Interview exercise: PerfScope investigation

PerfScope is a small performance-measurement tool this team inherited. Users have reported that some of the numbers it prints cannot be true. Your job is to investigate.

What you need

• Python 3.9 or newer, on Windows, macOS, or Linux. No other requirements.
• Any editor, IDE, or debugger you like. Use whatever you are fastest with.
• Any AI agent you prefer

Getting started

1. Run the tool:

   python main.py          # macOS / Linux (or: python3 main.py)
   py -3 main.py           # Windows
   

2. Useful flags while exploring: python main.py --list shows the available workloads, python main.py --help shows everything else. --verbose prints extra detail that may or may not be useful to you.

Your tasks

Three oddities have been reported against this tool. For each one: reproduce it, find the root cause, explain it out loud, and fix it.

1. Watch the [data] lines on a first run (delete the data/ folder to force a fresh download). Compare the reported download time and speed -- and the remote freshness check time on every later run -- with what your network can plausibly do. Something is not right here.

2. Run the full suite, then run python main.py --only scan_first_luxury. Compare both results with the reference outputs in docs/USAGE.md. Something is not right here.

3. Compare the mean ms column of the report with the [total] suite finished in ... s line and with how long the run actually felt. Something is not right here.

Each oddity has a single, small root cause. None of them require performance tuning, new dependencies, or large rewrites.

Ground rules

• Think out loud. We care about how you investigate at least as much as whether you finish.
• Expected effort: 10-15 minutes for each of the oddities. You may pick only 1 oddity to work on. It is fine to finish fewer than three -- depth beats speed.
• You need to understand and be able to explain your findings in your own words, no matter what AI agents you use to assist.

Good luck.