Deployment and Orchestration of ML Workflows

MLA-C01 · 66 questions

• A marina slip-application clerk UI must return a risk flag in well under a second, all day, with steady office-hour traffic. Which SageMaker serving mode fits?
• A county-fair pie-score model is invoked a few dozen times on fair weekend and then sits idle for months. The board will not pay for an instance that stays up. A cold start of a few seconds is acceptable. Which serving mode fits?
• Pier cameras drop multi-hundred-megabyte clips that take many minutes to score. Clerks can wait for a queued result; they cannot block a synchronous HTTP call for an hour. Which serving mode fits?
• Once a week the compost office drops a single large CSV of well readings and wants a scored file back in the same bucket. Nobody will call an API per row, and no endpoint should sit around between runs. Which serving mode fits?
• Ferry planners need a bunching-risk score for every stop by 06:00 from last night's AVL dump. The SLA is a completed file, not a sub-second UI call. Which deployment strategy fits?
• A hatchery lab uploads one long raw-instrument file per sample. Scoring plus preprocessing overshoots the real-time timeout, and the file is too large for a serverless payload, but a same-morning queued result is fine. Which mode fits?
• A moth-trap count model is hit in short evening bursts, then silent until the next dusk. Finance flagged last month's always-on real-time bill. A short cold start is acceptable for the field app. Which tradeoff should they make?
• A dock-camera species tagger is a large vision network, and CPU hosts miss the ranger-app latency budget. A second, tiny tabular lot-full model on the same campus does not need a GPU. How should inference compute be chosen?
• Parks IT wants the staging reservation-no-show endpoint to mirror production instance counts so it feels real, even though staging traffic is a handful of synthetic calls a day. How should test compute be provisioned?
• A water-plant turbidity model must sit on a CPU family the agency already standardized, and each invoke ships a wide feature vector from an in-VPC store. The constraint is host generation and NIC bandwidth, not use a GPU because it is ML. What should select the inference host?
• The harbor-tug wait model is a supported scikit-learn or XGBoost-class artifact. A contractor wants to spend a sprint writing a custom image for control. Which inference container should they use?
• The lock-chamber occupancy scorer depends on a niche hydraulics library that no SageMaker provided inference image ships. The serving contract is still a SageMaker endpoint. Which container path fits?
• Each county boat-ramp has its own contamination model with the same framework and roughly the same size. Most ramps are quiet; only a few are busy at once. One endpoint per ramp would waste hosts. Which endpoint pattern fits?
• Visitor-center comment scoring always runs a tokenizer/featurizer container and then a model container on the same request. The agency wants one endpoint, not two hops the app team must glue together. Which deployment fits?
• The ice-rink routing model is invoked continuously during storms. A teammate proposes stuffing it onto the existing multi-model endpoint beside twenty rarely used precinct models. What should they use instead?
• A new compost-intake scorer has no existing container platform. The county already trains in SageMaker AI and wants a managed HTTPS endpoint, not a new cluster. Which deployment target fits?
• The bookmobile due-tonight ranker is a tiny serialized model, invoked from an existing Lambda-backed API, a few thousand times a day, well inside Lambda size and timeout limits. Standing up SageMaker hosting would be new platform work. Which target fits?
• Parks already serves three internal APIs on Amazon ECS. They want the trail-washout severity model packaged as one more service on that cluster, with the same deploy path, not a new SageMaker endpoint estate. Which target fits?
• The ferry operations team already runs Amazon EKS for internal tools and requires the new delay scorer to land as a Kubernetes Deployment next to those services. Which target fits?
• Pier kiosks run a small local board with a named target architecture and cannot call the VPC during weekend outages. The trained artifact must be compiled to that device, not hosted as a cloud endpoint the kiosk cannot reach. Which capability fits?
• One county team already runs dozens of Airflow DAGs on Amazon MWAA. A greenfield team has only SageMaker training and processing steps and no Airflow estate. Both need an orchestrator for the deployable workflow. How should each team choose?
• Finance will not fund GPU real-time hosts so rangers can click score this photo instantly. Overnight batch scoring of the day's dock-camera dump meets the posted-alert SLA at a fraction of the cost. Which serving choice fits?
• After a bad tug-wait model went live, ops needs to put the previous endpoint configuration back in minutes. The artifacts are already registered. What is the Domain 3 gap?
• Two approved dock-inspection models must share live traffic on one real-time endpoint from the moment of deploy (10 / 90 split). Ops is shifting invoke traffic, not reading a shadow copy and not running a weeks-long production A/B monitor. What should they configure?
• The pier-advisory serverless endpoint is cheap overnight but clerks hit a cold-start wall at 08:00 opening. Leadership will pay for a small reserved slice during business hours only. Which provisioning choice is that?
• Marina-permit traffic is smooth most days, spikes when a storm-repair window opens, and is predictably dead on Sundays. The team must pick a scaling-policy shape, not an instance family. Which policies match those shapes?
• Every compost-scoring environment was clicked together in the console and now drifts. The requirement is a repeatable template for the model, endpoint configuration, and endpoint. Which official IaC service fits?
• The ferry team already authors AWS CDK apps in Python or TypeScript for other services and wants the bunching endpoint expressed as constructs in that same app, not a new YAML dialect. Which IaC path fits?
• Two equally valid templates would stand up the lock-chamber endpoint. One group lives in reviewed CloudFormation YAML; the other wants CDK because they already review Python PRs. What should decide the IaC choice?
• Networking already owns a CloudFormation stack that exports subnet and security-group IDs. The ML stack must consume those exports so the new endpoint lands in the existing VPC without copying IDs into chat. What should they automate?
• The custom lock-chamber hydraulics serving image is built in CI and must be pulled by SageMaker (or ECS/EKS) in each account. Someone proposed leaving tarballs in a personal S3 prefix. Where should the image live?
• After the lock-chamber team chose a custom container, they still need the scripted path: image in ECR, SageMaker model image URI set, endpoint created. Which BYOC steps fit?
• Parks already picked Amazon ECS as the hosting target. The remaining work is a task definition plus service that runs the serving container from ECR, not a SageMaker Model.deploy call. What should they script?
• Ferry operations already picked Amazon EKS. The serving image must run as a Kubernetes Deployment on that cluster, with the image coming from ECR. What should they script?
• The brine-plant model must not be reachable on the public internet. Private subnets and a security group already exist; the endpoint just is not using them. What should they configure?
• An ice-rink engineer has a trained Model (or estimator) in a notebook and needs the snippet-level call that creates the real-time endpoint, not a CloudFormation rewrite and not a from-scratch Flask server. Which path fits?
• The marina-permit real-time variant is even on CPU, but each host is drowning in invoke count at noon. The scaling signal should be traffic per host, not a guess at CPU. Which auto scaling metric fits?
• The tug-wait hosts show high CPU while invoke counts still look modest, because each call does heavy featurization. Scaling on invoke count alone never adds capacity in time. Which auto scaling metric fits?
• Ice-rink scoring still returns 200s, but p99 model latency crosses the dispatch budget before CPU looks high. Ops wants capacity to follow latency, not a vanity CPU chart. Which auto scaling metric fits?
• Pier-festival ticket-risk invokes are near zero except 10:00–20:00 on event days, which are on a published calendar. Scaling only on live demand always lags the opening rush. Which auto scaling approach fits?
• Boat-ramp contamination traffic follows no weekly clock. A viral news story can 10x invokes on a Tuesday. A Sunday scale-down schedule would strand weekday spikes. Which auto scaling approach fits?
• The existing ice-rink mobile API is already AWS Lambda. The new no-show model lives on a SageMaker endpoint. Lambda should validate the payload and call the endpoint rather than exposing the endpoint to the phone app. How should Lambda be used?
• A burst of overnight batch-style EC2 scoring for dock-camera dumps can die and retry. The scripted fleet should add Amazon EC2 Spot Instances for that elastic slice and keep a small On-Demand core. Which capacity move is that?
• Fear of a Saturday festival surge led parks to pin eight real-time hosts all month. Average utilization is tiny. What is the maintainable fix?
• Someone turned on a CloudWatch alarm but never registered the SageMaker variant with Application Auto Scaling, so instance count never moves. What step is missing?
• A well-meaning intern changed the staging endpoint instance type in the console. The CloudFormation or CDK stack no longer matches reality, and the next deploy rolled the change back. How should hosting changes be made?
• The compost-intake model is promoted by hand: download the artifact, click deploy. Leadership wants a Source, Build, Deploy pipeline that releases what Git says, not what is on someone's laptop. Which service shapes that pipeline?
• The pipeline's Build stage must run unit tests and build or push the inference image. Someone tried to put that work in a deploy-only click. Which service is the build and test workhorse?
• After CodeBuild succeeds, the county still copies the image tag into the console. They need the deploy tool that rolls the new revision onto the target with a defined strategy. Which service is that action?
• CodePipeline is red on Build. Logs show the buildspec path is wrong and the expected artifact name does not match what Deploy is looking for. What should they troubleshoot?
• Ferry models go to staging when someone emails a zip. The official source of truth should be a Git repository, and a commit or merge should create a pipeline execution. What should start the pipeline?
• One team wants long-lived develop and release branches before production deploy. Another wants merge-to-main as the production trigger. Both must map onto CodePipeline. Which branching structures are those?
• The Git repo exists and the CodePipeline exists, but nothing runs on push. What piece is missing?
• The current pipeline only lints Python and then deploys whatever tar.gz is sitting in a shared bucket. A new commit never proves the model is still fit to ship. What CI/CD principle is missing for ML?
• The next dock-inspection model must stand up beside the live fleet and flip traffic only when the new environment is healthy, with a clean path back to the old environment. Which release strategy fits?
• Ops will not flip 100 percent of ice-rink dispatch invokes to a new revision. They want a small slice of live traffic first, then more only if that slice stays healthy. Which strategy fits?
• Leadership wants production traffic moved to the new tug-wait revision in equal steps every few minutes, not a tiny canary and not a single cutover. Which strategy fits?
• Five minutes after a canary, error rate on the new variant spikes. The pipeline must automatically return traffic to the prior revision instead of paging a human to click around. What should they configure?
• When a new labeled parcel lands in the S3 prefix, or on a cron, nobody should open Studio to click Train. Which service should start the SageMaker training or batch inference job?
• Processing, training, a metric condition, and registration are still run as four notebook cells. The team needs a repeatable DAG that builds the model when invoked. Which service automates that build?
• A merge to main should start the SageMaker pipeline (or a small AWS Step Functions deploy machine) and then update the endpoint. Today those are two bookmarks on a wiki. Which in-scope services should invoke that build and deploy?
• Featurizer helpers and inference serialization broke twice last quarter and were only caught in production. The Build stage has no unit-test step. What should the CI stage add?
• Unit tests pass, but the staging endpoint rejects the real request schema the clerk UI sends. The pipeline ships to production anyway. What test is missing?
• A green pipeline never exercised the path from the S3 drop through the orchestrated job to a scored output the downstream report reads. A silent contract break shipped last month. What should CI/CD add?
• New weigh-ticket labels arrive every Monday. Retrain is a calendar reminder. Official 3.3 skill is a mechanism that retrains when the trigger fires, then proceeds only if the evaluation gate passes. What should they build?
• The ML pipeline assumes a curated prefix that a person loads by hand. Official 3.3 knowledge includes automating data ingestion with the same orchestration so training or inference jobs see fresh data without a human copy. What should they add?