and key policies.
Document direct/quasi-identifiers, linkage and re-identification tests, permitted use and reviewer approval. Encryption alone does not de-identify data.
Use a managed API when its measured quality, customization, geography, latency and total cost meet requirements with less operating effort.
Use a simple, reproducible rule/seasonal/statistical model on the same uncontaminated split and business metric.
Underfitting often has poor training and validation results; overfitting has good training but materially worse validation, subject to distribution and data-quality checks.
Repeated tuning against test outcomes leaks information into selection; use validation and reserve the final test for the frozen candidate.
Use data parallelism when the model fits a device and useful batch work can be divided; use model parallelism when model size or architecture requires it.
The training job must support resumable checkpoints, an interruption-tolerant deadline and bounded retries; otherwise the discount may be outweighed by lost work.
A small false-positive rate can still overwhelm a rare-positive population. Inspect precision-recall and absolute error counts at the operating threshold.
Compare confusion counts, false-negative/false-positive costs and review capacity on representative validation data; then validate the frozen choice.
Include slice quality, bias/explanation evidence, data lineage, reproducibility, security, latency, cost, operational readiness and rollback.
Shadow inference does not select the user-visible answer; A/B changes user exposure. Shadow still needs privacy, cost and downstream-side-effect controls.
Batch Transform fits offline input files and a completion deadline without keeping an endpoint continuously running.
Async fits supported long-running/large-payload or bursty work where queued completion is acceptable; verify its current limits.
Model loading, cache eviction and shared capacity can produce cold latency and noisy-neighbor effects.
Tags can be reassigned; preserve an immutable image digest alongside code, data, preprocessing and model identities.
Check DNS, routes, endpoints or egress, security groups, NACLs, S3/ECR/log access and authorization including KMS; private placement alone is insufficient.
Metrics, scaling decisions, quotas and instance/model startup all introduce delay; test bursts and provide buffering or planned capacity.
Define traffic/cohorts, minimum evidence, quality and service alarms, bake time, rollback target and the authority to stop promotion.
Pipelines focuses on an ML job/artifact DAG; Step Functions coordinates broader service/workflow state, retries and error handling.
Registration preserves a candidate and its evidence; deployment changes production and requires its own gates and authorization.
Test malformed/stale data, leakage, failed evaluation, unauthorized access, bad images, missing artifacts, timeouts, duplicate events and failed rollback.
Drift may indicate a broken input pipeline or a harmless population shift. Diagnose first; any retrained candidate must pass quality and release gates.
Data drift changes inputs; concept drift changes the input/outcome relationship; label drift changes outcome prevalence; measured degradation needs valid ground truth.
Large easy cohorts can dominate the average while a small, important population fails; inspect meaningful slices and sample sizes.
It provides a versioned, representative reference for constraints and distributions; it must be appropriate to the population and intended comparison.
Trace application, queue, invocation, model processing and downstream timing with correlation IDs; compare resource and error metrics.
A larger instance may complete substantially more useful work per billed hour; measure cost per successful prediction at the required latency.
Include application, environment, owner, model/release, workload stage, cost center and usage volume so shared and idle costs remain visible.
Separate human development, ingestion, training, evaluation, build, deployment and runtime responsibilities; limit PassRole and audit privileged transitions.
Routes, DNS, gateways, service endpoints and allowed egress decide the path; a subnet label does not enforce it.
Validate source provenance and labels, restrict writes, version and hash artifacts, scan dependencies, isolate roles and gate releases with audit evidence.

Final review checklist¶

I can explain all four domains and their 28/26/22/24 weighting.
I can map a scenario from data contract through monitoring, security, cost, and rollback.
I choose formats, storage, batch/stream paths, transformations, splits, and feature-store behavior from requirements.
I distinguish leakage, imbalance, bias, drift, and ordinary quality failures.
I select a model/managed service, training strategy, metric, threshold, tuning plan, and approval evidence.
I choose batch, real-time, serverless, asynchronous, multi-model, container, or edge deployment deliberately.
I can construct a versioned pipeline with tests, evaluation gates, registration, canary, monitoring, and rollback.
I monitor data, model, service, workflow, business, security, and cost signals separately.
I reason through IAM, KMS, S3, VPC endpoint, container, secret, lineage, and audit boundaries.
I have checked the live transition page and know whether I am actually sitting C01 or the C02 beta/GA exam.

Places to learn¶

This is not a complete list, and it is not meant to be consumed end to end. Pick the explanation style, lab environment, and practice format that close your own gaps. Because C01 retires imminently, do not begin a long paid C01 path unless you already hold a valid C01 appointment; new learners should prefer C02-labeled material. On September 28 the public Pluralsight page still listed 20 hours, five courses and eight labs; this is catalog evidence, not a review of the paid lessons. Sign-in shells, blocked pages and video shells cannot confirm their current content. Times are provider-stated when stable and otherwise transparent estimates; access, catalog contents, and runtimes can change.

Resource	Access	Estimated time
Official MLA-C01 exam guide (version 1.0 fallback; HTML unavailable)	Public	2–4 hours for a careful objective map
AWS certification page and transition notice	Public	10–15 minutes; recheck immediately before booking
AWS Skill Builder MLA-C01 exam-prep plan	Mixed public/subscription	Allow 20–35 hours as a study-planning estimate; current plan contents/runtime require sign-in verification
SageMaker ML lifecycle overview	Public	30–60 minutes plus linked implementation topics
SageMaker Pipelines tutorial	Public; AWS usage may cost	3–6 hours hands-on
SageMaker Model Monitor	Public; AWS usage may cost	2–4 hours; existing-customer service, use the availability alternatives for a new account
AWS/GoDaddy point-in-time feature-history article (June 22, 2021)	Public	45–90 minutes estimated reading/exercise; useful time-semantics lesson, recheck old SDK code
Machine Learning Lens	Public	4–8 hours selected lifecycle review
Pluralsight MLA-C01 path	Paid/trial	20 hours; five courses, eight labs, and practice exam listed
AWS Certified Machine Learning Engineer Study Guide by Dario Cabianca	Paid/subscription	Allow 15–25 hours as a reading estimate; current edition, length and test-bank access were not reverified
LinkedIn Learning MLA-C01 Cert Prep	Paid/trial	24h 58m plus exercises
Nikolai Schuler MLA-C01 course	Paid	About 25–35 hours including demos and one practice exam; verify displayed runtime
Stéphane Maarek and Abhishek Singh MLA-C01 practice exams	Paid	Allow 6–10 hours for practice and rationale review; current bank size was not reverified
MLA-C01 full course by Tech With Lucy	Public	Allow 4–8 hours including pauses; runtime and video contents were not reverified

Use practice questions as a diagnostic: explain why each alternative fails the stated requirement, return to official product documentation, and implement the weak concept. Avoid recalled-question collections and any source marketed as an exam dump.