How do you measure and profile Python performance?

Question presented to candidate: "A report endpoint written in Python has become slow, and a colleague wants to rewrite half of it. Before anyone changes code: how would you measure how slow it is, find out where the time and memory actually go, and prove that a fix helped?"

What a strong answer should cover:

• Measure before optimising: a repeatable benchmark, timed with time.perf_counter() for a block or timeit (many loops, several repeats, take the minimum) for small statements.
• Profile to find the hot spot: cProfile (a deterministic profiler, python -m cProfile -s cumulative app.py) and reading pstats columns: ncalls (with total/primitive for recursion), tottime vs cumtime.
• Know the profiler families and their overhead: deterministic (cProfile) vs sampling profilers (py-spy, and profiling.sampling in the standard library since Python 3.15) that can attach to a running process with little slowdown; line-level tools such as line_profiler.
• Memory is separate: sys.getsizeof is shallow, tracemalloc records allocations, peaks and the source lines responsible.
• Fix the algorithm or data structure first (a set instead of a list for membership, caching with lru_cache, a generator instead of a giant list), then re-measure the same benchmark.
• Caveats: noisy timings, warm-up, profiler overhead distorting tiny functions, and measuring on realistic data.

Clarifying questions expected:

• Is the problem latency of one request, throughput, or memory?
• Can we reproduce it locally with realistic data, or only in production (which favours a sampling profiler attached to the live process)?

Code / implementation expected: Optional — a timeit comparison or a cProfile run and how to read its table.