AIMentor — AI Expertise Learning Platform

13 questions · pass mark 80% · 20 minute limit. The timer auto-submits at zero.

Time remaining: 20:00

Question 1easy
What is the prerequisite for scaling an LLM service with horizontal replicas?
A bigger modelThe app must be stateless — no per-user data in process memory, so any replica can serve any requestA GPU per replicaDisabling the vector store
Question 2medium
Under heavy load, adding more replicas doesn't help throughput. What's the likely bottleneck for an LLM app?
Disk spaceThe model provider's rate limits (tokens/requests per minute)The number of DockerfilesThe favicon size
Question 3medium
How do you handle hitting a provider rate limit under load? (Select all that apply)
Queue requests and apply exponential backoff/retry on 429sRequest a quota increaseUse async I/O so each replica handles many concurrent in-flight callsSimply add more replicas to raise the provider's cap
Select all that apply.
Question 4medium
In Kubernetes, what does a Deployment guarantee?
It stores vectorsIt keeps N healthy replicas of an image running, self-healing and doing rolling updatesIt is a stable network endpointIt provisions the cluster hardware
Question 5medium
What is a Kubernetes Service?
A single containerA stable network endpoint that load-balances across a deployment's (changing) podsAn autoscalerA secret store
Question 6medium
What does a HorizontalPodAutoscaler (HPA) do?
Encrypts secretsAdds/removes replicas based on load (e.g. CPU or a custom metric) within min/max boundsBuilds container imagesRoutes traffic between clouds
Question 7hard
For a single scalable service, what is usually the right platform instead of Kubernetes?
A self-managed Kubernetes cluster, alwaysA managed container platform (Azure Container Apps, Cloud Run, App Runner) — most of K8s' benefits, far less opsA personal laptopA spreadsheet
Question 8hard
Why is 'we'll put it on Kubernetes' often an over-engineering reflex?
Kubernetes can't run containersA cluster you don't need is ongoing operational cost and risk; most apps are fine on a managed platformKubernetes is illegal in the EUIt only works for training
Question 9medium
Where should secrets live at scale, and how do they reach a container?
In the YAML manifest as plain textIn a secret store (Key Vault / Secrets Manager / vault-backed K8s Secret), injected at runtime via a reference (e.g. secretKeyRef)Baked into the imagePrinted in logs
Question 10hard
What does Infrastructure as Code (Terraform/Bicep) provide?
A faster modelDeclarative, version-controlled provisioning of infra so environments are reproducible and every change is a reviewable diffAutomatic prompt tuningA vector index
Question 11hard
Which are orchestration/scaling pitfalls? (Select all that apply)
Holding per-user state in process memory (breaks horizontal scaling)Assuming more replicas raise the provider's rate-limit capReaching for Kubernetes before you need itProvisioning infrastructure with version-controlled IaC
Select all that apply.
Question 12medium
An app must be ____ (one word) — holding no per-user state in memory — to scale horizontally across replicas.
Question 13hard
Which Kubernetes object (acronym) automatically adds/removes replicas based on load?

0 / 13 answered

Orchestration at Scale — Final Exam