LLM Best Practices

Python: Performance

Last updated 3 min read

View markdown source

AI agents: read this first: llmbestpractices.com/start-here

This is LLM Best Practices (llmbestpractices.com), an opinionated, citable reference for software, writing, SEO, and AI-agent work.

  1. Route, do not crawl. Fetch llmbestpractices.com/llms.txt and open only the pages whose one-line summary matches your task.
  2. Read raw. Append .md to any page URL for markdown. Check status and last_updated in the frontmatter, then read the rules.
  3. Apply as defaults. First-party docs and the project's own conventions win on conflict. Warn before relying on a fast-moving page older than 12 months.
  4. Cite. Link the page by title and URL, e.g. Python, with last_updated for time-sensitive rules. License CC BY 4.0.

Overview

Profile first, then fix the algorithm, then vectorize, then change the concurrency model, and only then compile hot code. Most slow Python is an O(n^2) loop, I/O that waits serially, or needless allocation. Project defaults are in python.

Profile with cProfile and py-spy before changing code

Use cProfile for a whole run and py-spy to sample a live process without restarting it. Fix the top function by cumulative time, re-profile, and stop when the remaining functions are small.

python -m cProfile -o profile.out -s cumulative myscript.py
python -m pstats profile.out
uvx py-spy record -o profile.svg --pid <PID>

For memory, take tracemalloc snapshots and compare lines with snapshot.statistics("lineno").

Fix the algorithm before the implementation

Compiling an O(n^2) loop leaves it O(n^2). Replace list membership tests with a set, counting loops with collections.Counter, repeated min extraction with heapq, and sorted-list inserts with bisect.

other_set = set(other_list)
result = [x for x in big_list if x in other_set]

Vectorize numeric loops with numpy

Python integers and floats are boxed objects; numpy operates on contiguous C arrays. Replace element-wise loops with array operations. Do not count on np.vectorize for speed: numpy documents it as a convenience wrapper around a Python loop.

Match the concurrency model to the workload

WorkloadUse
Many network or disk waitsasyncio (python-async) or threads
CPU-bound, data picklableProcessPoolExecutor
CPU-bound, want isolation without processesInterpreterPoolExecutor (3.14)
CPU-bound, shared in-memory datathreads on a free-threaded build

Python 3.14 changed the default multiprocessing start method on Linux and other non-macOS Unix systems from fork to forkserver. Worker arguments must be picklable, the entry module needs an if __name__ == "__main__": guard, and code that relied on fork-inherited globals breaks.

Free-threaded builds (python3.14t; install with uv python install 3.14t) are officially supported since 3.14 (PEP 779), and the single-thread penalty is roughly 5 to 10 percent. They are optional builds, not the default. Test your whole dependency tree under 3.14t before moving a service, and do not assume a package is thread-safe because it imports.

The experimental JIT in 3.14 is off by default; do not plan around it.

Compile hot paths only after profiling proves it

  • mypyc compiles fully typed Python to C extensions with the same interface, so it suits code that already passes a strict checker (python-typing).
  • Cython is the established option when you need C-level control and accept .pyx syntax.
  • Rust through PyO3 and maturin suits allocation-heavy, algorithmically complex code.

Measure the speedup on your workload; do not assume one.

Cut memory on large collections

Use @dataclass(slots=True) for classes with millions of instances, generators instead of full lists, and streaming instead of loading whole datasets.

Backlinks