“Made the service faster” is a promising CV claim, but it leaves important questions unanswered. What workload improved? Was the comparison fair? Did the change reduce customer waiting time or only a convenient internal measurement? A systems or performance engineer should be able to defend the evidence behind an optimisation, including the conditions under which it stops helping.
For a Rust, C++ or GPU performance vacancy, a benchmark-review interview can reveal more than an algorithm puzzle. The purpose is to assess experimental judgement and systems investigation. It is not a competition to guess the bottleneck from an incomplete chart.
Identify the performance problem you are hiring to solve
Write a workload description before a language checklist. A database engine, a request-serving application and a GPU inference runtime need different depth. Name the unit of useful work, the operating constraints and the system layer the hire will own. If the team cannot reproduce its reported slowdown, establishing a trustworthy baseline may be the first assignment.
Decide whether you need sustained engineering ownership or a bounded diagnostic engagement. A specialist brought in for a short investigation needs access to measurements and someone who will implement the result. A permanent hire needs an ongoing queue of valuable systems work. Neither arrangement works well if nobody can explain which outcome matters to the product.
Build a comparison with deliberately incomplete evidence
Give candidates a fictional benchmark packet with two versions of a service. Version B has a lower median latency in a single-user test, but the packet contains no loaded test and no evidence about correctness. Add a configuration difference, such as a changed worker count, and label the input dataset clearly. Explain that omissions are intentional and questions are welcome.
Ask whether the team should ship B as a performance improvement. A useful first response is a request to define success, establish comparable conditions and inspect what changed. The candidate need not reject the change. They should distinguish a promising result from sufficient evidence for a production decision. Let them choose the next two measurements rather than asking for an exhaustive test programme.
Explore what would make the result misleading
Use follow-ups to test the candidate’s understanding of the workload. What happens when requests arrive concurrently? Does the benchmark include initialisation or only steady operation? Are the same input sizes represented? Could the faster path be doing less work? Which errors or timeouts are excluded from the report? A strong answer connects each question to a possible distortion in the conclusion.
Ask them to specify a repeatable comparison: versions, environment, input generation, load pattern, measured interval and checks on output. They should also state what remains uncontrolled. Do not require your favourite benchmarking tool if another approach can produce the necessary evidence. For a role that will use a specific tool daily, assess that familiarity separately from the quality of the experiment.
Ask for a profiling plan before a rewrite
Now provide a simplified timeline showing work on the CPU and GPU, or a CPU profile for a service role. Ask where the candidate would investigate first and what observation would change their hypothesis. Avoid rewarding an immediate proposal to rewrite code in Rust, change the allocator or add parallelism without locating the relevant cost.
NVIDIA’s Nsight Systems guide describes focusing profiling on performance-critical regions and using annotations to interpret activity. That supports a useful interview question: which region would you capture, and how would you distinguish setup from the work you intend to optimise? It does not establish that a particular tool or optimisation is appropriate for every system.
For a CPU-focused vacancy, supply an equivalent trace and let the candidate explain the evidence they would seek. Keep the packet aligned with the actual role. A person hired to improve database execution should not fail because the assessment unexpectedly requires detailed CUDA knowledge.
Separate investigation, implementation and release judgement
| Capability | Meets the requirement | Follow-up prompt |
|---|---|---|
| Measurement | Defines a fair baseline and includes correctness checks. | What could invalidate this comparison? |
| Diagnosis | Uses observations to narrow competing explanations. | What would make you abandon this hypothesis? |
| Implementation | Explains a bounded change and its maintenance cost. | Which new failure modes does it introduce? |
| Release | Plans a controlled rollout and regression detection. | When would you revert despite a faster benchmark? |
Use “not observed” where appropriate. A verbal interview can establish a plan but cannot prove the candidate can implement it. If hands-on depth is essential, review a small code change together or use a bounded exercise with a prepared environment. Keep the assessment proportionate and explain whether documentation and AI assistance are permitted.
Check that the claimed result was their work
Ask for an anonymised example of an optimisation the candidate delivered. What did they measure personally? Which part did a colleague own? What failed along the way? How did they discover regressions? Do not insist on confidential throughput figures or proprietary code. A precise explanation of the experiment and decisions can be useful without disclosing an employer’s secrets.
Listen for how the change survived beyond the initial result. Someone had to maintain the benchmark, explain the new implementation and notice when assumptions changed. Those responsibilities distinguish a durable engineering improvement from a one-off demonstration.
For teams hiring across the UK and Netherlands, clarify access to hardware, lab equipment and collaborators before the offer. Remote work promises are incomplete if essential profiling can only happen on an inaccessible machine. Pair this assessment with our Rust hiring guide when language depth is part of the brief, while keeping the performance decision grounded in measured work.