Design Lab — Pick the smallest compute that still fits
AZ-305 › Unit 4 › Design compute solutions
Design Lab — Pick the smallest compute that still fits
Lab brief
- 4 — Design infrastructure solutions
- 30–35%
- 20 minutes
- $0 — no subscription needed
- Core
Compute questions are decided by what the workload actually needs, not by what the platform can do. Every option here can run the workload; only one runs it without buying capability the team will never use and cannot operate. The reliable failure mode on this exam is recommending Kubernetes to a team of three.
Read the brief, commit before you reveal, then check.
The brief
Tailwind Trading — platform consolidation
A function fires on each queue message, runs for about $15$ seconds, and must reach a database over a PRIVATE endpoint — no public path is permitted. Traffic is spiky: zero most nights, thousands per minute at market open.
Commit before you reveal
Write four answers down first.
▶The trap in this area, stated plainly — open after you have committed
Most wrong answers here are not wrong about capability — AKS genuinely can run all three workloads. They are wrong about who operates it afterwards. When a scenario tells you the team size and says there is no platform team, that is not colour; it is the constraint the question turns on.
Question 1 — the private-endpoint clause
Workload A is event-driven, bursty, and must reach its database over a private endpoint. Which hosting plan should the design name?
Question 2 — what the container options are actually for
For Workload B — five microservices, three engineers, no custom Kubernetes resources — which statements are true and relevant to the recommendation? Select all that apply.
Select all that apply
Question 3 — the cheapest way to be interrupted
Workload C is embarrassingly parallel, tolerant of eviction, and measured on cost per frame. What should the design recommend?
Question 4 — the order that stops you buying Kubernetes
Order the compute selection method so that capability is bounded before it is bought.
Put the steps in order
- 1Choose the least capable option that clears every hard requirement
- 2Check who will operate it, and whether that team exists
- 3List the hard requirements — networking, runtime, scale floor and ceiling
- 4Confirm the cost shape matches the traffic shape, including idle
Why the runner-up loses
Four compute homes for the same container
| Attribute | Recommended | |||
|---|---|---|---|---|
| Buy it when | Event-driven code, bursty, scale to zero | Microservices without cluster operations | You need the Kubernetes API itself | Parallel jobs that tolerate eviction |
| Scale floor | Zero | Zero | At least one node, always | Zero between runs |
| Who operates it | Nobody — platform-managed | Nobody — platform-managed | A platform team you must have | Pool config, then nobody |
| What you pay for idle | Nothing | Nothing at zero replicas | The node pool, all night | Nothing at pool size zero |
| Disqualifier | Long-running or stateful work | Custom resources, operators, a mesh | No team to run it | Interactive or latency-bound work |
The row that decides most exam questions is who operates it. Capability comparisons make AKS look strictly better, because on capability it is. The scenario is what makes it wrong: a platform with nobody to run it is a liability the design created.
The shape of the answer
The traps that catch people
The method, for any question of this shape
How to attack a compute question in 90 seconds
Underline the networking clause
'Private endpoint', 'no public access' or 'VNet integration' eliminates plans before anything else is considered. It is the single most common disqualifier in this area.
Before you move on
Come back cold after the Unit 4 lessons. If you can state, for each of the three workloads, the single clause that eliminated the runner-up, the pattern has transferred.