Design the replica count together with shared failure boundaries
SLO (Service Level Objective): the measurable quality target a service aims to meet over a defined observation window.
Design is not a product list; it is the work of turning user requirements into measurable conditions and dependencies. The maximum number of users alone does not determine the GPU count. You also need the input length, concurrency, response latency, availability, data retention, and operations staffing.
Linking requirement IDs to SLOs, resource assumptions, failure domains, and verification items makes the reasons for design choices traceable. Mark items without figures as unconfirmed and define the tests that will resolve them. If you pick equipment first and fit the numbers afterward, the failure boundaries you need will be missing.
The fixture places two replicas on different nodes, but they use the same rack PDU. There is headroom for a node failure, yet both stop on a PDU failure. If rack-level availability is a requirement, either split the power paths or renegotiate the requirement through the approval process.
Read the placement fixture, and write down why it fails the rack failure requirement along with two design fixes.
- Why does this happen?
- Linking requirement IDs to SLOs, resource assumptions, failure domains, and verification items makes the reasons for design choices traceable. Mark items without figures as unconfirmed and define the tests that will resolve them. If you pick equipment first and fit the numbers afterward, the failure boundaries you need will be missing.
- When is it a problem?
- If requirements are unconfirmed or shared boundaries are unaddressed, there are insufficient grounds to proceed.
- Common beginner misconceptions
- Do not treat unmeasured throughput or power headroom as a firm figure in a quote. Write the assumptions and units in the cost table, and mark items that are not yet settled as unconfirmed.
- How to verify it yourself
- Organize the functional, performance, availability, security, and recovery requirements by ID. Attach assumptions and a verification method to each.