Python Performance

Python Performance & Concurrency

Profiling, optimization, and async patterns for Python.

Before Writing Code

1. Read references/pythonic-style.md for style conventions
2. Check Python version: 3.13+ enables free-threaded concurrency options
3. Profile before optimizing - never guess at bottlenecks

Reference Files

Profiling First

cProfile (CPU)

python -m cProfile -o output.prof script.py
python -m pstats output.prof  # Interactive analysis

import cProfile
import pstats

with cProfile.Profile() as pr:
    main()

stats = pstats.Stats(pr)
stats.sort_stats("cumulative").print_stats(10)

line_profiler (Line-by-Line)

uv add line-profiler
kernprof -l -v script.py  # Requires @profile decorator

py-spy (Production)

py-spy record -o profile.svg -- python script.py
py-spy top --pid 12345  # Live profiling

memory_profiler

uv add memory-profiler
python -m memory_profiler script.py  # Requires @profile decorator

Concurrency Pattern Selection

Async/Await Patterns

Basic Async

import asyncio

async def fetch_data(url: str) -> dict:
    await asyncio.sleep(1)  # Simulates I/O
    return {"url": url, "data": "result"}

async def main():
    result = await fetch_data("https://api.example.com")
    print(result)

asyncio.run(main())

Concurrent Execution

async def fetch_all(urls: list[str]) -> list[dict]:
    # Concurrent I/O eliminates sequential waiting: 10 URLs at 100ms each
    # complete in ~100ms total, not 1000ms. Critical for API aggregation.
    tasks = [fetch_data(url) for url in urls]
    return await asyncio.gather(*tasks)

Error Handling

async def safe_fetch(url: str) -> dict | None:
    try:
        return await fetch_data(url)
    except Exception as e:
        print(f"Error: {e}")
        return None

async def fetch_with_errors(urls: list[str]):
    results = await asyncio.gather(
        *[safe_fetch(url) for url in urls],
        return_exceptions=True  # Don't fail on first error
    )
    return [r for r in results if r and not isinstance(r, Exception)]

Timeouts

try:
    result = await asyncio.wait_for(slow_operation(), timeout=5.0)
except asyncio.TimeoutError:
    print("Operation timed out")

Async Pitfalls

# WRONG: Forgetting await
result = async_function()  # Returns coroutine, doesn't execute!

# WRONG: Blocking the event loop
import time
async def bad():
    time.sleep(1)  # Blocks everything!

# CORRECT
async def good():
    await asyncio.sleep(1)  # Non-blocking

# WRONG: Calling async from sync
def sync_func():
    result = await async_func()  # SyntaxError!

# CORRECT
def sync_func():
    result = asyncio.run(async_func())

Optimization Patterns

Data Structures

# Sets use hash tables for O(1) membership testing. For collections
# checked repeatedly or exceeding ~10 items, the overhead of hashing
# pays off vs linear scanning.
if item in items_list:  # O(n) - scans every element
if item in items_set:   # O(1) - hash lookup

# String immutability in Python forces reallocation on every +=.
# For N concatenations, this creates O(n²) total copying.
# join() pre-calculates final size and allocates once.
result = ""
for s in strings:
    result += s  # Creates N intermediate string objects

result = "".join(strings)  # Single allocation, O(n) total

List Comprehensions

# List comprehensions bypass method dispatch overhead (no .append lookup)
# and CPython pre-sizes the result list. For simple transforms, prefer
# comprehensions unless loop body requires complex logic or side effects.
result = []
for i in range(n):
    result.append(i**2)  # Method lookup + call per iteration

result = [i**2 for i in range(n)]  # Optimized bytecode path

Generators for Memory

import sys

list_data = [i for i in range(1_000_000)]  # ~8MB in memory
gen_data = (i for i in range(1_000_000))   # ~100 bytes (lazy evaluation)

# Generators yield values on-demand, preventing memory exhaustion when
# processing data larger than available RAM. Essential for log parsing,
# ETL pipelines, or any single-pass data processing.
def process_large_file(path):
    with open(path) as f:
        for line in f:  # File iterator is itself a generator
            yield process(line)

Caching

from functools import lru_cache

@lru_cache(maxsize=128)
def expensive_computation(n: int) -> int:
    # Cached: subsequent calls with same n return instantly
    return sum(i**2 for i in range(n))

# Clear cache if needed
expensive_computation.cache_clear()

Local Variables

# Python's LEGB resolution requires dictionary lookups for global/nonlocal
# variables, but uses direct array indexing for locals. In tight loops with
# 10k+ iterations, this overhead compounds significantly.
GLOBAL = 100
def slow():
    for i in range(10000):
        x = GLOBAL * i  # Dict lookup each iteration

def fast():
    local = 100  # Captured once at function entry
    for i in range(10000):
        x = local * i  # Direct array index

Multiprocessing (CPU-Bound)

import multiprocessing as mp

def cpu_task(n: int) -> int:
    return sum(i**2 for i in range(n))

if __name__ == "__main__":
    with mp.Pool(4) as pool:
        results = pool.map(cpu_task, [1_000_000] * 4)

Async + Sync Integration

import asyncio
from concurrent.futures import ProcessPoolExecutor

def cpu_bound(n: int) -> int:
    return sum(i**2 for i in range(n))

async def main():
    loop = asyncio.get_running_loop()
    with ProcessPoolExecutor() as pool:
        result = await loop.run_in_executor(pool, cpu_bound, 1_000_000)
    print(result)

asyncio.run(main())

Testing Async Code

import pytest

@pytest.mark.asyncio
async def test_fetch():
    result = await fetch_data("https://api.example.com")
    assert result is not None

@pytest.mark.asyncio
async def test_timeout():
    with pytest.raises(asyncio.TimeoutError):
        await asyncio.wait_for(slow_operation(), timeout=0.1)

Benchmarking

import timeit

# Quick benchmark
time = timeit.timeit(lambda: my_function(), number=1000)
print(f"{time:.4f}s for 1000 runs")

# pytest-benchmark (uv add --dev pytest-benchmark)
def test_performance(benchmark):
    result = benchmark(my_function)
    assert result is not None

Workflow Integration

Best Practices

1. Profile before optimizing - find real bottlenecks
2. Use async for I/O - network, disk, databases
3. Use multiprocessing for CPU - heavy computation
4. Consider nogil (3.13+) - threading for CPU-bound
5. Use appropriate data structures - set/dict for lookups
6. Cache expensive operations - lru_cache
7. Use generators - for large data iteration
8. Avoid blocking in async - no time.sleep()
9. Handle cancellation - catch CancelledError
10. Benchmark changes - prove improvement