Python cProfile: inspect call counts without asserting timing

Operation contract

A settlement calculation runs three times under one profiler. The resulting Stats object is queried for the function's call count, while the last business total is asserted independently. This verifies instrumentation without printing timing numbers that change with hardware and process load.

Failure boundary

Profiling adds overhead and can change scheduling. A deterministic call count says nothing about production latency, input-size growth, or memory allocation. For a bottleneck decision, profile representative data and inspect cumulative time across the actual call graph, then benchmark the revised implementation separately.

Working program

import cProfile
import pstats

def net_total(amounts):
    return sum(amount - 3 for amount in amounts)

profile = cProfile.Profile()
for _ in range(3):
    last_total = profile.runcall(net_total, (47, 48, 49))
statistics = pstats.Stats(profile)
calls = sum(record[1] for key, record in statistics.stats.items()
            if key[2] == "net_total")
print("last_total", last_total)
print("calls", calls)

Output

last_total 135
calls 3

Costs and limits

Instrumentation records function events and adds runtime overhead. Its storage grows with the number of distinct recorded call sites, not with this printed call count alone.

Common Mistakes

• Do not assert exact profile seconds in a portable test.
• A call count is not a latency measurement.
• Small synthetic inputs can hide the expensive path in production.

Connected lessons

• Python tracemalloc: distinguish retained Python allocations from a process memory claim
• Python unittest: test rejected input and preserve state after failure
• Python compatibility checks: parse syntax and execute contracts as separate evidence

Test this contract.